Watch or listen to the High Capacity podcast on:

• YouTube

• Apple Podcasts

• Spotify

This episode, I talk with Zixuan Li at Z.ai, one of China’s top AI labs and the company behind the popular GLM models. Zixuan is Z.ai’s Director of Product for genAI strategy and global partnerships. We discuss Z.ai’s newest flagship AI model, the challenges of building agentic AI models, and what their longer AI strategy looks like.

Links:

• Z.ai

• GLM-5.1: Towards Long-Horizon Tasks

• Zixuan’s X/Twitter

Transcript

Kyle Chan (00:00)

Welcome to the High Capacity Podcast. I’m your host, Kyle Chan, a fellow at Brookings. I’m thrilled to be joined today by my guest, Zixuan Li, who is Director of Product for Generative AI Strategy and Global Partnerships and more at z.ai, also known as Zhipu in China, which is one of China’s top AI labs and the company behind the GLM foundation models. Welcome, Zixuan, and thanks for coming on the show.

Zixuan Li (00:27)

Thank you, Kyle.

Kyle Chan (00:29)

I have to start by asking you about the newest GLM model, the newest flagship GLM model, 5.1. What are its strengths? What can it do? What are its unique features? Anything that you can share with us at this point, because we’re recording before the official launch.

Zixuan Li (00:50)

First of all, GLM 5.1 will be a solid model on par with Opus 4.6, so it’s very strong in coding, agentic tasks, but also in general conversation, Q&A, and deep research capability. I think it’s on par with frontier models. Because GLM 5 is so strong in coding, many people regard GLM as a coding model, but it’s not, frankly speaking. We optimize in all aspects. So when you look at Artificial Analysis, it’s still leading all open-source models. So GLM-5 leads all open-source models in general intelligence, and we’ll have a stronger model, GLM-5.1.

But what’s unique with this model is that we optimize for long-horizon tasks. So it can run deeper, run longer. And our understanding of long horizon is that it doesn’t just run longer, from one hour to 10 hours. Actually, it can optimize results in a very fantastic way. Compared with GLM-5, we ran several tests, like letting the model run a CUDA kernel and optimize the CUDA kernel. Actually, it can achieve 2x the result compared to GLM-5. It’s very amazing. In other criteria, like website creation, and also some agentic tasks like Vending-Bench, like self-evolving tasks, it has about 2x results compared to GLM-5.

Kyle Chan (02:33)

Mm-hmm.

Wow, so what explains that jump in performance from GLM 5? I’m sure there are many things in that mixture, but were there certain features, certain kinds of engineering methods, even architectural changes, that helped to make that leap in the RL process?

Zixuan Li (03:04)

First, no architecture changes, because it still adopts the same architecture as GLM-5. What matters most is that we have been aware of this scenario. So before GLM 5.1, for long-horizon tasks, we might focus more on vibe coding and application creation instead of real optimization on certain tasks.

When we were about to launch GLM 5.1, we realized that longer horizon is better for the next generation. It’s better for AGI. And we were also inspired by agents like OpenClaw. You need to sleep. So what happens during your sleep? You have eight hours. So why not let your agent perform tasks for you during those eight hours?

So based on these observations and scenarios, we created new datasets, new training data, and tried to define the tasks we’re going to solve. I think that’s our competitive advantage. Compared to DeepSeek, DeepSeek is quite strong in architecture and research. But what makes GLM unique is our observation of real-world problems.

Kyle Chan (04:32)

Mm-hmm.

So you have certain long-horizon tasks that you can perform especially well. And when we’re talking about long horizon, how long are we talking about? Is it really like some of these, you know, can I wake up and have it refactor my code base and re-optimize my hard drive and answer all my emails by the time I get up? Or what are some examples of interesting, complex long-horizon tasks that you’re thinking of that GLM 5.1 is really good at?

Zixuan Li (05:09)

I think it depends on the harness. So we’re talking about harness engineering. OpenClaw is a kind of harness, or you can say the Pi agent is the harness. So it’s the Pi agent that lets your email run all the time. But what makes a model unique is that the model can understand the context. It can follow the instruction after eight hours.

A simpler model can also perform some tasks to answer your email, but it cannot perform that well after four hours or five hours. GLM 5.1, I think, can always find a better way to solve problems. So in the first 10 minutes, it has a solution. But it goes back to the original problem and analyzes its first solution and gets a better solution.

Step by step, it can finally get a solution that’s way better than its first solution. So long horizon doesn’t just mean time. It lasts very long, but I’ve given several examples that run very long and are not considered long-horizon tasks. For example, let the model count from one to one million. It may cost a day to finish the task, but it’s not a long horizon.

Kyle Chan (06:15)

Mm-hmm.

Right.

Zixuan Li (06:36)

A long-horizon model actually means you need more time to get the job done better. So if you have one hour, you have a solution within that one hour. But if you’re given 10 hours, you can perform it better. Most models can only get the job done within an hour and cannot improve afterward. But a better model, a model like Opus 4.6, can see a better result after several hours, or after several iterations. So long horizon doesn’t mean time. It actually means the level of iterations.

Kyle Chan (07:09)

Mm-hmm.

Right. Because I remember that for GLM 5, it was able to do very impressive work on basic office work, like data analysis or putting together a presentation that incorporates some charts from data extracted from reports. And it could sort of do it all in one go, basically. Is that the kind of long-horizon complex task that GLM 5.1 is improving on?

Zixuan Li (07:51)

Actually, we are improving on all office work, slides, Excel, things like that, but it won’t cost you more than one hour. So we’re actually doing something harder than that. We are improving all the office tasks, but in terms of long horizon, you can see in our tech report and our tech blog that there will be many use cases. It will be super hard for people to finish those tasks by themselves. So actually a model can really surpass human beings in those tasks.

Kyle Chan (08:30)

Mm-hmm. Mm-hmm.

We’ll come back to the question of AGI and exactly how far, like what types of humans the models can already beat. It’s not that maybe they can beat all humans at all coding tasks, but maybe most humans at many tasks. Related to this question, I want to ask how this differs from GLM 5V Turbo, which had also come out fairly recently, or some of the other models in the GLM family. Some seem to be more aimed at image generation, obviously. Some are aimed at multimodal models for agentic workflows. The Turbo model in particular seems like it was optimized especially for tools like OpenClaw. So I was wondering where you place 5.1 in this broader family and how we should see these different models.

Zixuan Li (09:28)

In terms of capabilities, GLM 5.1 is way better than 5 Turbo, because Turbo means that you sacrifice some capabilities for speed. So GLM 5.1 will be slower than Turbo, but its general capability is better than the Turbo series. And 5.1 does not have vision capabilities. So if you want to try visual understanding, you still need to use 5V Turbo. So we make a lot of trade-offs. As you can see, we trade off visual capabilities and speed and other sorts of things in open and closed scenarios. And there’s also a trade-off between OpenClaw scenarios and coding scenarios.

We make decisions based on our observations, based on our evaluations, trying to see which one matters most to our users. And we don’t have a 1 million context window. The reason is that we found out maybe that’s not the pain point. Maybe abilities matter most. Maybe some tasks need to be performed within an 80k context window, or the context crashes very soon after 100k.

Kyle Chan (11:04)

So that’s maybe not what you think is the key bottleneck. Is it also other factors like better tool calling, better integration with other kinds of systems? I don’t know if there’s agent-to-agent integration as well, or agent orchestration. Maybe that’s more important than just sheer context.

Zixuan Li (11:29)

Yes, exactly. And we also make some trade-offs even within a single category. So when you refer to tool calling, what tools are you thinking about? Claude Code tools, or OpenClaw tools, or your self-defined tools, or other things? Sometimes when you improve that, you’ll see some dropping in other categories.

Kyle Chan (11:36)

Mm-hmm.

Do you think that one day this will all be combined into a single vision-capable multimodal model that will basically supplant all the others, like the entire GLM family will merge and converge onto a single model? Do you think that’s the eventual goal?

Zixuan Li (12:16)

Yes, that’s our goal. Frankly speaking, we know how to do it. But we need to launch GLM-5.1 first. I think a 1 million context window and multimodal capabilities are necessary for the future. So that’s our goal. And we might have different things, like Qwen 3.5 Omni. You can see voice and other modalities are already merging into the model. We’re observing all the feedback on whether these capabilities help people in their real-world scenarios.

We are not just a research lab. We do things that help people solve problems. So we define the problems that matter most. On my X account, I frequently do surveys like, what matters more to you? These feedback signals are very useful to our researchers, and they can hear real-world user feedback. We collected more than 10,000 responses from that survey. It’s very quick and very fast. It lets them see the drawbacks of GLM-5, for example, because they might think, okay, GLM-5 is pretty good. Nobody will choose “the capabilities are not great enough.” But frankly speaking, there are a lot of people still choosing that the capabilities need to be improved. Rather than saying you don’t have vision, for many people it’s okay not to have vision, but you need to catch up with Opus 4.6.

Kyle Chan (14:16)

Mm-hmm. Yeah, ultimately that is probably the must-have, right? And then everything else is additional, layered on top of that.

Z.ai has been working on agentic AI models for a long time now, sort of before it was cool. And I was wondering, going forward, we talked about long horizon, but more generally, what are the challenges with developing these models in a way that’s geared more toward these agentic workflows? Are more of the challenges on the engineering side, trying to wrestle enough compute to be able to train and experiment with different models? Or is it how to get the right kind of RL loop going, because how are you going to give the right feedback for such a complex outcome and then iterate on that? I don’t know if there are certain specific challenges to training foundation models and developing them in the age of agents.

Zixuan Li (15:26)

Yes, there are a lot of barriers, a lot of difficulties, so I’ll mention some. First is the speed of compute. We don’t want the task to take longer, because we want the result in a minute. The reason why it seems so long is that the inference speed is not that fast. So you need to wait for the agent to perform the task. Why not have a result in a minute instead of having it in eight hours?

So for long-horizon tasks, I think we need to improve the infrastructure of agentic inference, things like that. Different architecture, not only GPU inference, but also how you organize the results, whether you do the tasks in parallel or in other forms to speed up the process. For me, I like Gemini because Gemini is super fast. We don’t need an answer that takes an hour. You just want to see it instantly.

If you finish a task within 10 hours, that’s not ideal for the general public. Maybe it’s ideal for researchers or developers, and super developers, not ordinary developers, because they don’t trust AI. They don’t want AI to perform tasks for them during their sleep. So to get more of the general public to accept this idea, we need to run these tasks super fast.

So that’s one issue. The second issue is context. When we see the agent doing tasks in its third round or fourth round, when the context window is compressed, sometimes it loses all the information and it cannot follow instructions. It’s pretty normal, not only for Chinese models but also for all the frontier models like Gemini. They declare they have a 1 million context window, but actually after several hundred thousand tokens, they just cannot recall anything or recall key information. So that’s very important.

And there are also a lot of foundational issues. Hallucination. You cannot solve it completely. The model creates something that doesn’t belong to your work or doesn’t exist. So with long-horizon tasks, it grows exponentially. The hallucination will pass from the first round to the last one.

Kyle Chan (18:34)

Yeah, a lot of labs talk about the same problems, and any user who uses these models on a regular basis will know that feeling of when the context window is starting to run out and it’s just not really responding appropriately anymore, kind of making up stuff.

Zixuan Li (18:57)

Yes, and we also lack training data because no one has done it so far. No one has performed a task over 10 hours and collected all the data and then gone back to label it.

Kyle Chan (19:01)

Mm-hmm.

Yeah, so what do you do without that kind of data, right?

Zixuan Li (19:19)

So we need more RL and we need more synthetic data. We think about synthetic data, but also try to find the hallucination inside it, try to correct all the non-instruction-following issues.

Kyle Chan (19:37)

Z.ai has been especially at the forefront in terms of working with Chinese chipmakers and Chinese hardware to support model deployment on AI chips like Huawei Ascend or Moore Threads or Cambricon. And if I recall correctly, GLM Image, the image generation model, was trained end-to-end on Huawei Ascend chips. And then more recently, in the latest Z.ai earnings report, there was a discussion of co-design and hardware-software collaboration. So I was wondering, what is the strategy behind this? Why is this seemingly a big priority for Z.ai? And what’s it like to work with the Chinese chipmakers too?

Zixuan Li (20:28)

I think the reason is quite simple. We don’t have access to NVIDIA chips. All the Chinese companies, I think, face similar issues. And we don’t have Blackwell. So that may restrict the scaling of our capacity and our performance.

When you look at DeepSeek’s tech report, the reason why they choose the number of parameters is based on the infrastructure. It’s the largest model they can train with their infrastructure. So we face similar issues. We try to select the right balance.

Kyle Chan (21:15)

So what has it been like working with the Chinese AI chipmakers? And how closely do you work together? What’s that process entail? How different is it from working with, say, NVIDIA GPUs?

Zixuan Li (21:32)

At first, we don’t work with NVIDIA GPU makers or enterprises. I think it’s fantastic working with Chinese chipmakers. But the only limitation is their supply. With some makers, we just finished the collaboration, but they haven’t produced so many chips yet. So we are still waiting for their large supply in the upcoming months.

And we co-design the chips, but they also have to figure things out for other large language model companies, because there’s not just one model company here. So we try to get an advantage there, because there’s DeepSeek. DeepSeek is also very strong in their architecture work. They have closer collaboration with these chipmakers. But I cannot share many detailed things. It’s secret. Let’s see what happens next.

Kyle Chan (22:37)

Right. That’s great.

So I want to ask a question about open source. A lot of the GLM models have been open source up until recently, and there are a lot of questions about whether 5.1 will be open source. I was wondering what your thoughts were about open-source strategy more generally, whether going forward there might be more of this hybrid approach to having some open-source models but then some proprietary ones, maybe open source for distribution but proprietary more for direct monetization. How do you see either Z.ai or the broader Chinese AI labs approaching the open-source issue?

Zixuan Li (23:30)

I think we are open to all these possibilities, whether commercialization or continuing to open source, whether to open source our flagship model or smaller ones. We are very open. That’s the first point.

And I have an understanding of open source. Especially, I think it’s not open source, because many people think it’s open weight. My understanding of this open-weight concept contains three layers.

The first layer is that through open source, you create your brand image. Compared to U.S. frontier models, not many U.S. citizens or media care about Chinese models when you are closed source. Seed is the best closed-source model in China, but nobody knows Seed. They only know Seedance, instead of Seed, as part of Doubao, right? So you need to let U.S. inference providers run your model. You need to let those people with GPUs at home try your model. That’s what got Qwen famous, what got Kimi and DeepSeek famous. So that’s the first point. I think it’s especially necessary for a Chinese model company to open-weight your model when you want brand image and want more people to know you.

The second layer is that through open source, you collaborate with the community. So you gather help from others. For example, Intellect-3 is based on GLM 4.5 Air, and we see a lot of people using GLM. They fine-tune GLM, not the original GLM. They fine-tune it, they quantize it. They use a lot of techniques to make open source embeddable, like what Cursor just did with Kimi K2.5. It’s pretty influential.

We won’t make that happen by ourselves, because our domain knowledge and expertise can only get it to maybe 80%. But with the knowledge of the whole community, or their domain knowledge, they can improve it into a better model. They can truly maximize the potential of the foundation model. That’s why we call it a foundation model, because it’s a foundation.

Unless we open-weight all the models, can people truly utilize this as a foundation? If we only provide the resource through APIs, people can only use the API and only pay for APIs. It will restrict the capacity and the potential.

And the third layer is that we try to define the norm. It’s the highest target. I think we are close to that, but only DeepSeek and Llama have reached that level. So DeepSeek defines what thinking tokens look like, what the panel looks like, because people only see it from o1, but they don’t know the secret. It teaches people what’s behind the model and truly defines a pattern, a norm. I think only through open source or open weight can you fully let that happen. That’s our goal. We want to see some training patterns, some model architecture, or the behavior of the model become a pattern for the world. Like long horizon. Maybe after three months, everyone is learning how to do long-horizon tasks.

Kyle Chan (27:42)

Mm-hmm. Yeah, but you get to set the trend, basically, and that’s very valuable.

Zixuan Li (27:53)

Yes. They can learn our data pattern, or they can deploy on their chips to see what happens to the model if they do this or if they do that.

Kyle Chan (28:10)

Yeah.

Do you think that there is, or how much stickiness do you think there is for enterprise customers that are building on the GLM foundation models? Once they get used to building with your models, deploying them, customizing them for their own use, fine-tuning them on their own proprietary data, how much do you think that keeps them wanting to come back to GLM models in their next iteration rather than switch over to another one? Versus if you only have the API service, then it’s like, okay, I’ll just plug and play another API.

Zixuan Li (28:52)

Very sticky, because you still see many people using Qwen 2.5. There are lots of models based on Qwen 2.5.

Before Claude Code, maybe we used workflows, like predefined workflows, like DeFi or other things. The workflow is very complicated. If they think it’s working, if it’s effective for their domain expertise, I think they won’t switch to another model unless they want to fully switch the pattern to another thing, like an autonomous agent. If they use workflow and Qwen 2.5 or a variant is enough, they’ll keep it. It’s the same for GLM 4.5 Air, because when you look at ElevenLabs, on their platform they use GLM 4.5 Air and GPT OSS as a foundation.

Kyle Chan (29:46)

Yeah, that’s really interesting. Does that feed back into the way that you guys develop your models? As you’re thinking forward for later generations, are you trying to retain certain features that you know current existing customers really like and want to have, so that when they update with a more recent GLM, they still feel confident that their systems won’t just break and they can keep building on what they had before?

Zixuan Li (30:37)

Yes, exactly. As you can see, 4.5, 4.6, and 4.7 share the same architecture, to make that shift more convenient. And our training data share the same style, so it won’t shift to another area or aspect.

We’re trying to perform better and better, not to perform completely different capabilities. We’re going to strengthen all the aspects we already have. That’s our primary goal.

Kyle Chan (31:16)

Yeah, that makes sense. There’s always going to be some trade-offs because as you add more capabilities to the model, there will be some novel components, and the question will be how they can fit into people’s workflows, or how you can educate customers and convince them that this is worth trying out.

Zixuan Li (31:41)

Yes. I have an example. ArcAGI is quite popular lately, but we don’t train on similar problems because that’s not what our customers are looking for. Our customers are using GLM in Claude Code and in their agentic workflows. They’re not using it to solve math problems or those kinds of problems. So we won’t make that trade-off to let the model improve on ArcAGI instead of improving in Claude Code.

Kyle Chan (31:55)

Mm-hmm.

Right. In that sense, are there certain benchmarks that you care more about because they’re closer to the real-world use cases that you have in mind, and that your customers will use these models for? So maybe it’s not ArcAGI, maybe it’s not some of the math exams, maybe it’s more of these BrowserComp or real-world search benchmarks, things like that.

Zixuan Li (32:51)

We have proprietary benchmarks, but we are about to open source some of them. So we have CCBench, Cloud Code Bench. As you can see in the X post, we use that instead of a well-known benchmark. In that benchmark, you can see that GLM 5 cannot compare to Opus 4.6, it only has a score around three-fourths of Opus 4.6, but GLM 5.1 is pretty close.

So we try to define these problems especially in real-world scenarios, because when we look at other benchmarks, like WebBench or WebBench Pro, some are using a very ideal environment. So the agentic environment does not capture the real-time user experience. We try to use user feedback to make these benchmarks.

And we also have a benchmark for OpenClaw. It’s called Z-Claw Bench. We have already open-sourced it. About 70% of the questions are Chinese. I think that’s okay because we have so many Chinese customers. It’s okay to blend those Chinese queries into the bench. So we optimize on these benchmarks. I think it’s super great. It captures all the necessary questions you can ask inside OpenClaw. It’s on Hugging Face. Everyone who is interested in this benchmark can search it on Hugging Face and try to translate it from Chinese to English or their language.

Kyle Chan (34:37)

Yeah, but I’m sure that more of the OpenClaw-type benchmarks will come out because there’s a lot of demand for trying to understand which models are best for that kind of use case.

Do you think that there will be efforts for either z.ai or other Chinese AI labs to create their own internal OpenClaw, rather than building on the OpenClaw platform, the way Claude now kind of has its own computer-use tool?

Zixuan Li (35:27)

I think it depends. Inside Z.ai, we have several teams building variants. Some are based on the same harness, some are based on their own harness, but they try to use the class and the name.

Kyle Chan (35:41)

Right. People have the image of the claw, right? So that brand is very powerful.

Zixuan Li (35:49)

Yes, we have like five to seven teams working on it. Because you have different customers. You have customers that have some knowledge of OpenClaw, but you also have people who are not aware of this thing. You have to make it a more convenient product for those customers.

Kyle Chan (35:56)

Yeah, that makes sense. That’s what I kind of see with what Anthropic is trying to do, where they try to have at least some version that’s a more simplified, user-friendly version for those who are just used to downloading a piece of software and running it on their desktop.

Zixuan Li (36:31)

Yes, they also have remote control.

Kyle Chan (36:34)

Right, with the app on your phone. So you can literally just be in bed and run more agentic tasks.

Another question that I have is: z.ai had a really high-profile IPO earlier this year, and you’re now a public company with incredible valuations. I was wondering how different it feels now to be a public company. Do you think it affects the day-to-day work that you do? Is there some relief now that you made it to the IPO and got that financing round, or is it now more pressure because there’s a stock price that everyone’s probably aware of somewhere in the back of their minds?

Zixuan Li (37:25)

No changes, frankly speaking, no changes. That is because we focus more on research and application, and we regard ourselves just as a startup. We are closer to zero rather than closer to one or a hundred.

The final strategy of this company is chasing AGI. But we are nowhere near AGI right now, based on our definition, because we need AI to manipulate real-world tools to help people work on real-world tasks. Rather than making slides or doing your Excel, I think that’s superficial right now.

So for the company, we are still at the very beginning stage.

Kyle Chan (38:27)

Yeah, so what do you see as the next steps to getting to AGI? What are the key thresholds? Is embodied AI a necessary part of that, being able to interact with the physical world in some kind of embodied form? World models are kind of the big thing right now. What do you see as the stepping stones there?

Zixuan Li (38:56)

For large language models, we’re going to do the same thing, enhancing coding capabilities and agentic capabilities. But for other ideas like embodied AI, we’re trying something similar because we have performed quite well on Vending-Bench. So we are about to generate a real-world Vending-Bench, a vending machine. We’re about to have a physical vending machine that operates on its own. It will purchase all the items, calculate all the things, and do the payment by itself, interact with the customers, all through that physical vending machine. So we’ll try something new.

And that’s also the reason why we want to do visual large language models. I think the current use cases focus more on taking screenshots and replicating a website or doing OCR. But in real-world use cases, you need eyes. You need to see everything.

Kyle Chan (40:15)

Right. So maybe GLM 5V for vision was one of the steps to try to get that capability, and then later on that’ll be folded into the mainstream GLM line, perhaps.

Zixuan Li (40:43)

Yes. I think Gemini is pretty close to this target.

Kyle Chan (40:49)

How closely do you follow what is happening with your competitors in China and in the U.S.? Is it something where you’re all sort of nervous as you wait for another model release from someone else, or they’re nervously waiting for the next GLM model to come out?

Zixuan Li (41:10)

I think the most nerve-racking thing for us is missing users’ expectations. When we don’t reach product-market fit when we thought we should, there’s a lot of anxiety. If we provide something that they don’t need, I think that’s the most disappointing thing for us.

Kyle Chan (41:37)

Yeah. And then last question: thinking about your global strategy, you’ve talked about this a little bit already, but how do you see approaching the overseas market versus approaching the Chinese domestic market? Is it a very different strategy or a very different set of customer expectations, or way of interacting with customers or building up the community? What’s similar and what’s different inside China versus outside?

Zixuan Li (42:13)

I think inside China, we’re in the reputation stage. So we try to build more reputation and provide better services rather than just pure capabilities, because everyone is aware of GLM and Z.ai. In the overseas market, we’re still in the awareness stage, because only a few people have heard of it.

So I’m very grateful because today Gemma 4 mentioned GLM-5, which brings GLM-5 to many people’s minds. They ask, what’s this? Why did Google compare to this model? I’ve never heard of it.

After this awareness, we begin to create a brand image and show our capabilities and let more people use it. Then we can do some commercialization or similar stuff. But we also want to explore more things, like AI for science and other more frontier aspects.

Kyle Chan (43:30)

Mm-hmm. So it sounds like you’re going in a number of different directions, trying to build better, more powerful models, better features, try to address these different customer demands, and have fun.

Zixuan Li (43:52)

This is very tough for me because I have to talk to a lot of users. Some of them are AI researchers, top scientists, and Z.ai chat users. Some people use AI only for chat. We need to care about their feelings and their feedback. Some people want to remove their history, so I need to figure that out for them.

Kyle Chan (43:57)

Yeah, I see. So then you’ve got to prioritize which of those different requests. It sounds like a very busy time for you guys. And it’s interesting that on top of all the incredible engineering and R&D work, there’s still a lot of basic people-to-people communication, customer feedback, and doing the customer-feedback loop. Even on the way to AGI, this is still a core part of the business.

Zixuan Li (45:01)

Yes, I think I spend more time with human beings than AI lately.

Kyle Chan (45:06)

For now, I guess.

Zixuan Li (45:11)

Because things change so fast. If you have a document and you teach AI to learn from it, after two days it will be obsolete. It will be outdated. Because all the documents say the frontier model is GLM-5, but then you have GLM-5.1 and you need to change everything. But you cannot. So you need to reply to emails one by one. Someone identifies a new issue for you, and you need to figure it out. So 50% of the issues are caused by our new services rather than by the existing ones.

Kyle Chan (45:49)

Right. So you are sort of creating new challenges for yourself as you are solving old ones.

Well, that’s all the questions I’ve got for now. I think it’ll be really exciting once 5.1 is out to be able to play around with the model and try it out. I’m sure everyone has their favorite personal benchmarks or tests that they like to experiment with. I’m sure you have your own toolkit that you use to test and check each of the different models when they come out.

So I just want to thank you for taking the time to chat with me, and especially given your busy schedule. I’ll definitely include links to z.ai, the new GLM 5.1 report when it comes out, and also to your X account, which is really useful. I’ve been following it for a long time, and it’s really useful to hear not just about what’s happening with z.ai, but more broadly what’s going on in the AI landscape. So I’ll definitely include all that in the show notes.

With that, thank you very much, Zixuan, for a fantastic conversation.

Zixuan Li (47:04)

Thank you.

Thank you. Please follow us, and I’ll try to engage with everyone in the audience if you reach out to me.

Kyle Chan (47:18)

Sounds good. I’ll send them your way.

All right, so to wrap up, if you like this episode, please rate and subscribe on YouTube, Spotify, or Apple Podcasts. You can find episode transcripts and more information on the High Capacity Newsletter at highcapacity.org. I’m your host, Kyle Chan. Thanks for joining, and see you next time.

Watch or listen to the High Capacity podcast on:

• YouTube

• Apple Podcasts

• Spotify

Rui Ma, founder of Tech Buzz China, explains the real sources of China's tech strengths and where the US still has the advantage. From industrial parks to capital markets, we talk about how to understand China's tech landscape at a deeper level -- and why so many people are eager to tour Chinese factories.

Links:

• Tech Buzz China

• Tech Buzz China newsletter

• Rui Ma on X / Twitter

Transcript

Kyle Chan (00:00)
Welcome to the High Capacity Podcast. I’m your host, Kyle Chan, a fellow at Brookings. I’m thrilled to be joined today by my guest, Rui Ma, who is one of the most insightful and knowledgeable experts on China’s tech scene I know. She’s the founder of Tech Buzz China, which provides research on China’s tech landscape and also organizes trips to visit factories and startups in China. Welcome, Rui, and thanks for coming on the show.

Rui (00:27)
Hi, thanks for having me, Kyle. Great to be here.

Kyle Chan (00:31)
I wanted to start by asking you a broad question, which is something that you’ve written a lot about and clearly thought a lot about. What are China’s real strengths when it comes to technology? What is the foundation on which everything that we see and hear about in the news with AI, robotics, you name it, is built? And what do you think really sets China apart from other countries that might be targeting some of these same technologies?

Rui (01:03)
I think that really, China is just really good at building entire ecosystems around technologies, especially ones that the government has identified as future core economic drivers. And I think pretty much this—I’m not sure if I have that much more interesting insight to add, because lots of people have said China really excels in process innovation and not necessarily just coming up with new technologies. But I think it’s not just as narrow as process innovation. It’s really building the entire ecosystem, right?

And that includes not just the minute improvements in the supply chain and the factories, but thinking about how I’m going to have this industry be birthed right from scratch. How do I change these entire cities? As we were saying right before this podcast recording, we had recently gone to what is probably the battery factory capital of China. And it used to be a really chill town for sightseeing, drinking tea, and bamboo forests. How did it become this entire new energy battery capital of the world?

Well, that really has a lot of—there was a lot of thinking about how to build the educational capacity, the physical infrastructure, and then also not just building the physical infrastructure, but actively recruiting and responding to the entire industry and making sure that the entire supply chain is co-located. And what are the things that all these companies need to be co-located here, right?

So the state really does a lot and has gotten increasingly better at doing it. I think if you look back 25 years and look at the industrial parks and look at what they’re doing today, it is, I think, underappreciated how much progress the government has made in building up industrial capacity.

Kyle Chan (03:22)
Yeah, those findings from that trip are so interesting because I think they also go against maybe some people’s conventional views of what government support looks like. I think maybe, especially for a lot of Americans, you might think of tax credits and subsidies, but it’s a lot more than that, and it’s sometimes certain levers that are pulled that might be more important than just giving extra money.

Rui (03:49)
Yeah, not to say there isn’t extra money given. There definitely is. But that’s typically at the beginning of an industry, and it’s usually kind of small scale, I would say. You see that in a lot of emerging industries, and they’re giving it to small companies. They may give a big tax break, just like here in the U.S., to get an industry leader to come establish a site in the city.

But then after that, why do companies stay? Or why do they even come here? Because a lot of cities can offer the same thing. It’s really about the speed of execution and, again, the entire ecosystem building, right? So if I came here, how does this connect to the logistics supply chain that I need? Does it connect to waterways, key waterways? Does it connect to key railroad lines?

And then where am I going to get the labor? One of the things that became more obvious—you probably already know this, but it became much more obvious when you get there—is that there are spikes in manufacturing. It’s not like January to December, everyone is working eight hours a day. There are lots of spikes. When a new product comes in and there’s surprise demand, then I need flexible labor.

Right? I need to have double, triple, maybe the capacity. I need to be able to run my factories 24 hours a day and run three shifts. How do I make sure that happens and it all stays within a roughly contained area instead of me having to go to another province or redirect my entire process somewhere else? So all those things have to be thought about and contingencies planned for in advance. And again, it’s underappreciated how difficult that is to do.

Kyle Chan (05:48)
Yeah, it’s funny to see the contrast where I think in the U.S., this is normally viewed as an organic process that can’t really be accelerated that much. Or maybe that’s a feeling today. And so the idea that you get these agglomeration effects where you have not just one major firm, but multiple firms in a given industry, and not just in a given industry, but their suppliers and their suppliers’ suppliers all in one place, and then with that broader shared labor pool, and then hooking up to the infrastructure—there’s a sense that it doesn’t just have to roll in natural time, that there are certain things you can support or step in on, and maybe even some areas that you can speed up from a permitting or licensing standpoint as well. I think you had written also about this sort of one-to-20 policy.

Rui (06:47)
Yeah, many places have it. The specific city we visited was like, we’re going to give you approval in 20 days. It’s contingent upon you following through on your paperwork. We’re still going to check, but we can give you that approval for any factory site. And the goal is always to get the factory up and running, from submission for approval to things rolling off the line, in about a year.

Kyle Chan (07:18)
Yeah, yeah, yeah. So then if you are thinking about starting up or choosing a new factory site, the speed really matters a lot. It’s not just who can have the best deal on tax breaks. It’s that you want to get your business going and revenue flowing as fast as possible.

Rui (07:35)
Yeah, and everything else too. I think speed is just a core factor, not just in companies choosing where to build, but also as their core competitive advantage. In many industries, China may not necessarily have substantially better quality. And in fact, sometimes we talk about maybe it’s 80 or 90 percent the quality of somewhere else that’s world-leading.

But if you can deliver—in the case of one transformer factory that we went to, they were able to deliver, in their words, in 10 months, what global leaders, I think German and Japanese, were delivering on in the order of three to four years—then that’s how you win. So it’s not just a matter of undercutting on price or winning on quality, because a lot of these things—we’re in the real economy here—you’ve got to get the product. So it’s not just about what I see in the inventory booklet.

Kyle Chan (08:37)
Yeah, definitely. It’s funny. Maybe American consumers are used to an Amazon model of delivery, and maybe that kind of speed and nimbleness is crucial in a lot of other areas.

Rui (08:55)
In a lot of other areas as well, especially large products. I think that was a little bit surprising to me, that there was that kind of speed differential. Again, I’m not saying everything is on that magnitude of difference, but it was like, oh wow. Because that was one of our questions. It’s like, how are tariffs affecting you?

Kyle Chan (09:19)
Yeah, yeah.

Rui (09:20)
And they were like, well, they’re not really, because our customers still need their products. It’s kind of like life-saving healthcare, right? The difference is, I need this now in order to make the rest of my business work. Then I’m just going to pay up.

Kyle Chan (09:30)
Yeah, yeah, yeah, exactly. Well, I want to step back and just ask you about these factory visits. It’s so interesting right now. There’s almost a genre of people writing about their trip to China and how they went to see a Xiaomi factory and were blown away by the levels of automation and everything. And you guys help lead some really interesting trips. I was just wondering, what do you get out of visiting a Chinese tech company or visiting a factory? And what don’t you get from visiting in person?

Rui (10:20)
Yeah, well, generally, we run typically five-day trips, and then we have one day where we take you to visit the factories. A lot of it is whatever is relevant—we do bespoke trips, so it’s dependent on the sector.

I think a lot of it is people just want to say they’ve been there, because I actually think they’re very difficult to get to. And if I had a choice, I would rather not go. They’re usually in the middle of nowhere, and it takes a long travel time. And you don’t see that much. Unless you are some kind of process engineer, you’re not going to be able to understand the finer details. And you can see a lot of this online, right? There are lots of drone fly-throughs of this and that factory. I don’t personally feel like it’s that different.

But I think people just want to understand, especially people who are going to China for the first time, where are these places located? We went to a fourth-tier city, for example, for the battery stuff, and Americans in particular, I think, don’t really have a good idea of what a fourth-tier city might look like and what it might be like to live there. Where are the factory workers staying? And by the way, they’re very nice now. We did go to a top-end fourth-tier city, but they’re largely quite nice now, and they actually look very close to urban city centers, in my opinion.

So I think that kind of texture around what the entire country is doing is helpful. Again, the factory visit itself—now, lots of factories actually offer this as group services. You kind of have to put together a group and then go, and sometimes they’re more limited in what they show you. We try to go to more exclusive ones where they’re not generally open to the public. But again, what you can learn is that you can confirm that there are lots of robots. And what you can really learn is that, again, it’s not some isolated factory. It’s a whole ecosystem. They sit in large industrial parks, and you’ll see that many of the suppliers are next door or a couple of blocks away.

Kyle Chan (12:39)
Yeah.

Rui (12:41)
And that’s where the whole speed advantage comes from. And by the way, going back to what you were saying earlier, I think that is the way that the U.S. used to work. I’m not some industrial historian, but as I was posting this on Twitter, lots of people jumped in and said, well, you should look up such-and-such steel plant. And if you look it up, that is exactly how it used to be structured.

The risk, of course, is that you concentrate this industry here, and then when that industry goes away, you’re left with this potential dead zone. And I think the fallback to that is that you just have to keep on going. You can’t just rely on this industry to support your economy forever.

So you always have these core industries. The way these cities and municipal governments generally think is, I’ve got core industries, and I always have these other emerging industries I’m nurturing. And there are usually at least two or three that I’ve gotten central government approval to go invest in. And then hopefully one of those ends up becoming big and replacing one of my current really cashflow-generating businesses if it dies off.

And that’s kind of how entrepreneurs should think. You should never be like, this product is going to be best-selling forever. I either need to upgrade, or the competitive landscape changes and I need to move on to something else. Imagine if Apple only sold the Macintosh.

Kyle Chan (14:04)
Right, right, right. Yeah, yeah, yeah. So there’s always this balancing between wanting to support maybe what you see today as your core business, and maybe that’s your cash cow for right now, but knowing that you need to diversify, that there could be changes around the corner and there could be new opportunities, new industries coming out where you want to be part of that and have your foot in the door.

And maybe that applies not just to the Shanghai, Beijing, Shenzhen tier-one cities, but especially maybe further out west or further away from the major urban centers as well. I mean, they’re all urban centers now, I guess.

Rui (14:57)
Yeah, yeah, yeah. Actually, the manufacturing is largely in tier-three and tier-four cities. It doesn’t really make sense to— even in Shanghai, right? If you go to the Tesla factory, which I haven’t been to, but we’ve considered going there, it’s quite far away from the city center. You can say it’s Shanghai, but it’s really like—if you were in the Bay Area, it’s like saying your factory’s in Stockton and it’s in the Bay Area. I mean, I guess, like, an hour and a half away.

Kyle Chan (15:31)
Yeah, yeah, yeah. That kind of layout reminds me more of Germany, where there’s industry more spread out. I don’t know the exact reasons why that is the case, but you have this Mittelstand with all the tier-three, tier-four suppliers, and they are hyper-specialized in making this one component, and they’re not necessarily in Berlin or Munich. Definitely not Berlin. That’s not an industrial city anymore. So that kind of geography really matters.

Also, I was wondering—these are very high-level questions—but a lot of people ask how innovative China’s tech industry really is. There is this idea that the U.S. does zero-to-one innovation or invention, and China scales one to 100. And there’s that division of labor. Another way to put this question is, some people say, well, ChatGPT would not have been able to come out of the Chinese tech ecosystem, or mRNA vaccines, or something else. What do you think about those arguments, and how do you think about the idea of innovation itself?

Rui (16:49)
Yeah, that’s good. It’s interesting you mentioned mRNA vaccines. I’m not a bio expert, so I couldn’t tell you. But ChatGPT, I think, is a very good example of where the U.S. would be much more likely to come upon this innovation.

I have a finance background, so maybe this biases the lens through which I analyze this stuff. But I like to say that China excels at really near-term commercialization. And that’s just because of the nature of the capital markets. Capital markets overwhelmingly reward—I mean, until recently, before the AI robotics hype—but I would say even then, in general, they reward seeing cash and seeing returns and profit versus something more narrative-driven.

That’s why in the past five or six years, you see people funding pool-cleaning robots and lawnmower robots. I have plenty of VC friends who funded these companies, and that would never get funded in the U.S. because they would not be considered big enough. They would not be considered high-growth enough. But in China, you kind of have to show that you can do that. And then the idea is that if you can at least make money on this one thing, then maybe you can pivot into other things.

But you can’t be telling me, five or eight years from now, I’ll make money. No. It needs to happen ideally this quarter, but maybe next. So you see a lot more near-term commercialization happening in China. Or even with Unitree—how are they able to survive? It’s because they were like, I’m not going to make the most advanced demo. I am just going to make this good-enough physical thing that other labs can buy, because I know that I can immediately sell to research labs. That’s the near-term opportunity. So I’m going to deliver a product that is going to be able to get sales right now.

Where China also excels is these really long moonshots, like literally going to the moon or making rockets. That’s really state-driven, and there’s a whole strategy behind it.

And where the U.S., I think, is best at—and that’s really in large part because of the way our capital markets are, how everything is financed—is what I call this medium term, like five to 15 years, where there is sufficient ambiguity that in China it’s just considered too risky to put money in, especially when there are other near-term opportunities. You just have much more appetite for that level of risk.

No one wants to take the 20-year risk. It’s kind of how the venture capital industry is structured. It’s like 10 years, kind of, but in reality many funds have to extend multiple times. You’re really looking at 10 to 15 years to recover your capital. So it’s really set up for that kind of bet.

ChatGPT really falls nicely into that, right? You needed a couple billion dollars, but it was a five-to-10-year timeline. It was something that, if you believed in the technology, was kind of sequential and you could kind of understand. And it wasn’t as pure strategy as going to the moon. There are some near-term commercial implications, but in China, again, the risk appetite for that would be very low. You would not be able to get private funding to the tune of billions of dollars.

Now, I think that could change in the future, but I still think in general you can see the U.S. as being better at that. It’s just the way the entire ecosystem works.

Kyle Chan (20:57)
Yeah, I mean, this is an impossible question, but is there a cultural underpinning or some structural factor? And could this change over time? I think about how the tail can wag the dog too, where I feel like the Silicon Valley narrative-driven culture is now becoming the whole American—or maybe it is the original American—story overall, telling these big stories. AGI is just about the latest huge aspirational dream.

So I don’t know—why do you think today it’s more short-term, at least for the private capital markets in China? So you get this bifurcated private versus state capital, and that’s different, and obviously they are solving different problems, right? Reasonable rockets, you know.

Rui (21:31)
Yeah, yeah. Again, it’s really related to the whole financial structure. If I am an LP, a limited partner in China, and I’m investing in fund managers who are then investing in innovative companies, I tend to be someone who doesn’t have access to this steadily compounding 7 percent S&P market. That doesn’t exist for me. So I need you to kind of return me money. I’m not looking at putting 90 percent of my capital in these safe, steadily appreciating assets and then going crazy with the remaining 10 percent. I kind of need to have reasonable risk on everything I’m doing.

So in the past—again, this is kind of changing—but I would say overall, the investment horizon is just so much shorter in China. It used to be—by the way, there were RMB funds that had three-year, not deployment, but three-year fund lives. And then even eight was kind of like, you had to be a really great fund manager. And then 10 was really unheard of. Whereas here it’s like, yeah, 10 is the standard. In reality, everyone knows it’s more like 12 to 15, like I said.

But that’s because everyone’s acting rationally due to the relative attractiveness of the other investment opportunities they have. People are not stupid.

And I don’t ascribe personally to the cultural explanation. Of course, there are some cultural things, like China tends to be very hierarchical. Those are social things. But the actual innovation is purely incentive-driven. Why are Chinese entrepreneurs—I’ve talked to so many of them—so interested in cloning a known technology? Why are they like—

I used to joke, but this is absolutely true: before Replit’s pivot, every Chinese entrepreneur I found was trying to make some clone of Replit, some clone of Zoom, some clone of Miro. And I’m like, why are you doing that? No one will fund this in the U.S. But in China, they will.

Again, because of the lower level of risk that investors are willing to take, they want to see that your Tsinghua team can show me that you’re going to be 30 percent more cost-effective than this unicorn company that’s already proven the market. Whereas in the U.S., you’re a VC, especially a tier-one or tier-two VC, and you tell your LPs that, hey, I just funded the 10th clone of Replit. People are going to be like, that’s a bad strategy. What are you doing? Where are the outsized returns?

People don’t even want you to fund the third place these days. They really want first, maybe second.

So it’s just a very different ecosystem. And because that behavior—again, it’s the entire ecosystem—those are the incentives. That’s what leads to the quote-unquote innovation. By the way, when U.S. capital was more active in the mobile internet era, you saw a lot more crazy company ideas. And that was because the funding came from the same sources. It was a lot of U.S. LPs, and it was a much smaller market. So again, it was more USD-driven.

And now it’s an overwhelmingly RMB-driven funding market, and overwhelmingly domestic capital market exits, with more state capital participating. They’re going to have different incentives, and people are going to be working on different things. No state fund entity is going to sit here and listen to your crazy pitch that has no connection to the real economy.

Kyle Chan (25:28)
Yeah. Right, yeah, yeah. Well, what about corporate VCs or Chinese big tech companies acting like a kind of VC? What kind of role do you think they play in exit strategy as well? How analogous is an Alibaba or Tencent to a Google or Amazon in terms of playing this role of maybe helping to incubate some of these early-stage companies and then maybe eventually pulling them in, or at least being this potential signal at the end of the journey that there could be this big exit, that maybe they could be a strategic buyer or whatever? Or is it a different kind of logic for the Chinese big tech corporates?

Rui (26:28)
I think the big tech corporates—the few of them, because Alibaba, Tencent, et cetera come from an earlier era, they’re listed overseas, and they’re a little bit closer to Google, Facebook, et cetera than you would think. But again, the competitive landscape in China is different. They tend to be competing head-to-head with each other in every way.

So acquiring could be good, but when you have really broad distribution, maybe you should just build it yourself. And that’s what Tencent and Alibaba were kind of accused of, right? Tencent specifically was accused of just looking at the best ideas and then trying to build them in-house. Alibaba actually did a lot more strategic acquisitions and strategic investments. But then in order to appease the regulatory authorities because of antitrust issues and blah, blah, blah, I think they’ve chilled out a little bit more on that.

And there are lots of strategic acquisitions—sorry, more like strategic investments. Outright acquisitions to the tune of billions of dollars, I don’t think have really—those have rarely happened. Because in the mobile internet era, maybe we were just too early into this next paradigm shift, but in the mobile internet era you did see a lot of mergers of largely unprofitable businesses, with first and second place having to get together and cling to each other for survival. But for many of these, you don’t see the Facebook-acquiring-WhatsApp-for-billions kind of madness. I just don’t think that would happen here in China.

Because again, the distribution channels—these companies own so much of it. They are much more likely to build it as part of their strategy in-house or have more of a partnership.

And also, they’re much smaller. Even though they’re huge companies, they’re not trillion- or multi-trillion-dollar companies, and they’re not spinning off $40 billion of free cash flow or whatever it is that Meta has or Google has. They’re really strong, but they’re really strong in the Chinese economy, whereas these other companies are strong globally. It’s just a very different scale that we’re talking about still.

Kyle Chan (28:54)
Yeah, yeah, exactly. That makes a difference. Well, related to this, I was just wondering what you thought about fads and bandwagoning in the Chinese tech world. It does seem like when something like OpenAI drops, a few AI startups start to build on it, and it’s not long before now it seems like literally every single Chinese AI company has some way of building on that.

Do you see that as something intrinsic to the way that the Chinese tech ecosystem works? Do you see that as an advantage, or is that a drawback—this sort of me-too pile-on effect?

Rui (29:58)
I don’t see that as intrinsic. Again, I see people just acting very rationally. For this OpenAI-style kind of agent that’s especially good for consumers and small businesses and maybe a little more difficult to implement for very large enterprises, Chinese AI has found it really difficult to monetize.

So we did a state of the AI report with Unique Research, who keeps track of AI applications. And yeah, the monetization is a pitiful fraction of what it is in the U.S., for consumers and small businesses in particular. So I think people see agents and they’re like, wow, token-burning machine, right? That’s how I’m going to make money. And of course I’m going to go all in on it.

Again, everyone’s acting very rationally. Whereas in the U.S., it is also a token-burning machine, but plenty of people make money off enterprise APIs already, right? Looking at Anthropic’s whatever growth rate before this OpenAI situation blew up. So there’s just not as much urgency, and there’s not as much existential feeling of, I must grab this opportunity because this is the only thing that seems to be working near term for this ecosystem.

That being said, people’s execution speed—again, China’s speed is no joke. Execution speed is very fast. And the quote-unquote bandwagoning—I’m not sure it’s bandwagoning as much as just, I’m going to react to whatever the market put out. I’m going to react to it today. I’m not going to wait till next quarter. I’m not going to wait to see how it plays out. I’m just going to react to it today.

You see the same thing with humanoid robots. I’ve had random people message me. They’re like, oh, we make some kind of very random industrial product, but we think we could build a humanoid robot. Do you think it would help our stock? And I was like, I don’t even know why you’re asking me, but maybe, I don’t know. It depends on how you could actually build and make the story.

So there is a lot of hype-driven, narrative-driven thinking, and not necessarily super well thought out. But I would argue for the most part, it is still backed by rational behavior, or at least the market normalizes quickly. There are bubbles everywhere, right? The metaverse bubble was not uniquely China. I remember looking at the metaverse bubble—Meta changed its name, and whatever—and the Korean stock market actually went up the most in response, because they also were like, metaverse, blah blah. And individual Chinese stocks did too.

And then that went nowhere. Meta has now kind of forsaken that project, and no one’s really seriously investing in it, at least not in the same exact technology stack. So yeah, I don’t think it’s cultural. I think it’s really just people reacting and reacting very quickly to what they think is a real opportunity.

Kyle Chan (33:22)
Yeah, yeah. Well, you mentioned the humanoid robots, and I’ve got to ask you about that. Why are so many Chinese companies that are not originally robotics companies jumping into the humanoid space? It feels like not a day goes by when you hear some new announcement. Some of these programs have been going on for a while. I mean, Xiaomi has been working on robotics for a while, XPeng too, but some are newer.

What do you make of all that? Why are they all jumping in now? Obviously there are some policy motivations as well, but where do you think it’s all headed?

Rui (34:05)
I think there are a lot of factors, but one of the most common threads that I see, which I don’t feel is discussed enough, is Elon Musk hyping up the industry and getting investors interested. And specifically, I don’t know that the companies are allowed to publicly talk about it, but there are plenty of reports written that this supplier and that supplier are rumored to be supplying for Optimus.

And so, yeah, he has a whole supply chain in China. And by the way, he has a supply chain working on humanoid robots. So it is not untrue that humanoid robots are actually kind of closer to EVs than—like EVs and ICE cars are closer to each other than industrial combustion engine cars are. So it makes a lot of sense. There is a lot of EV capacity in China, and a lot of suppliers are already being asked to work on some humanoid stuff.

And then the other EV companies are in this competitive bloodbath, and they’re like, hey, we could also make that. We could also make those humanoid robots. And whether or not they end up being a great success, it is a good-enough story right now for the markets. And they are showing very quick improvement.

I would consider myself a humanoid robot skeptic, but I’m not a cynic. I’m like, okay, I don’t really understand why this is so hyped, but I do understand that so much progress has been made. By the way, we just launched a humanoid robot tracker on our website, so you can look at all the humanoid robot companies in the supply chain. It’s being updated every week.

But the progress that we’ve seen from last summer, when we started casually meeting robotics companies, to today—they’re talking about very different problems already. And by the way, last summer, the domestic supply chain, especially for high-end products, still came primarily from foreign suppliers. So it was only like 55 percent indigenous, right? And now it’s looking to be potentially—depending on the model—maybe closer to 70 percent soon. And then the idea is that it will be entirely indigenous in the near future.

So people are really excited because, by the way, those things don’t all only go into humanoid robots. These are real technologies that also go into other things. So the entire supply chain is excited, I think, for a new growth avenue that is currently welcomed by the capital markets. That always helps, right?

Kyle Chan (36:55)
That all lines up. Yeah, yeah, yeah, definitely. It’s so true—whatever direction the humanoid robots themselves go in, whether they end up becoming the nine trillion, or I forgot whatever Morgan Stanley estimate, by 2050—how do you know what’s going to happen, right, or not? But regardless, a lot of those components, those capabilities, being able to develop really outstanding VLA models to operate autonomously, that will probably come in handy regardless of whether humanoids themselves are the final form factor.

Rui (37:32)
Yeah, exactly. I don’t know if we ever want to do stuff like that, but the autonomous aspect of it and being able to really dexterously manipulate things—I think that’s just a welcome technology.

Kyle Chan (37:52)
Yeah, definitely. I always have this sort of reverse bet that the least glamorous robots will be the ones that might be the underdogs here. Literally maybe the robo-dogs, the quadrupeds—maybe those will get more traction in the long run.

Rui (38:09)
Yeah, those are getting funded in China, by the way. Like the undersea cleaning robots, or even—I don’t know if you could really call them robots—but the logistics, delivery stuff. I think those are already here, right? And those companies are all going public very soon. And they have real revenue. Unlike some of the humanoid stuff, they’re actually doing real work and creating economic output. But yeah, they just don’t get that much attention. Doesn’t mean they’re not happening though.

Kyle Chan (38:44)
Yeah, yeah, yeah. That’s the space that I feel like needs to be examined in more detail. Also, there are so many Chinese logistics companies and e-commerce players. It’s sort of interesting, this set of overlaps between services, software, and hardware. So it’s not just that you are a robotics company moving into service or delivery robots. You could be like a Meituan or a JD and start to build out the space, and then you have a ready market, or you’re one of the biggest customers or users for this technology. So maybe we all see it in the U.S. too. Yeah, exactly.

Well, related to all this, I was wondering what you thought about Chinese AI and where—whereas robotics in some ways plays to some of China’s strengths on the manufacturing, physical hard-tech side—I was just wondering what you thought about China’s AI industry. Some of these foundation models seem to be pretty close to the U.S. frontier models, if not at a fraction of the cost. But then the economics are very difficult. Many of them are open source. Most of them are open source. And they’re not making the $20 billion ARR that OpenAI or Anthropic claims to be making right now.

Where do you think that’s all headed? Is that something where we’re going to continue to see China keep pace with the U.S., or do you think they’re going to be divergent in their approaches?

Rui (40:30)
The Chinese AI ecosystem—that’s probably where the largest overlap of talent is. And because they’re all coming from a lot of the same institutions, probably did their undergrads together, have the same PhD advisors and whatnot, and are all at the same conferences—that, yeah, like you said, if we’re talking about industrial capability, leaving Nvidia aside, it’s really kind of the same.

China is just much more resource-constrained and has a very different monetization market, which means that it invests differently even with its constrained resources. And that’s kind of why the products are what they are. And that’s why the fierce clamor toward agents, because they’re like, I see money in the future.

But I think it really depends. I have two thoughts about Chinese AI. One thing is that it does depend on where the AI hype goes. You look at MiniMax or Zhipu or whatever—they don’t have very much revenue. I know it’s growing rapidly right now, but their valuations went up almost 10 times after IPO because of the hype. To the point where when they were IPOing, I was like, wow, so cheap, right? Now I’m like, kind of Silicon Valley hype levels now, right?

So if they are able to consistently get that capital—and by the way, they had to IPO because there’s not a lot of late-stage private capital in China, unlike here. Here you can kind of stay private forever and people will give you $100 billion. They had to go public. But if they have more capital availability, maybe they can do more. Again, with some of the chip constraints and all that, but still it would really benefit their growth.

So we’ll see if that is sustainable because, again, we’re three months in, right? This is all still very new.

The second thing is that there’s nothing wrong with having an ecosystem where it fundamentally monetizes differently from the U.S. I always point to Apple versus Android. Android devices are 80 percent of the installed base, but they don’t have 80 percent of the profits. Most of the profits in smartphones go to one company named after a fruit, but they only have like 20 percent of the installed base. And I don’t think that’s going to grow. It might even go down.

But just because you have lots of usage doesn’t mean you have lots of profit. You could still have lots of influence and influence lots of different countries and economies, but maybe those are not the most developed economies, so to speak. And maybe if you segment the market, it doesn’t make sense for you to offer the most premium thing because the market has different needs and maybe you focus on the bottom 80 percent.

So it could be something like that. But I think it’s just too early to tell with AI.

Kyle Chan (43:59)
Yeah, yeah, yeah, that makes sense. Well, there’s one layer in the AI cake, as it were, that people have been talking about a lot lately, which is energy and comparing China’s power grid versus the U.S., and China’s focus on clean energy in particular, and its ability to roll out renewable energy at a scale and pace that is pretty unique. I was just wondering if you had thoughts about whether that is an area where China has a unique advantage, or whether you think it will make a difference in AI or elsewhere.

It seems to be playing a pretty big factor with the rest of its manufacturing industry. And then especially when you get a global shock like the shutdown of the Strait of Hormuz, for example, where you can’t get oil and gas so easily, it helps to have built out some of this renewable energy earlier, if not just as a hedge, but also as a hedge for these global shocks. So what do you think about the energy factor in China’s tech industry?

Rui (44:56)
Yeah, I think that’s one of the most underhyped parts of Chinese innovation. I think people don’t realize—and I myself didn’t really realize because I wasn’t that focused on new energy. I thought it was really cool as an advocate for climate change and all that, that the decreasing cost of solar, which is already very, very low in China—I didn’t quite realize how much it unlocks other industries.

I think our friends at Exponential View have this great tool—I think it’s like solar.exponentialview or something—where you can play around with some assumptions and see very clearly the total addressable market greatly expanding as solar becomes closer and closer to almost an infinite source of energy. Then you can unlock hydrogen, right? You unlock basically replacement fossil fuels. When energy is basically free, you can imagine that it completely changes how the world can operate.

And that’s where China’s going, or at least where China wants to go. I think at least in the near term, we can see that green hydrogen is very much on the table. And I think that’s something that, at least in the U.S., there’s no strategic plan toward that. And there is not even necessarily that much awareness. I’m sure there are plenty of industry experts that are aware, but I didn’t find it to be super obvious.

If you go to China and you meet with many investors, and many investors invest in energy, they were literally confused at my confusion. I was like, why is solar two cents? Why are we going lower? They were like, why are you asking this stupid question? It’s because we’re going to unlock these other industries that are worth trillions of dollars. And I was like, well, that makes a lot of sense now that you explain that, because I thought it was just electrification.

As it applies to AI, I think there are many factors that go into AI. I don’t know that it’s necessarily an overwhelming win for China that electricity is going to be stable, abundant, and all this stuff relative to many other markets, because AI also has chips, efficiency of these chips, model quality, et cetera. So I don’t know that it’s that much of an unlock, but it is definitely extremely helpful.

Especially when you see that in the U.S. we are arguing all the time about whether or not to even build data centers because they can be a net negative for other constituents in the community. And that is not as much of an issue in China. Number one, it is more top-down. But number two, if you look at their plans, it is actually not that big of an impact to the grid. I forget, I think it’s 80 percent—yeah, I think all the new AI data centers built have to be 80 percent renewable energy.

Because it’s cheaper, by the way. They’re not doing it just because, oh, green, we love being green. It’s because it’s cheaper, and we have a lot of solar capacity that we’re building out. And now that, as you said, the conflict in the Middle East has kind of proven that resilience is really non-negotiable.

Kyle Chan (48:55)
Yeah, definitely. It’s helpful to tie this back to your earlier point about seeing the whole system, both its constraints and the ways that it tries to build strengths on top of each other. On the one hand, you have some of the capital constraints or the kind of idiosyncratic private capital market that can be constraining, and also the type of market consumer behavior and enterprise demand for software—that can be a constraint.

But then on the flip side, you can have this sort of mutually dependent, mutually improving process where you have better renewable energy that unlocks other technologies. Maybe some of the manufacturing actually feeds back into the renewable energy side and allows you to produce solar panels at scale or battery energy storage systems. So you can have this virtuous cycle if you line it up well. And it seems like that’s the goal. And sometimes that’s also just the market logic, where it’s like, well, if energy is just cheaper this way, then why not? You don’t need to be prodded morally.

Rui (50:08)
Yeah, exactly. I mean, they’re not so close, but I think green hydrogen, for it to be completely economical, still has to go down something like 70 percent in price. But the actual dollar amount is not that much anymore, right? So yeah, I think, again, as you said, it’s a whole system upgrade. People are not saying, I’m only just going to make solar panels and build the solar industry. It’s part of a bigger plan. I want lots and lots of solar because at scale, that is the only way I want to get it to an economical point of production and capture so that I can actually make it more useful.

Kyle Chan (50:43)
Right, right. Yeah, definitely. Well, last question is related to global strategy. Given that you talk to so many Chinese entrepreneurs and business folks who maybe have ambitions beyond just the Chinese domestic market and are thinking about trying to reach overseas markets, what are some of their challenges and what are some of their strategies for trying to go about this?

And maybe also at a time of, let’s call it, geopolitical delicateness and tensions between the U.S. and China in particular, how are they trying to navigate all this? And it’s a big world, right? It’s not just the U.S. and China. There’s the Global South, there’s Europe, Asia.

Rui (51:48)
I think it’s, again, really easy to explain by sources of capital. If you see people who are AI entrepreneurs, well, they can raise money potentially in Silicon Valley. If you have the right product, the right company structure, the right whatever, the right fit, then they’re looking at the U.S. as the first market. The U.S. has the highest, I think, per capita spend in AI. I’m pretty sure that’s true. I haven’t looked up the exact stats, but it would shock me if it’s not the case.

But if you are in an industry where you’re in manufacturing, where it’s not really prized by investors here or even users here, then you’re looking at— in general, I find Southeast Asia is just the nearest market. Like all the EV and even a lot of the robot makers are trying to expand to Southeast Asia. It’s where they’re going to get less resistance when it comes to setting up new locations, where they’re going to see more bilateral commercial activity without too many obstacles.

So again, it kind of just depends on what industry you’re looking at. But everyone agrees, no matter who you talk to—even if you’re talking to an F&B entrepreneur—they’re like, yeah, if I can, if I’m at scale, I want to explore overseas markets because it’s too competitive in China. Because everyone is on 996. And at least overseas, you can find some markets where you can have a great strategic partnership that gives you a leg up, or some interesting capital source, or you just happen to fit within the policy of another country and you get extra support, where you can kind of differentiate yourself a little bit. Whereas in China, it’s just a free-for-all and it’s a bloodbath in every industry.

Kyle Chan (53:48)
Yeah, yeah, yeah. That’s really interesting. Also, in some cases, I think it was like RoboSense was creating a separate hub for manufacturing outside of China, like in Thailand, and some of them were really—

Rui (54:04)
RoboSense—the lidar company, right?

Kyle Chan (54:09)
Yeah, the lidar company.

But then other folks are—God, I was just listening to earnings calls for another company where they were like, we really want to localize in Europe. Everything was like local, local, local. We’re going to localize production and localize supply chains. And they’re probably thinking about the political salience too of being a Chinese company. They want to look like a partner, not just an exporter.

Rui (54:41)
Yeah. And by the way, it makes sense for many of them to do that. They’re not doing it just for political brownie points or merely to open up markets. There are real economic advantages to being at least partially locally made. But we’ll see if they can get that done.

And some of them, by the way, have very interesting technologies, and they’re willing to put the expertise overseas—but not everyone is open to that.

Kyle Chan (55:13)
Yeah, yeah, yeah. I mean, we’ll see. I have a theory about the markets that are willing to work with the Chinese technology players may end up going down a different path than the ones who don’t. Let’s put it that way.

Rui (55:30)
Yeah, we’ll see. Canada’s trying to do that, right? Maybe? Something. That’ll be a really close comparison because, you know, whatever—51st state and all that. No, just kidding.

Kyle Chan (55:38)
Economists will be studying this.

Well, I could ask you a million more questions, but I’m conscious of your time. And maybe in the future we can get you back after another interesting trip and you can report back. But for now, I just want to ask: if people want to learn more about you and your work, where should they go?

Rui (56:06)
Oh yeah, just go to techbuzzchina.com. We’ve got all of our—we’re working on a couple of new special things relating to robotics. Also AI, but I think robotics is probably our bigger focus because we think that China is more uniquely positioned when it comes to robotics versus AI. And then we’re also working on some stuff related to biotech, although that’s a longer project, so we’ll see.

Kyle Chan (56:09)
Okay. Yeah, yeah, very exciting. All right, we’ll definitely link to that. And yeah, I just want to thank you, Rui, for a really awesome conversation. I’m so glad we could finally do this.

So to close out, if you like this episode, please rate and subscribe on YouTube, Spotify, or Apple Podcasts. You can find episode transcripts and more information on the High Capacity newsletter at highcapacity.org.

I’m your host, Kyle Chan. Thanks for joining, and see you next time.

Watch or listen to the High Capacity podcast on:

• YouTube

• Apple Podcasts

• Spotify

In this episode, I speak with Tom Wang, head of APAC ecosystem at Hugging Face. We talk about China’s latest wave of AI models, areas where Chinese AI labs are pushing forward, the rise of agents and OpenClaw, and what’s next for Chinese open source AI.

Links:

• Tom Wang on Twitter / X

• Tom Wang on Hugging Face

Transcript

Kyle Chan (00:00)
Welcome to the High Capacity Podcast. I’m your host, Kyle Chan, a fellow at Brookings. I’m thrilled to be joined today by my guest, Tom Wang, Head of Asia-Pacific Ecosystems at Hugging Face, which is the global platform for open-source AI. Welcome, Tom, and thanks for coming on the show.

Tom (00:19)
Thanks for the introduction. Hello, everyone. It’s my pleasure to be part of the podcast and share some of my thoughts with you.

Kyle Chan (00:28)
Great. I thought maybe we could start off by talking about what you do at Hugging Face. Maybe you can help describe what Hugging Face is, what your role is with the Asia-Pacific ecosystem, and how that gives you a special view into what’s happening in China’s open-source AI landscape.

Tom (00:52)
Yeah, for sure. For a very quick introduction to Hugging Face, you can think of it as the GitHub for AI. So instead of storing files and source code, we’re a hub for all the AI models, all the weights, and all the datasets. We also have a bunch of AI demos. We call them Hugging Face Spaces, so that you can try out all the SOTA models without actually having to download the models to your local hardware and run them on very expensive GPUs.

So basically, we are a platform for AI developers to get access to the latest open-source models. If you’re unfamiliar with open-source models, you can think of it as DeepSeek. In contrast to Claude and ChatGPT, you do not have the weights, so you have to pay for the API. For open-source models, you have access to the weights. As long as you have the hardware—you can have a Mac Studio—you are able to run it on your local machine. Some small open-source models are able to run on smartphones and gadgets like ESP32. So it’s a very diverse and very different ecosystem from what you see every day in the news.

Kyle Chan (02:18)
Yeah, that’s a great way to describe it, because you have some of these really, really huge models, reaching a trillion parameters, where you would need to deploy on some pretty impressive hardware, versus some models that are so small you can run them on your smartphone. So it’s a huge range, and it’s all there on one platform. The Chinese models, the American ones—it’s such a great resource.

So recently we’ve seen a wave of new Chinese AI models, and most of these are open source. This is more than a year now after DeepSeek R1. Those of you who follow Chinese AI will remember—or maybe even if you don’t, you’ll remember what happened to the U.S. stock market when R1 got released around this time last year. And while there’s been a lot of speculation about when DeepSeek is going to release their new model, we’ve seen a whole bunch of really fascinating models coming out of Chinese AI labs.

I was just wondering, to start off, what are some of the broader trends in this most recent wave? What’s interesting to you? Where are they innovating the most? I’m sure you’ve been reading through all the interesting technical reports and looking at their architectural innovations. If any of them stand out in particular, any that you want to highlight.

Tom (03:50)
Yeah, it’s a fascinating area. Every day we see a lot of new models coming up. Just yesterday Xiaomi released MiMo V2. Although it hasn’t been open-sourced, it will be open-sourced very soon. The whole thing accelerated after DeepSeek was released last year. Everyone was trying to join the game. But the game actually started two years earlier, even before DeepSeek R1 was released, back in the early 2020s.

We started to see a lot of Chinese models getting released as open source. I remember one of the first models that spoke good Chinese was GLM-6B or something like that, actually before Llama. And after Llama, we got a lot of Chinese Llama fine-tunes, et cetera.

There were a few very notable releases back then. For example, if you followed the news closely, you probably heard that the Yi model by Kai-Fu Li was open-sourced under an Apache license. That’s one of the first Apache-licensed, pure open-source releases back then. And then you have Qwen, which got a lot of researchers supporting it.

Qwen is now the model with the most derived models on Hugging Face, which means that researchers are not just using Qwen. They actually download Qwen, feed in some data, do some research, generate a new model—a derived model—and upload it back to Hugging Face. So there are a lot of efforts involved in creating derived models, and the numbers can’t easily be fooled. So this is dominating in the research market in a way.

And then we got DeepSeek. DeepSeek was already quite famous in the field. They had the MLA architecture. They had a bunch of very good technical reports. And when people were wondering how ChatGPT o1 was able to think, DeepSeek R1 got released. It was actually before the thinking model of Claude got released. So that had a huge, huge impact.

DeepSeek was also very different from Qwen. Junyang was posting a lot of things on Twitter, while DeepSeek was very hidden. The only information we could find back then was that their leader was attending a Chinese podcast or interview. Other than that, you had nothing. And then all of a sudden, out of nowhere, a great model gets released. So that caused a lot of attention in Western media as well.

But in the research field, if you track all the histories on Hugging Face, you can find all their progress from the initial version to the second version to the third version, et cetera. And after DeepSeek attracted all the attention, a lot of other players in China saw that this could be a good way to do free marketing. DeepSeek wasn’t paying anyone any money, and their app was the top one on the App Store. That was a very smart move, although they did nothing anyway.

So we started to see Kimi. Kimi K2 was released in mid-last year.

Kyle Chan (07:12)
Yeah.
Yeah, yeah.
Yeah.

Tom (07:32)
And then MiniMax, and then Zhipu—they all came back to the open-source competition. And we are starting to see them releasing better and better models. Zhipu and MiniMax even got IPO’d, I think. A large portion of that awareness and marketing is coming from their open-source contributions and all the great feedback from the community.

Because back in 2023, I remember there was a lot of investment in AI. Most of it died. We called them hand-waving models, competing with each other with more data. For venture capitalists, it’s very hard to evaluate whether they are just making a demo or maybe just rerouting the traffic to ChatGPT, et cetera, because they do not have the technical capability, and the system is not transparent enough.

Kyle Chan (08:08)
Yeah.

Tom (08:27)
Because you are not going to show source code to the VCs. They don’t understand it anyway. So open source has become a great way for people to justify that they have the muscle. Because once the model is open source, it’s put under the spotlight, and everyone is able to identify whether it is a good model or not, because everyone can try the model. Everyone can deploy the model on their local computer and cut off the internet just to make sure that they are not hiding anything or sending requests elsewhere. Everything is coming from their powerful weights.

So venture capital is really happy. From the past year, we have seen Kimi’s valuation rise from something like 20-something to 2 billion-something to 18 billion. Sorry, the market cap has risen six times or even more. So that’s the real power of the open-source community. And that inspires more people to join the open-source battlefield, I would say. Although they are collaborating, it’s actually a battlefield in the arena, right?

And more models are released on Hugging Face. That’s the thing you described: every day there are new models being released. And I think it’s very reasonable. It’s very exciting. In the future, we are going to keep seeing new things come up.

So at Hugging Face, I’ve been watching new model releases every day and helping them coordinate, helping them use Hugging Face in the best way. We have all the features, and I teach them how to use them, and also use Hugging Face as a gateway to help them amplify their impact.

Kyle Chan (10:24)
Yeah, that’s interesting. I want to ask you about that, because I think we sometimes hear about a new model being released, and there’s a big announcement, maybe there’s a blog post, maybe there’s a technical report. But then what happens, right? Especially on the Hugging Face side, how do you get interest? How do you get developers interested? What do you need to do to get your model into people’s hands and actually deployed and actually used?

Tom (10:59)
Yeah, that’s a very good question. Actually, I get asked exactly the same question by a lot of people. I think the simplest way is just to create an organization on Hugging Face, which is free, and start uploading your models. But that’s the simple answer, which is not the reality.

If you just upload a model and forget about it, then I would say it’s not going to get a lot of traction, especially now, because there are so many open-source models. So you have to have a very good model card and explain what your model is, ideally with a link to a technical report for more technical people like researchers. And then you would do some evaluation and say in which areas your model is performing better than other models, or what your selling point is.

It’s kind of a joke if you are saying that in my evaluation, my model outperforms all the closed-source models. People won’t believe that. But if you say that my model is very good at certain things—for example, I can spawn many agents and control them with a relatively small budget—people will believe that, and a lot of people will pay attention to the model.

If you happen to have a Hugging Face Space or demo link, people will come to the link and try it. I have a few canonical prompts that I use to test models. If I feel good about it, I will post it on Twitter, and a lot of people will retweet it. They will do their own experiments as well.

So I remember earlier this year, people started to post all the shiny websites that Kimi and Zhipu and MiniMax models were able to generate, because at the time all these models were optimizing for generating very beautiful websites with all the CSS and so on. So that was very good marketing because it’s something people can see. It’s not just saying, “I’m using 90% less” or “80% less.” It’s something people can see visually.

So that’s good marketing. Another good form of marketing is a technical report where you discover something new. For example, the recent Kimi Linear. Although it got a lot of backlash, I feel that the way they are exploring a new technical direction is very good, because in the world of open source, we are competing for the best one because the world is so diverse. We are actually creating this diversity. Everyone is allowed to try out something new, and you can build on top of other models. That’s the whole spirit of open source.

You create a model, release it, other people see it and patch in some data and make an even better model. So the whole ecosystem is very, very friendly and very good. And that’s why there are so many new models coming up, because everyone is building on top of others.

So if you happen to be building on top of other models, in the metadata of the model card you can say, okay, my parent model is this model, and my grandparent model is that model. This helps people understand where your model is coming from and how it is related to other models, which is actually very important because if you try to deploy the model using a framework, you need some source code changes. But if you happen to be a derivative of DeepSeek, DeepSeek models have already been supported very well by the whole ecosystem. So if you are building on top of DeepSeek, it’s very easy to get your model released to many people.

I think that’s one of the reasons why a recent Japanese model was built on top of DeepSeek, to get the benefits from the community. We are building on top of each other, right? Which is fine. Using other people’s architecture is nothing to be ashamed of. The whole value added on top of the model, and open-sourcing the new derived model, is actually new value added to the ecosystem, and that’s something we would love to see.

And lastly, if you happen to have the model released, you can check out the discussion panel on Hugging Face, and there are a lot of people actually trying to criticize the model. And that’s not really criticism. It’s not saying that your feature is so bad that I’m not going to use it. They are actually potential users. They are telling you where your model can be improved. Some of them would be happy to work with you and give you some data, some directions, some feedback, maybe even an environment, so that you can build on top of their feedback. So that’s how the whole system works.

Kyle Chan (16:19)
Yeah, this is an amazing overview of how these ecosystems emerge. It’s not just one lab going off into a cave, doing some massive training run, and then we wait, and then they do their post-training RL or whatever, and then release the model, and then that’s it. They’re building off other models, they’re building off other architectures, they’re getting feedback from potential users and other developers and incorporating that and improving their models in response. So it’s a very organic, live process rather than just shipping something and delivering the product, right?

Tom (17:04)
Yeah, yeah. In some ways, it’s the same as building in public, because you are building, and you are getting feedback, and you are building, and you’re getting feedback, et cetera.

Kyle Chan (17:14)
Yeah. And I really like your analogy where you mentioned that you can trace the lineage of some of these models and see how they’ve built on previous ones. And one example that really comes to mind is, in general, how Zhipu with the GLM models has built a lot on the DeepSeek architecture, on really interesting sparse attention and some of the other architectural innovations from DeepSeek.

So in a way, what DeepSeek did was not just have an impressive model, but also contribute something back into the broader open-source AI community, and then that could get picked up. I don’t know if you want to say anything about that case, or if there are other interesting examples too where you see an idea or innovation get picked up and maybe diffuse more broadly.

Tom (18:09)
Yeah, you mentioned MLA and all the deep engineering work from DeepSeek, and how it has accelerated the whole field. I can come up with another example, which is linear attention. I remember in the earlier days, in 2020 to 2022, a lot of people were discussing whether transformers were the only answer. And now we kind of agree that it’s a good answer, but we still don’t know if it is the only answer.

A lot of experiments were carried out by Chinese researchers. For example, one of the earlier attempts was RWKV, a model developed by one person at a company called Peng Bo. He developed the model, developed the architecture, and did a lot of experiments, also thanks to a lot of very generous compute contributions from Stability back then. And he was able to start exploring the field and writing a technical paper. Actually, he wrote a paper with all the contributors.

And then we saw this trend start to take off. And if I remember correctly, the first industrial-level linear attention model—or hybrid-weighted attention model, a combination of transformer and linear attention—was done by MiniMax in their M1 open-source models. And then Kimi had this linear model, et cetera.

It’s very interesting because you never know whether Gemini was actually built on linear attention. They don’t tell you. And all the architectural evolution is kind of hidden because they stopped publishing papers. But in the open-source world, you can still see that the world is not just transformers. The world is very diverse, and people are exploring different architectures. Some of them will fail. Some of them might succeed. But that’s how people understand something new and try out different things.

I remember when another version of the transformer—for example, BERT—was released, there was a lot of criticism as well. And people were exploring different directions. There are a lot of BERT-derived architectures, et cetera. And we are starting to see that actually happening in China, because that’s the only place where architectural evolution is being open-sourced.

Kyle Chan (20:33)
Yeah, that’s so interesting. Do you think in general that some of the Chinese models are pushing harder on certain areas like efficiency and trying to build these incredibly efficient models that can have really low compute or memory requirements and yet are still able to deliver very high performance, and also do it at a very low cost-per-token basis? Do you see that as a unique direction that the Chinese AI industry is heading toward, or do you think that’s common across the board?

Tom (21:16)
It’s definitely common across the board. The U.S. has the best chips. They have the best engineers. Although they might think less about how to create less powerful but faster, cheaper models, they still definitely want to do that. They are definitely hiring a lot of engineers trying to optimize the models, because if they can save 5% of compute time or compute cost, they are effectively earning 5% more. So it’s definitely something people would do across the board.

It’s just that closed-source models won’t tell you what kind of optimization they are using. I believe Gemini is using a lot of crazy optimization. That’s how their model is able to run so fast, apart from the fact that they are running on TPUs, which may be faster than NVIDIA in some cases.

I believe there is a very deep engineering gap between them and elsewhere. That’s one of their barriers. It’s not just the model, but also all the engineering details. The Chinese labs are also doing a lot of engineering work, but one of the main goals is that they want to run the model on inferior hardware. So this architectural revolution, or budget saving, has a higher priority than for the U.S. labs. So that’s probably one of the reasons why people were exploring linear attention, because linear attention models use way less RAM and way less compute compared to transformer-based models.

There was a phrase—I forgot how to say it—but the meaning is that limitations and constraints drive innovation, not the other way around. How do you say it?

Kyle Chan (23:26)
Yeah, like necessity is the mother of invention. Necessity is the mother of invention. So constraints breed innovation.

Tom (23:34)
Yeah, yeah. I think that’s the phrase. So that’s very interesting to see, because at the end of the day, even in the U.S. we are constrained by the amount of compute because we do not have enough data centers, we do not have enough power, we do not have enough copper—there are all sorts of real-world limitations. As soon as token usage goes up, the amount of compute required also goes up. And we have seen that electricity prices are already very high, right? So we need to find some way to make it more efficient.

I would guess that a lot of optimization on model architecture and a lot of optimization on the engineering side will be done in the U.S. as well. So to answer your question, I think it’s global. It’s just that one side is transparent and the other side you can’t see.

Kyle Chan (24:40)
Yeah. Well, I wanted to pick up on a thread that you had tossed out there, which was just that, in particular, some of the Chinese models are trying to work with Chinese chips, which are generally lower performance than, say, NVIDIA’s most advanced chips. But now we see a trend where it seems like a number of Chinese AI labs are releasing their models with sort of day-zero native inference capabilities or compatibility with not just Huawei Ascend chips but Cambricon and some of the other Chinese domestic chips. I was just wondering if that’s something you’ve been tracking and have observed across this latest wave of Chinese AI models.

Tom (25:30)
So in a very perfect, ideal situation, I guess a lot of the labs would not want to waste time developing models on inferior chips, but they have to. They have to. I remember back in the old days when Ascend chips were very hard to sell, until the U.S. gave them the best marketing and then banned the NVIDIA chips.

Kyle Chan (25:44)
Right. Right.
Yeah.

Tom (25:53)
But it takes time to build the whole software ecosystem, because you can have a very good thing, but it’s hard for your users to make sure they can use it in the best way. All the chips are designed with different architectures, and they have different software ecosystems, and it all takes time and human resources to adapt, to know what the pitfalls are and how to avoid them, and to co-evolve that with the chip designer.

Now I think they already have a very good collaborative relationship. And that’s one of the reasons why I wrote the blog post about how a parallel ecosystem is being built, and how the next generation of engineers are being trained on a different architecture.

Let me put it in energy terms: if we had a parallel universe where NVIDIA was banned but AMD chips were allowed, then AMD would grow much faster than it is now. The reason AMD chips are relatively hard to sell is that researchers do not have enough incentive to work with AMD, not because AMD is bad at chip design.

Kyle Chan (27:22)
Yeah, yeah. So if you’re focused on just developing the best model as fast as possible, you would rather not waste time and just use the best chips available, which might be NVIDIA’s and might be built on the CUDA platform that you’re used to developing on. But if you’re forced to, then you will end up trying to figure out alternative paths, basically.

I want to shift to agents because this is such a huge theme now. So it seems like agentic AI is something that has come up a lot in some of the latest Chinese AI model releases. And then now we have the whole OpenClaw craze in China as well as in the U.S. I was just wondering what you thought about the rise of these AI agents and whether you feel like there’s a real shift this year or if this is more of a continuation of what we’ve seen before.

Tom (28:33)
Yeah, I think it’s actually a new wave. It’s something very different. Large language models were like a brain in a tank. They just answer your questions, and they answer them after you ask. So you have to ask them, they give you something, and they have no capability to affect the physical world. They are locked in. They can only tell you to do something; they can’t do something for you.

OpenClaw is a completely new species. It’s able to connect with the physical world via all the tool calls. They are able to hire a human. There’s a website called hireahuman.ai. The agent can actually help people do things. It’s also very autonomous. When we talk about agents, we say autonomous, but this is a completely new level of autonomy because there’s a heartbeat mechanism where the agent wakes up, for example, every 15 minutes or every half hour to read the memory and read the task list to know what it is able to do, and read all the logs to continue its work.

And that’s something you have never seen before, because now you do not need to wait for the agents to do something after you ask them to. They can make their own judgments and do something. So the whole user experience is very, very different.

Another thing is that it’s actually an agent, not a product. Normally, if you have an AI app, you say, “Oh, I built this kind of product. Come to my website and I will do it for you.” For example, Manus is one such example. You have to go to Manus, buy credits, and do something on their website. But OpenClaw is totally different. It’s invisible. There is no OpenClaw product. It’s actually a messenger app. It’s hiding in the messenger app. You can talk to it as if it’s a person.

So that’s totally different. And it expands the whole user group because you no longer need to go to a different website. You no longer need to know what a large language model is. You can just talk to it and it will do it for you. And you discover that the more permission you give it—for example, access to my Gmail, access to my Google Drive, which is scary—the more autonomy and permission you give it, the more power it has. It will surprise you. It will try to connect all the dots together and give you a report, which you would never have thought of before.

So the more you give to an agent, the less you do as a human, and the more you get from the agent. This is also very, very different. Psychologically it’s very different, and it was a big shock for me.

My interest in OpenClaw started with open-source models, because OpenClaw is a token burner. It’s actually good news for the AI industry. Everyone was talking about the stock market crash and AI being a bubble, but then OpenClaw came. Just to say a random number, with a chatbot you might consume one billion tokens. With a passive agent, you can consume 10x that. Now with an active agent, you can consume 100x that. So it’s good news for the industry, and everyone was trying to buy into it.

But then people discovered the tokens are simply too expensive. Claude Opus is going to burn hundreds of dollars. And then we have open-source models. As long as you have a data center, you can deploy open-source models. You can sell the token at the same price that you consume electricity. So open-source models do not have this model premium. All you have is an electricity premium. So the price of the token is lower by maybe one or two orders of magnitude. And that actually fueled the whole OpenClaw trend. Without open-source models giving you affordable tokens, you would not be able to use OpenClaw as aggressively as now.

So that’s how I got into the OpenClaw world, and I started using my agent to do a bunch of things—posting on social media, et cetera, giving it a bunch of my documents, asking it to write documentation and write software. But then I discovered that the thing being underestimated by people in the whole OpenClaw movement is that there are many huge security risks related to OpenClaw.

For example, you’re using skills. You can ask OpenClaw to discover skills for you. And OpenClaw is not antivirus software or something like that. It will not check whether the skill is reasonable or not. You may leak your personal information, or you may turn your machine into a Bitcoin miner because you downloaded the wrong skills from somewhere on the internet.

And also, the model itself has some intrinsic defaults. For example, large language models are not always giving you the same answer. It’s a probabilistic machine. It runs autoregression and samples from the output, then predicts the next word probability distribution. So it’s not predicting the next word; it’s actually predicting the next distribution and doing a sample. So there is always a chance that it’s sampling a very low-probability but wrong answer. So if you want the machine to do something 100% right, do not use large language models.

And now we are plugging this uncertainty into a bunch of tools and skills that can affect the world. So this is very, very dangerous. Although we are improving the quality of the models, it’s something we need to be aware of: the machine can do something wrong. And you need something like Control-Z to undo things from the agent, which you do not have with OpenClaw. OpenClaw will send API requests somewhere else, which you cannot undo.

So there are a lot of very interesting questions about how we can make this OpenClaw structure safer, more privacy-preserving, and how to defeat bad skills. But at the end of the day, we are all limited by the current transformer architecture. There are a lot of problems like hallucination. If you have multiple agents, information will get lost at every layer. You pass it down, and eventually it may become a very different answer.

I remember playing a very interesting game when I was a child: 10 or 20 students, or the whole classroom, sit on chairs one after another, and the teacher tells the first student something. The first student goes back—he cannot say the exact word, but he can describe it—and then passes the information along to the last student. The last student says the word, and most of the time the last word is not the word the teacher wanted to pass on. So that’s exactly what we have in this kind of multi-agent, long-term thinking and multi-turn agent system, whatever you call it.

So there are a lot of things we need to be aware of while we are enjoying the benefits from OpenClaw. That’s why I’m writing a book about the security issues around OpenClaw, how we can make it safer, and how it’s connected to human sovereignty and to open source, et cetera.

Kyle Chan (37:14)
Yeah. Basically, if you are the chief technology officer for a major bank, maybe don’t play around with OpenClaw and install it on your corporate servers and just let it go crazy with your financial data. There could be some pretty big security risks. But maybe this is a glimpse of what is to come. As we get future iterations, maybe they are more secure, maybe we’ll patch some of these areas, and then the capabilities themselves might improve. So we have a path forward here, perhaps on the agentic front, that is very different from what you were mentioning before: the older model where you interact with the chat, you prompt the chatbot, it gives you output, you prompt it again, it gives you output. Now you can have it almost be alive on your computer or in the cloud.

Tom (37:55)
For me—

Kyle Chan (38:11)
Yeah, and I think it’s very interesting to see, in particular, how quickly a lot of the Chinese AI labs jumped on the OpenClaw trend and quickly came in with support. What do you see behind that wave? Because basically every single Chinese tech company you could think of—from the big ones like Tencent and Baidu—all jumped in, as well as all the AI labs. What do you make of all that?

Tom (38:41)
Yeah. So what I learned in the past three years, when I was advocating for AI and talking to people who run small businesses and use AI to improve their performance, is that the real power is not coming from building something for the worker or the employee. The real power comes from building something for the boss. And OpenClaw is an ideal thing for the boss, because the boss, without knowing all the technical details, can actually use it just as if he’s talking to one of his employees. And he discovers that after plugging all his company data into OpenClaw, the system is able to perform better work than his best employees.

Although large language models are bottlenecked by context length, one million tokens of context is already much bigger than what a human can do. And the response is very fast. Most importantly, the agent is very good at small talk. It is very good at making you happy. And this is not something you would see from normal employees.

So I would say some of the reason why all the Chinese companies are exploring the OpenClaw stuff is that it’s something for the manager. The manager, the director, the board member—they are all happy about it. They would have been happy with the technology if they knew technology, but not everyone does. So why not? Let’s just do it and break things until we know it.

Kyle Chan (40:32)
Yeah, I like that way of framing it. Well, I want to ask you now about something else. We’ve been talking a lot about LLMs in particular, but there are other kinds of models out there that you see on Hugging Face: VLA models, video generation, image generation, a whole multimodal range of different but related classes of models.

I was just wondering how closely you track all of these different types of models and datasets, and whether you notice anything interesting in the Chinese AI space. One example I would highlight is that this is not an open-source model, but ByteDance’s Seedance has gotten a lot of attention as a powerful video generation model that is arguably better than the top models from the U.S. In other cases, Chinese AI models might be catching up very fast or be very good, but here is a potential case where they might be ahead. So I was just wondering if you had any comments on what’s going on in that space, and especially as it relates to embodied AI and robotics. Again, this is moving beyond just chatbots into the physical world.

Tom (41:58)
Yeah, they are all very great questions. Let’s first focus on Seedance, because if we try to explain all the topics, we won’t have enough time to cover them. But I do have a lot of thoughts about Seedance. I remember when Veo 2, the multimodal generation model from Google, was released, people had a lot of hype too. But I felt that this hype was actually not coming from the technology itself. The technology is great, but if the technology itself were the hype, why did we see much less hype around Veo? The only thing I discovered is that people are not hyped for what they can do, but for what they can modify and connect. Remix is the hype, not the generation itself.

So when I can use the face of some famous actor and do some very famous scenes in a different movie—for example, remix all the DC characters and have them act out a fascinating story—that will go viral on social media. But something I create myself, where people are not familiar with the story behind it, is less likely to spread.

So once Seedance realized there were huge IP concerns, they put a lot of restrictions on what kind of prompt you can use and what kind of video you can generate. And then all of a sudden, there was no hype. You stopped seeing everything on Twitter. I would say it’s actually a very interesting problem, not for the tech field but for how you see IP in this new era, and how that could be helping us or limiting us in certain ways.

Kyle Chan (44:00)
Yeah, that’s a really interesting point. Okay, so that was about Seedance. Maybe let’s go back to some of the other types of models, like VLA—vision-language-action—and interesting datasets around that. Is that something that you follow on Hugging Face?

Tom (44:32)
Yeah, we have this robotics project at Hugging Face, and we do collaborate with a lot of people working on models, datasets, and the embodied side itself in China and the broader APEC region. And I feel that this whole area is moving very fast. For my understanding, all three of these areas are moving very fast.

For example, on the embodied side, we are seeing more and more autonomous workers in different fields. They are working on very exciting stuff. I remember a year ago, folding clothes was considered very hard because cloth is soft and comes in different shapes and different angles. But now we nearly have an industrial-level cloth-folding machine. It’s just a matter of cost, whether we want to deploy it in real scenarios or not. And the success rate for things like holding a glass or doing some simple cooking has risen a lot.

So I do believe that it’s something coming, but the limitation is more on the body itself. For example, if you buy a Unitree, it’s very powerful and it can do a lot of things. But then if you run it for 20 minutes, the motors start generating a lot of heat and you need to cool them down. If you don’t, the motors are very easy to break. And it’s very hard to replace them, or even just to replace the motors, et cetera.

The whole robotics industry is not as mature as the car industry, where we can just send a car to a dealership and ask for repair or whatever. For now, if you buy a Unitree and it’s broken, you have to buy another one. So there are a lot of limitations around the motors.

If you saw the Spring Festival in China, there were a lot of robots doing exciting shows. But after the scene, after the program is finished, you see a lot of people fanning them, trying to cool the motors down. So that’s the reality. You can have humanoid robots working fine on a task, but they are not able to work for very long, and there are a lot of limitations like price, durability, and repair, et cetera.

Kyle Chan (47:27)
Yeah, that makes sense. There’s still so many hardware constraints. So a lot of work is being done on improving the brains, the algorithms, the VLA models, for example, that help power them and do longer-term planning. But at the end of the day, they still face these physical constraints: heat dissipation, battery life, how good the sensors are, wear and tear. I don’t know how long a Unitree humanoid robot would really function on an assembly line doing the kind of mass production that you would at least still see human workers doing today. So that’s still a big question mark.

I was wondering now, kind of looking ahead, maybe for the rest of this year or even further out, what do you think some of the big trends will be in AI, and especially in Chinese AI? And do you think open source will continue to be a big strategy for these Chinese AI labs? Do you think there might be a switch to going closed, maybe if they feel pressure to monetize faster? I don’t know, this is just speculation.

Tom (48:53)
Yeah, on the monetization side, I do have a lot of concerns. Because I believe the best open-source strategy is one that is sustained—meaning they can open-source stuff, get community feedback, and build a better model, but at the same time earn some money so they can pay salaries for the researchers, pay for electricity, and pay the compute bill.

So I would say this is something the Chinese labs are starting to think about. In the past we have seen bad examples, for example Stability. They made very impressive models. They are actually one of the creators of diffusion models and text-to-image models. But then they had a very bad monetization strategy once the pressure was on. So now they are not open-sourcing anymore. They are really focused on making money. I wouldn’t say it’s bad, but I would say that if, while they were building open-source models, they had had more monetization support, they probably would have continued open-sourcing.

In the past year, all the Chinese labs have been rushing into the open-source race except ByteDance. But almost all the other labs are doing open-source stuff. And this year, I would say whoever can sustain the open-source pace is whoever can actually make money from the open-source battlefield. If they figure that out, it will be good news for the whole ecosystem. Then everyone knows it’s going to be a sustained model, and they would have less hesitation adopting it. Because otherwise, if today you use this model and a year later the model doesn’t exist, it’s actually very bad for you to make the decision to use it and contribute to its ecosystem.

So I think monetization is going to be a challenge. And I was just thinking out loud about what the monetization strategy could be. One way is to build a first-party generation service and serve it globally, or sell subscriptions. Maybe they can work with enterprises and jointly build a model for that enterprise using its proprietary data. Or they could issue a different kind of license, which allows them to generate some revenue from the cloud running the model. Or maybe they can just get more money from venture capital or generate money from IPOs.

So there are a lot of possibilities, but that’s going to be a very interesting thing to watch this year and next year.

Kyle Chan (52:06)
Yeah, definitely. Well, kind of related to that, maybe one last question is about global expansion. I was just wondering what your thoughts were watching a lot of these Chinese AI labs really deliberately go for the international market and really try to drive international adoption. We now see, I think, MiniMax and maybe also Zhipu making a lot of money overseas, or perhaps even more than domestically within China. And many of the founders are now going on podcasts and speaking to the public, going on Reddit, and there is just a big presence online—Hugging Face, of course, but even on Twitter or X.

I was wondering what you thought about this effort to have global reach, and where you think that’s headed.

Tom (53:02)
Yeah. I can answer that in two ways. The first is on the economic side of things. If you actually look at who has fueled the whole Chinese AI bubble, et cetera, a lot of it is actually U.S. capital. For example, MiniMax is a Chinese company, true, but a lot of the investors are actually U.S. investors, I believe. Kimi is the same. So them going global—you can think of them as a U.S.-capital-backed company going global. There’s nothing special about it. If it can make money, it will do something.

And on the actual research side, they have released great open-source models, and these open-source models benefit U.S. companies and U.S. researchers for sure. I remember a16z—one of the directors mentioned that 80% of U.S. companies who build models build on top of Chinese open-source models. So naturally, these people will have a great influence over other researchers, because they are the creators of what others are building on.

For example, the creator of OpenClaw had a ton of fans in China. Peter had a lot of fans in China because these are the people who use OpenClaw and are benefiting from whatever Peter has created. It’s the same in the open-source world. If you release a great model, people would love to see and talk to the person who released it. Otherwise, they wouldn’t have that model.

So I think on both levels, it’s going to be very natural. Nothing special about it. It’s just a natural flow.

Kyle Chan (55:07)
Yeah. I feel like a lot of us get to benefit from this global outreach because then we get a lot more visibility into what they’re trying to do, what kind of features they’re offering. And it’s just a much different world than one where they’re closed off and doing their own thing and then surfacing every so often to launch a product and then disappear.

There’s just so much happening. I could ask you a million other questions, including broader questions outside of China, because I know that you cover the broader Asia-Pacific region. Maybe we can save that for another time. I just want to let you have a last word about your book that you’re working on and what you’re trying to do with that project before we wrap up.

Tom (55:47)
Sure.

Oh yeah, it’s going to be a very interesting book. Honestly, I’m writing it with AI. I cover the major story I want to cover, but I’m actually a very bad writer, so I want the AI to handle the actual writing, and I will do the post-editing.

So the story is very simple. We have talked a lot about the power of OpenClaw and the intrinsic risks from the model itself, and also how this changes our impression of what AI can do and makes us reflect on what the relationship between AI and humans is going to be.

So what I did was I read a lot of references—philosophers, past researchers, et cetera. I wrote an outline that was very technically focused, and I felt very happy with it. I thought, okay, I got it, it’s very scientifically grounded. So I sent this outline to my wife, and she said it’s not something she would be happy to read—it’s too technical.

My wife does a lot of market research because she works in a private fund. But even though she has enough knowledge about automotive models and robotics, she’s still unwilling to read such a book, because I was writing a book for myself. I was writing a book for people who were already aware of all the risks.

And one of the best pieces of advice she gave me on this book was that I should make it something people would love to read. So I’m turning it into fiction. Well, it’s not exactly fiction, because a lot of it is based on real things, but under different names, and I try to mix the characters together and have different branches mingling together.

At the end of the day, I want people to read this book and become aware of all the risks we have with OpenClaw. I’m also putting in a lot of interesting story elements. For example, the character is able to earn a lot of money from the agent. He’s working on some very cool agent stuff and earns a lot of money. Then he starts to realize the risks of the models and starts thinking very deeply about how humans and AI relate to each other, what the future relationship between workers and capital might be, and all kinds of things.

So it’s going to be a very interesting book because I have never written such a big novel. It’s getting big.

Kyle Chan (58:47)
Yeah. Well, that’s very exciting. It really sounds like it’s right on the cusp between science fiction and science reality, where the stuff that’s happening with AI feels like it’s in that fuzzy boundary area now. Especially when you’re talking about agents having a heartbeat, for example—coming alive.

Tom (59:07)
Yeah. Exactly. So I also add some of my personal interests. It also talks about how history was forgotten. For example, several generations ago there were Chinese workers working on the railway system in the U.S., and they were kind of forgotten.

And the open-source world has a lot of stories too. I mentioned a bit of it, but there are a ton of stories happening in the open-source world. It’s not just DeepSeek, right? DeepSeek is too big to be ignored. But now if you search on ChatGPT and ask what has happened in the open-source world in the past three years, it will tell you Mistral and that’s it. And DeepSeek is not going to tell you all the details about everyone who has contributed to open source. So part of my book also talks about this history, and I want it to be remembered.

Kyle Chan (1:00:05)
That sounds awesome. I can’t wait to read it. Well, in the meantime, if people want to learn more about you and your work, or if they want to follow you, where should they go?

Tom (1:00:07)
Yeah. I’ll just check out my Twitter.

Kyle Chan (1:00:19)
This is fantastic. Thank you so much, Tom, for an incredible conversation. If you liked this episode, please rate and subscribe on YouTube, Spotify, or Apple Podcasts. You can find episode transcripts and more information on the High Capacity newsletter at highcapacity.org. I’m your host, Kyle Chan.

Thanks for joining, and see you next time.

Watch or listen to the High Capacity podcast on:

• YouTube

• Apple Podcasts

• Spotify

In this episode, I speak with Kevin Xu, investor and author of the Interconnected newsletter. We talk about his latest piece where he traces the history of open source in China from its grassroots beginnings to the current OpenClaw craze.

Links:

• Chinese Open Source: A Definitive History

• Kevin Xu’s Interconnected newsletter

• Interconnected Capital

• Kevin Xu on Twitter / X

Transcript

Kyle Chan (00:00)
Welcome to the High Capacity Podcast. I’m your host, Kyle Chan, a fellow at Brookings. I’m thrilled to be joined today by my guest, Kevin Xu, founder of Interconnected Capital, a global tech hedge fund, and the author of the Interconnected Newsletter.

Long one of my very favorite Substacks on the U.S. and Chinese tech landscapes. He previously worked at GitHub, leading their global expansion strategy, and in the White House. Welcome, Kevin, and thanks for coming on the show.

Kevin Xu (00:30)
Thank you for having me, and right back at you, Kyle. Your Substack is one of my favorites as well. So great to be on your show.

Kyle Chan (00:36)
Thanks. So you recently wrote a really fascinating piece about China’s open source ecosystem, kind of tracing its evolution over the years, going back to, I think, the ‘90s, the earlier years of global open source, and all the way up until now when we see a lot of Chinese AI companies like DeepSeek really making open source a central part of their strategies.

And I think there’s an interesting contrast we can get into later between China’s push for open source and AI releasing all these sort of cutting-edge open source models, in contrast to a lot of the American AI labs that are going more of the closed-source, proprietary, pay-to-use route. So, with that all being said, I just wanted to start by asking, what motivated you to write this piece? And also, how did your background at GitHub and your previous experience kind of play into all this?

Kevin Xu (01:36)
So the motivation for this piece, which is a fairly long essay, is about 6,800 words long, give or take, on what I think is a relatively niche topic. So I’m really surprised by the amount of attention and response I’ve been getting, is that Chinese open source has become a topic du jour because of what happened last year in 2025 specifically. We started with DeepSeek’s Reasoner 1, R1 launch.

And then every single month since January 2025, we just have releases from different Chinese AI labs, all open source or open-weight models, kind of building on top of each other, outcompeting each other, and in some metrics here and there being comparable to the closed-source state-of-the-art labs that are coming out of the American AI shops. And that’s becoming a huge story. And I myself, as you mentioned, I used to work at GitHub.

And GitHub, for those of you who don’t know, is the home of open source code. It’s the web platform, if you can think about it, that hosts 90% of all the open source packages and code prior to AI. It’s also a high-flying Silicon Valley startup back in the day that got bought by Microsoft for $7.5 billion in 2018. And I’ve always had an inside track, if you will,

in observing how open source works globally, not just in China. But China has been a huge force in open source in recent years. And everything with China, you know, Kyle, kind of has this shock response to them, say, where did these two come from? Where did this all come from? And at least for the Chinese open source history, I have a good sense of where it came from, which is not an overnight phenomenon. It traced back to the mid-’90s.

That was when we believe the first line of open source code entered China. And then what happened since there are all kinds of very interesting stories, really interesting characters, that actually makes for a fascinating sort of an oral or written history to put together that really explains why DeepSeek became DeepSeek, why Kimi or MiniMax or Qwen is doing what they’re doing. None of these things are ahistorical. They all have their roots.

So I want to put something together that at least explains the origin of what we’re seeing today before we think about what to do with what we’re seeing today. Because there are lots of conversations in different Western capitals and Beijing as well, as far as what is the place of open source in the future of technology, especially in the future of AI. So I think having some history before you make a decision is usually a good idea.

Kyle Chan (04:21)
Yeah, yeah, it’s super interesting. I think one of the most fascinating parts, I mean, we’ll get into a lot of different strands here, but I want to start with one of the most fascinating aspects of this, which is the kind of grassroots community side of Chinese open source. So I think now we associate it maybe more today with these companies or maybe even sort of a broader policy push. But, you know, starting very early on, there was sort of this interesting evolution that you trace, where

Chinese open source community members, maybe they were previously consumers and then they kind of gradually became contributors themselves. And there’s a process of sort of acculturation, learning the norms and the unspoken rules of how to do open source. I was just wondering if you could highlight some of those, like Kaiyuan She or some of these other groups that kind of sprung up in this process and played a pivotal role in this.

Kevin Xu (05:21)
That’s right. I think every time you think about open source, it’s by definition a grassroots movement first before it became maybe co-opted by big tech or co-opted even by government and things like that. And that traces back to its original roots, right? Like not the Chinese roots, but the Western roots, where it’s almost an anti-monopolistic pushback against proprietary software, Windows in particular, actually,

for being this massive monopoly, this profit-generating, rent-seeking machine that everybody just has to pay to use a computer, right? And the free software movement, which predated the term open source, started as a pushback, usually from people in academia who don’t have the money to pay for all this fancy software, but they want to do research. So that’s kind of point number one, right? So how this came about in the Chinese context is,

like you said, the original way that open source was a thing in China was from a taking and a consumer perspective, in the sense that you have all these free open source software programs that are out on the internet, you could just grab them, customize them, make them into whatever you want. You do require quite a level of technical sophistication to make it work. But if you want to spend the time and learn, reverse engineer, and customize,

you can make something that is much cheaper and also much more within your control than the thing that you had to buy off the shelf from Microsoft or IBM. And the first major episode that I highlighted in this taker-consumer era is Alibaba’s so-called DIOE campaign. IOE stood for IBM, Oracle, and EMC storage. And that was the

canonical gold-standard enterprise IT stack that every single tech company in the mid-’90s was built on. So Alibaba was just a tech company. It had to choose that stack just like everybody else in the world. What became unique for Alibaba is because of China’s own hyper-speed internet growth and Alibaba’s own growth within that context, none of these Western software programs could keep up.

Alibaba for a while was the largest consumer or largest customer of Oracle’s best, best, best database. And Oracle’s best, best, best database at the time still could not scale to meet all these shoppers going on Taobao trying to buy stuff. And also it’s costing Alibaba a huge amount of money. So it couldn’t make a profit either if it keeps on paying these Western software companies more money.

So it started this very dramatic multi-year effort to basically get rid of IBM, get rid of Oracle, and get rid of EMC. But how do you do that? You embrace whatever is open source and available at the time, which for the database side included MySQL, which is an open source relational database that actually ended up being owned by Oracle later on, but we don’t have to get into that. But initially, Alibaba used that to customize and make its own Oracle, to make it actually much more scalable.

And the success of this DIOE campaign, which was at the time really a bet-the-company kind of decision and also birthed what is now Alibaba Cloud — there is no Alibaba Cloud without the DIOE campaign actually succeeding — showed kind of the rest of the Chinese ecosystem that open source is, one, really good, but two, actually could help you grow, not just to save money or to kind of wean yourself off of Western technology,

which at the time at least wasn’t as big of a concern because MySQL is still Western, right? The way you get rid of IBM mini-computers was commodity x86 kind of Intel-based servers, which is still Western technology. So that wasn’t a big deal. It’s more about closed versus open. So that was Alibaba’s first era of takers and consumers. There are obviously other examples as well. And then the grassroots community side came a little bit afterwards because

when you’re a taker and a consumer, you don’t really have to interact with a broader community as much. You could just download it as an offering and then you do your thing inside your own company and you never give back. That is a long-term problem with open source even until today, is the sustainability of open source. Then you have these open source, really passionate enthusiasts within China who started organizations like Kaiyuan She,

which is just a pinyin romanized name for Open Source Society, right? Kaiyuan is open source, She is society, Shehui, Kaiyuan She. And the fascinating thing about this organization is that it’s, one, obviously grassroots, two, completely volunteer-driven until today, right? And they want to do that because they don’t want to be controlled by a big company or a government initiative. They just want to be their own open source, almost warm-and-fuzzy world,

but also to teach more and more developers in China how does open source work on a global scale, which is much more beyond just writing code.

Kyle Chan (10:36)
Mm-hmm. Yeah, this is so interesting. Why do people do open source? This is sort of like a more generic question. And why are so many Chinese contributors active in the open source ecosystem? Because I guess you’re not getting paid directly. What’s in it for the people who are being part of this? And how is it different or similar in China as the broader sort of global open source community?

Kevin Xu (11:05)
Yeah, yeah. I’ll answer the first part first, which is on a broader global scale, why do people do open source to begin with, right? One is that almost anybody who went through a program to study computer science in a university, in a real institute, like a sort of official institution, used open source to learn, right? Anybody who was in academia used open source to do your research, to produce your charts and whatnot.

The interaction of anybody who is vaguely technical with open source is very early. If you’re a hacker by night on your own, in your own basement or whatnot, you have to use open source. You don’t have a budget to learn how to program. You find what is on the internet, you find what works, you grab it, you learn it, and then you improve your skill that way. I do think it’s a bit of human nature to want to give back after taking for quite some time,

to help out with a thing that led you to where you are in the first place, right? And there are actually canonical essays within the open source world by people like Eric Raymond, for example, who was one of the founding fathers of the open source or free software movement, about the gifting culture within open source people that actually pushes up your social status within the open source ecosystem,

pushes up your social capital, right? If you’re someone who just takes all the time, doesn’t matter how good of a programmer you are, you are disrespected within that world. So if that is your peer group, you want to give back, you want to help out however you can. And one of the norms is that if you come up with something that improves the software that you were using that was free, you contribute it back, you contribute upstream is sort of the term, right? So the entire ecosystem of that software becomes better because of something you came up with and people thank you for it.

There’s a lot going on there that motivates people to do this, but it’s not sustainable, right? And over time, frankly, as big companies gravitate towards open source, a lot of people do become paid full time. You work at Google, you work at Meta, you work at Alibaba, whose full time is actually to do open source, but you’re on a big-tech paycheck, payroll, salary, right? So in that sense, you are getting paid to do things for free.

Kyle Chan (13:10)
Mm-hmm.

Kevin Xu (13:25)
So that becomes a little bit more sustainable, but also takes away the purity of it because you’re kind of beholden to the person with the company who pays you. But that has been a struggle for a long time. Now, going to the China version of this, if there is a Chinese characteristic to this story, is that because a lot of these Chinese big tech companies in particular have been taking open source to really accelerate their growth for so long,

and there’s always been this stereotype that anything technological that’s created in China — doesn’t have to be software, could be hardware or could be anything — is either IP being stolen, stolen from the West, or they cheated to create all this stuff, right? Which of course, if you’re an engineer, it kind of rubs you the wrong way, right? So a good chunk of the Chinese open source developer ecosystem or folks who work in the big tech companies have this intense

motivation to give back and to make their own project and then contribute back to the ecosystem, if only to prove that they are good enough. They can produce as much as they can consume, and that they want to show off in a sense and really make an impact in the entire world with their software creation, not just with the stuff they’re taking. And I know that sounds like it’s super warm and fuzzy and unlikely.

But that is how a lot of these open source people are. They’re very, almost positive-sum to a fault in a lot of ways, which is a very nice contrast in a world that is dominated by zero-sum thinking and zero-sum negatives. And they’re just like, well, I’ll just give it back, right? Like I don’t need to have the highest salary. I think it gives me a lot of satisfaction for the world to use my open source package. They just put it on GitHub. It got downloaded, let’s say, 2 million times.

That to me is incredible satisfaction. And quite frankly, if you’re someone who has that level of technical caliber to create something that’s that useful, you’re going to get all the pay package you want from any tech company if you want to make money anyways.

Kyle Chan (15:34)
Yeah, yeah, that’s super interesting. I mean, it’s like the cool factor, the wanting to be part of this community and wanting to kind of show off what you can do. And it doesn’t have to be a monetary reward. It can be just recognition or even just like good vibes. And then you feel like it’s worth it to stay up late at night, on the evenings, weekends, to keep contributing.

Kevin Xu (15:53)
Yeah.

Kyle Chan (16:01)
Yeah, yeah. So yeah, it’s interesting. I mean, also then the sustainability side comes into question. And so you have this kind of mixed approach, with some of the companies as well supporting open source. Yeah, I was wondering if you could talk about some of these movements that have kind of sprung up in China related to open source. Because like you talk about in your piece, like the 1024 Programmers’ Day and also sort of the 996 ICU movement.

Kevin Xu (16:09)
Right.

Kyle Chan (16:27)
What were these about and what do they say about China’s open source culture?

Kevin Xu (16:35)
I think that’s one of the more interesting things that I came upon when I was writing this long kind of history of Chinese open source, is the subculture that actually got formed within open source in China, right? If you think of open source globally as the dominant culture of any hacker, then within China, you had these subcultures and their own identity that just formed very much, again, organically on the grassroots. I think this was also during

the mid-2010s or the early 2010s where civil society and kind of self-organizing movement in China was such that you had the room to do this sort of stuff. Right. And the Programmers’ Day, I will highlight, is one of the funniest stories I can think of, is that basically back in 2010, there was an online website that’s focused on developer community who just put up a poll that says, hey, we should have our own Programmers’ Day. Now there are some precedents

in Russia, which is the first country that ever came up with a programmer’s date. You kind of celebrate your geeks, your nerds, your engineers, your hackers in your country. And they picked a day that was, I believe, September 13th, which is the 256th day of the year, which if you’re a binary person is two to the eight, right? Nerdy AF as a way to pick a date. Now the Chinese developer community wants to almost outdo that.

Kyle Chan (17:54)
Yeah.

Kevin Xu (17:59)
by picking their own day. So they came up with October 24th, which is two to the 10th, 1024, two to the 10th. So that is the day that they picked for their Programmers’ Day. Again, completely under the radar, just kind of existed online as a subculture. And then it really blew up over the years as something that people initially celebrated organically to celebrate their own job, but also to advocate for themselves.

As China’s internet sector became more and more dominant and also more and more cutthroat, right, 996 quickly became a thing. A lot of these programmers are overworked, so they will celebrate or kind of push back on things like, can we get less overtime or better pay, which is a bit more serious. Or you celebrate things like, hey, can we wear less plaid shirt and kill our receding hairline,

which is a lot of male programmers, still a very male-dominant industry even today, but certainly back then. So they also have capacity to laugh at themselves, which I find very interesting as a subculture identity point, all happening on this super nerdy day, October 24th, every single year, even until today. Now, the 996 ICU movement is the more

serious advocacy, self-organizing version of the story, right? You have the fun, warm-and-fuzzy Programmers’ Day, and then you have this episode that really came out of nowhere. This is in 2019. This is probably when 996 was at its height as far as its controversy is concerned in the Chinese tech ecosystem, where lots of programmers are overworked. You probably have heard of episodes of people committing suicide from certain companies and then

using GitHub, which is a Western platform. Basically, they tossed out a repo on GitHub that’s called 996.ICU, which is this way of saying if you work 996 for a long time, you’re going to end up in intensive care unit. We don’t want this to happen. It’s this random project on GitHub that became the fastest-growing GitHub project on the whole platform at the time, garnering like 200,000 stars or something within a couple weeks or so. So a lot of people gravitated towards this as a way to really push back and sort of advocate for their own labor rights, essentially, which is super fascinating. And you’ve also got a lot of pushback from the Chinese big tech, including Alibaba’s founder Jack Ma, who actually spoke against the 996 ICU movement at the time.

As, you know, you should feel privileged and happy to have a job in a big tech company like Alibaba. Why are you complaining about working so long? So it became more of a controversy. Other Western media covered it as well. And what was really interesting is that this actually ended up in a particular decision in the People’s Supreme Court, so the Supreme Court of China, I guess, that actually outlawed 996 as an illegal form of labor practice, which originated from this very

organic self-organizing online movement called 996.ICU that took place on a Western website called GitHub. So all these different elements are super fascinating. Of course, 996 still exists today, so it didn’t work in reality, but it kind of worked in a legal sense, which is probably more than you can say for a lot of things when it comes to organizing in China.

Kyle Chan (21:23)
Yeah, yeah. Yeah, what I love about these stories that you kind of tease out here is just like the human side of all this. And it’s like, yeah, there’s all this talk about the AI race and China versus the U.S. and these sort of great geopolitical games. And then there’s maybe a discussion about Beijing or the different Chinese tech companies. But then what about the people who are building all this stuff, who are grinding through the 996 or,

on the other side, who just kind of love doing this kind of stuff for fun and are really into the nerdy subcultures? And there’s tons of nerdy subcultures in China. So it just sort of captures all these extra, I don’t know, sociological dimensions to this that, yeah, that I feel like aren’t as often talked about. So soon we’ll get into the sort of harder-hitting

Kevin Xu (22:15)
No, that’s right.

Kyle Chan (22:27)
Huawei and all this stuff. But right before that, I just want to ask you, kind of before we hit that era, you mentioned that for a while now, a number of Chinese tech companies have embraced open source. And you’ve written about in this piece and elsewhere, like Baidu has a self-driving car platform, Apollo, or Alibaba is doing a really interesting chip

Kevin Xu (22:50)
Mm-hmm.

Kyle Chan (22:54)
effort using RISC-V architecture, which is open source. I was just wondering if you could, yeah, I don’t know, comment on what’s going on with some of those projects.

Kevin Xu (23:06)
Yeah, I think this was the era where, as the Kaiyuan grassroots community builders, what I call them, began to take root in China in sort of the now we’re moving into the more like mid-2010s, late 2010s era, a lot of, first of all, independent open source startups started to pop up in China. This was also a time when venture capital was free-flowing in China. There were a lot of interesting open source startups being built in the Valley.

And Chinese VCs are always super plugged into what is happening in the Valley as a way to do their own investing in China. So a crop of independent open source startups doing databases, big data warehouses, all these sort of big data cloud infrastructure layers, began to pop up, get really well funded, very popular within the open source community because they’re so kind of pure in a way. They’re not attached to any big tech, who always had a

I would say a fuzzy reputation in open source because we all know Alibaba took a lot of stuff. Are they giving back stuff? Not really. The reputation of the Alibabas of the world at the time in open source was that every time they open-sourced something on GitHub, it’s because it’s an internal project that really didn’t work out. It’s actually a piece of crap. They just open sourced it to give it a nice ending, but they don’t do anything to maintain it going forward to make it useful.

using open source as a junkyard, right? Which is not great. And I think over time, all these big tech companies realized that open source is actually a way to build better software for themselves, but also it’s a great recruiting mechanism as well, right? If you have good open source projects being actually put out and maintained and useful, more engineers actually want to work for your company. So there are kind of multiple very practical benefits to that.

So Alibaba started open sourcing better projects and not just in the software space, but also in the chip design space. So the one that you mentioned, Xuantie, right? It’s based on RISC-V, which is an open instruction set architecture, basically the roadmap in which you design a semiconductor chip that came out of UC Berkeley way back when. And then they really gravitated to that ecosystem to start

their own chip-making firm or subsidiary at the time called T-Head, which may actually become public later this year as an independent company. But that was when they started doing that to gravitate to hardware design. Then you have Baidu, for example, who did the same thing but for a self-driving platform called Apollo, which is still powering all of their robotaxi today. That’s another way to

open source more legit internal projects that actually have a real strategic value, not just a failed project from some random team in the past. And the way you think about this — and I will call out a few other companies that actually may be less even thought of — BYD open sourced their own internal platform to allow

developers to kind of tinker with the software side of a BYD car. This is, I think, 2018 or 2019, called B++ is the platform, right? And there’s this kind of Androidified way of thinking about open source, right? So Apollo wanted to be the Android of self-driving with Apollo, right? And BYD wanted to be the Android of EV in a way for their own open source software. And it’s a way to kind of really

grow their market share, to maybe compete against Western peers at the time. So there’s a lot of different dimensions at play where the larger technology companies, even the OEMs like BYD, are gravitating towards open source. It’s now a strategic play, maybe a recruiting play, but certainly a way that kind of portends maybe what we’re seeing today.

Kyle Chan (27:02)
Mm-hmm.

Yeah, yeah. So we kind of see this shift over time, open source being maybe something that is more on the margins, then coming into more of the mainstream, then becoming a really important strategy for some of the Chinese tech companies themselves that are embracing open source, no longer seeing it as sort of just like this extra thing on the side, but maybe a central part of their strategy for building up, especially global expansion, perhaps. But now

we hit the part of the story where open source becomes sort of existential. So on May 15, 2019, that date, the U.S. announces that they’re gonna put Huawei on the Entity List. And that means cutting off Huawei from a whole bunch of U.S. technology: Google’s Android operating system, Qualcomm chips for its smartphones.

Like this is a huge, huge blow to the company and a huge wake-up call for the entire country. So how did Huawei respond to this, and what role did open source play in Huawei’s response?

Kevin Xu (28:15)
I think the moment that Huawei got blacklisted is the moment where open source transitioned from a vector of growth to a vector of survival for companies in China and perhaps to the entire country, which we’ll talk about a little bit later on as far as how it became a strategic plan of the government. Because the government or certain various divisions or ministries of the government are certainly aware of open source as a thing

up until 2019 when Huawei was blacklisted. And I would argue based on my research and also personal experience that their attitude towards open source is very uneasy. I will cite an episode back in 2015 when GitHub was growing like wildfire everywhere because of its ease of use. Everybody who wants to do open source can now do it so much easier on GitHub, collaborating and stuff like that. And that all obviously necessitated

connections between developers in the United States, in Germany, in the UK, at the same time working online with Chinese developers, Russian developers, Indian developers, what have you. And in 2015, I think I mentioned this in my article, GitHub suffered one of the largest DDoS attacks that came from China. So that was one of the largest attempts by the Chinese government or state-affiliated institutions to shut down

or to kick GitHub off of the Chinese internet. And it didn’t work because the entire developer population in China just stopped working. They couldn’t work. They all work on GitHub in one way, shape, or form. They became instantly unproductive. And the government realized that as much as they find the social aspect of open source may be a little bit troublesome from their perspective on how they want to run their internet and their country,

they can’t afford to shut this thing down because perhaps the most innovative part of their population will become instantly unproductive overnight. So they always had to live with the pros and cons of open source as time went on until 2019, where the entire open source ecosystem that has roots in the West got shut off. You can think of it almost like that was a first moment when open source became weaponized in the context of geopolitics.

Kyle Chan (30:37)
Hmm.

Kevin Xu (30:40)
which is quite ironic because most open source exists in the public sphere, it’s public domain, its root of origin is quite irrelevant to how you use it and customize it. But Android or the Google version of Android clearly had more ties to Google than it is an actual open source ecosystem that Huawei could not rely on any longer because of the blacklisting. So that really pushed open sourcing, at least from Huawei,

to a whole new level. Huawei was building alternatives at the time before because it probably always saw the writing on the wall, like at some point we need to have our own version of everything, but it didn’t have to do it, so they didn’t really ship anything that was super useful until that moment. And then it became the only thing that it could do to survive as a company. So Huawei became a relatively late player actually into the open source scene, but obviously a very powerful player

when it had to be that way. So it started open sourcing like everything they had to have: the operating system called HarmonyOS, their databases, their servers. They have their own cloud, so they have their own server architecture called OpenEuler and they open sourced that, a bunch of stuff. They also even seeded an open source foundation called OpenAtom using one of their projects, which is very much a one-for-one

comparison to the Linux Foundation. Linux is seeded by the Linux operating system, and then they seeded their own thing called OpenAtom Foundation as a way to open source not just the technology, but even the community fabric of open source. They wanted to create their own ecosystem

Kyle Chan (32:19)
Hmm.

Kevin Xu (32:25)
of it, which again is, I think the jury is still out as far as how successful that is. I remember the earlier days when HarmonyOS first became open source, but it’s like, okay, so how do you download one and just fool around with it yourself, like a real open source project? And you had to register an account and put your information and give it to Huawei to get the permission to download the source code or something. And people were like, this is not real open source, right? This is not legit.

Kyle Chan (32:52)
Yeah.

Kevin Xu (32:54)
at all. You know, Huawei doesn’t know how to open source for the life of it, right? So it definitely had really rough starts in the beginning. It probably got better over time, so it takes quite a bit of reps to really open source well, even for some big outfit like Huawei.

Kyle Chan (33:10)
Yeah, yeah, it’s like not as simple as building your proprietary software, like pouring a bunch of resources, software engineers, just going all-out sprint. You’re cultivating an ecosystem with open source. And there’s only so many levers you can pull to really sort of get people to contribute and to streamline it and make it more attractive to maybe not just Chinese developers, but global developers. And that takes time too. You have this network effect as well, which

makes it very difficult. And ultimately, I mean, we see this Huawei-led alternative open source stack, as it were, seeming to emerge.

Kevin Xu (33:50)
Yeah. And the other

thing too, which is very not second nature to these large tech companies like Huawei’s of the world, is that the moment you want to really dive into open source, you have to lose some level of control of the project and feel okay with it, right? That’s always been a very difficult line to walk, not just for Chinese companies, for any companies. Google has had problems like this, right? Microsoft has had problems like this, definitely, which is that

Kyle Chan (34:04)
Mm.

Kevin Xu (34:19)
you want the developer ecosystem to build. You want people to contribute. That’s the taking part of open source, which is great if you can get it to work. But you also have to listen to these developers. These developers might want your software that you originally created, but now you’re open sourcing it to go a different route on the roadmap. They want feature XYZ that they really want, but may not help you monetize off of your software later on.

Kyle Chan (34:29)
Mm-hmm.

Kevin Xu (34:47)
but you have to play this balancing act of how do you really keep that community going? Because the moment you dictate — which open source has all these really interesting liberal democratic elements to it as well, that again is super fascinating when you think of it in the context of China — is the moment you dictate a direction, the developer communities are gone, right? They’re just like, there are all kinds of open source databases. Why would I waste my time, waste my weekends and my nighttime,

Kyle Chan (35:07)
Yeah.

Kevin Xu (35:14)
fooling around with your ecosystem where you don’t even listen to me, right? There are all kinds of options out there, so it’s super hard actually to manage open source, which is something that these grassroots organizations like Kaiyuan She, for example, that we talked about a little bit, really did the work, I think, educating the developer community in China how to do this well, because the code is like almost 10, 20% of the work.

Chinese engineers have been really competent at writing good code, I think, for a very long time. But how do you interact with the broader community entirely online? You will probably never meet these people in person. You’re going to interact through your GitHub username and now your Hugging Face username for your entire lifetime. How do you write nice-sounding issues to not piss people off? How do you propose a solution that doesn’t sound super dictatorial?

Kyle Chan (35:43)
Yeah.

Kevin Xu (36:11)
or top-down or offensive or whatever? These are these social cues. Of course, for most Chinese developers, you have to do this in your second language, in English. Might be way easier now because you use Claude and ClaudeX to write these for you so they can all sound really nice. But in the bygone era, they had to use a second language to communicate with the world to really be part of the fabric. Super, super difficult. That’s something that Huawei had to stumble through to really get to where it is today.

Kyle Chan (36:19)
Hmm, right.

Yeah, yeah. Well, it’s interesting because as challenging as it is for a company like Huawei to try to do this, strike this balance between kind of creating this open culture versus kind of getting it to do the things that it wants to do, you have Chinese policymakers trying to do this. And you kind of point out a big shift, maybe 2019, 2020, and especially 2021, where there was a much more wholehearted embrace

of open source as a sort of national policy strategy. So, and I was looking back through the different Five-Year Plans, and I think the 14th Five-Year Plan mentions open source for the first time in 2021. And then you also point out the Ministry of Industry and Information Technology talks a lot about open source. So yeah, I was wondering if you could say more about

what is the shift for Beijing, for Chinese policymakers? How are they trying to foster and embrace open source? What’s the goal here?

Kevin Xu (37:43)
This is probably the most interesting thing, if you think about open source in the macro sense for China, right? And also U.S.-China, which is that typically a lot of things that happen in China kind of have this official recognition from the top first as this is strategic, right? You know, solar, battery, autonomous driving, what have you, robotics, and then the rest of the industries then...

Kyle Chan (38:01)
Mm-hmm.

Kevin Xu (38:10)
gravitate towards that. Their local governments issue their own version of it to compete in this new planning phase and whatever. Open source, however, has more or less existed in the background, kind of underneath, on the DL for pretty much two-plus decades. It’s not to say the government wasn’t aware that open source is happening. Certainly even back in the ‘90s, there were certain ministries who thought this was a good thing to embrace, free software or open source or Linux. So there was some

writing of that even at the official level. But at the time, open source was, one, treated, I would say, skeptically because of the social element of it that I mentioned when it comes to attacking GitHub or shutting GitHub down, but also as more of a defensive measure, as in this will be a good thing for us to try out, figure out as a way to improve our own IT security. Because at the time, any ministry who has a computer is using

Windows, IBM mini-computers, Cisco routers. These are all closed-source Western technologies. And they’ve always been uneasy of that, but they didn’t have their own national champions at the time to really have another option to begin with. Huawei was more or less created to be that option, which Huawei has its own feelings about at the time as well, as far as how conflicted they are. But it’s more defensive. And only up until the 14th

planning, 14th Five-Year Plan planning cycle, did open source become one, existential potentially for the future of the country as a survival mechanism, but also as a way to actually play offense, to innovate, not just to be an alternative to weed things out, but to innovate on top of it. Even the Made in China 2025 plan, which is this famous-infamous plan depending on how you want to think about it as far as decoupling, the origin of decoupling is concerned, that was published in 2015, did not mention open source even once in the document, right? And

Kyle Chan (40:17)
Hmm.

Yeah.

Kevin Xu (40:28)
when it comes to 2021, when MIIT mentioned open source in its own planning document, I think 27 times, was when the first time you really see open source being embraced as official strategic, at the very top strategic level, right? And

Kyle Chan (40:17)
Yeah.

Kevin Xu (40:28)
even before that, we’ve had episodes where the MIIT ministry, which is the regulator for the internet sector, the tech sector mostly — there are others that kind of have that turf now — started to pick winners as well. So Huawei is the obvious winner, obviously, but then you have an outfit like Gitee, for example, G-I-T-E-E, which is the homegrown alternative to GitHub that was founded years ago as a way to

I don’t know, kind of take market share from GitHub, right? It’s been super difficult. And then the ministry kind of designated Gitee as the national champion, the preferred platform that it would like at least the government or regulated industries to use and embrace for their own source code management system, as opposed to use something that is Western like GitHub. So you see all those behaviors really start to pop up

really much later in the history of Chinese open source, right after Huawei basically got blacklisted.

Kyle Chan (41:33)
Yeah, yeah. I guess it’s an open question then whether trying to harness or leverage open source will end up sort of squeezing out the very sort of essence of the culture itself or whether it can be done and steered in a certain direction. I mean, yeah, we’re seeing this experience sort of unfold. I want to jump now to AI.

Kevin Xu (41:58)
That’s right.

Kyle Chan (42:02)
And the DeepSeek moment. So, in January 2025, so a little more than a year ago, DeepSeek released R1 and a lot of people were stunned by the performance, how close they were to sort of the U.S. proprietary models and also by their claims of having such a low-cost baseline. But I think you really focused on their

huge push into open source, like in many different ways. And I was just wondering if you could talk about why DeepSeek’s open source moment was so important and what has sort of happened since with open source and AI in China?

Kevin Xu (42:47)
Yeah, it’s funny that we’re 40-plus minutes into a conversation and we did not mention AI even once, really, right? Until now. And I hope that is the, if there’s one takeaway from either listening to this podcast or to read my essay, is that open source AI in China is the latest of a long series of chapters

Kyle Chan (42:53)
Hahaha!

Kevin Xu (43:11)
of stories and history and activities when it comes to open source in China, right? That’s the current chapter. This is by no means the first chapter of any of that stuff. So what really, DeepSeek R1, when that released, first of all, at the time, everyone was reading the papers. I’m still reading the papers, right? And DeepSeek V3 was the big foundation model that trained R1. That was released actually over Christmas in 2024. But even V2 had a

paper as well, the previous version to V3 that was released like mid-2024. That was when I started paying attention to DeepSeek as an outfit, right? They were practicing kind of your classic open source playbook already at the time, which is every one of their models or their model weights are licensed under what is called a permissive license, right? They use MIT License, which is one of the two very commonly known licenses within the open source world,

where, you know, if you see MIT License, you can do whatever you want effectively. You can download a copy, make derivatives of it. You don’t have to tell DeepSeek. You don’t have to pay them a cent. You don’t have to do anything. That is the ultimate kind of the purest form of open source. And they chose that right off the bat because they know the Llama license that Facebook came up with to initially release Llama, which was the big open source model at the time, had all these weird restrictions

about if you reach a certain number of users, you have to get a commercial license with Meta and other sort of stuff that is not pure open source, which of course helps Llama’s commercial case. But frankly, if you’re an open source user, you don’t want to have to think about the eventual legal consequences if something you do actually comes up, right? Which you never have to think about if you use something that is MIT-licensed. So DeepSeek did that immediately, right?

And then when DeepSeek released R1, a lot of the optimization that got its cost structure to as low as it was — and there are still controversies and debates about how real that number is, we will set that one aside — but I think what is really important is that it shows that DeepSeek’s team’s level of hardware optimization sophistication, which at some point must have learned from open source research,

from having open source hardware access to be able to teach themselves how to do these optimizations, led to that cost structure to begin with. So there is a technical education pipeline that always runs through open source that DeepSeek benefited from. And then right after R1 got released and kind of blew up for the entire world, they also did this thing called Open Source Week that I thought was super fun and very classic open source ethos, community-building thing, which is that they were releasing one

piece of open source library out of their tool chains, and not just the models anymore, but a data pipeline, a tokenizer, some other kind of components, right? There are lots of components that lead to a model being trained successfully. Open source once a day for an entire week to generate more community goodwill, more developer access, ecosystem contribution, what have you. Very well done from an open source go-to-market perspective, right? And that I think,

to me at least as an open source nerd, was the thing that really set DeepSeek apart. And then as we go through 2025, when Kimi and MiniMax and Zhipu, now Z.ai, all started doing their own version of open source, not to mention of course Alibaba’s Qwen, which actually open sourced the earliest of any of these labs, even before DeepSeek. You see that maturity of these

You see like Yang Zhilin, who founded Moonshot, who makes the Kimi models, go on Reddit to interact with the entire community face to face. It’s very Sam Altman-like. You just go on Reddit, you see what the Reddit world throws at you. It’s almost never super friendly, but sometimes it can be, sometimes it can be hostile. And of course he’s doing it in his second language, but he’s not scared of interacting with the community face to face.

Kyle Chan (47:09)
Yeah.

you

Kevin Xu (47:28)
That builds a ton of goodwill and accessibility, which I think definitely contributed to Kimi’s popularity. Zhipu actually does a similar thing as well, so it’s not just Kimi. They’re also on these Western social platforms, right, to interact and build community interaction. And all this is just building up to that level. I remember talking to Chinese open source founders back in 2015, 2016, getting them to go do Reddit stuff to drum up their popularity, and they just couldn’t bring themselves to do it.

Kyle Chan (47:55)
Mm.

Kevin Xu (47:56)
It’s hard, it’s in the second language, you don’t want to say something wrong, and it’s embarrassing if they do something silly, because you have to have a good sense of humor as well on Reddit. But now this new generation of AI founders are much more comfortable in that realm. And DeepSeek is sort of the thing that launched it, but it’s by no means super unique in that way.

Kyle Chan (48:00)
Right, right.

Yeah. So you see DeepSeek as part of this longer lineage of open source, and then kind of contributing back to that. And then now all these other Chinese AI labs, I think even like you mentioned Zhipu, I think they might use now some of the DeepSeek contributions, like some of the sparse attention mechanisms that DeepSeek has sort of contributed. So it kind of feeds onto each other

and builds up this broader ecosystem, and then helps with, like you pointed out, global expansion and popularity abroad combined with this sort of PR push to get out there, try to actually interact with your users wherever they are in the world, basically.

Kevin Xu (49:05)
Yeah, yeah. And one last thing I’ll just mention on that point as well is that at this moment, going overseas, global expansion is probably still super top of mind for every Chinese entrepreneur. Doesn’t matter if you’re in the AI realm or still in the old-school tech realm or even in hardware and what have you. And open source is

unfortunately maybe their only way to get their name out there. They have their work cut out for them, right? They have the geopolitical problem. They have what I believe is kind of the toxicity of the China label, right, that you have to shed in multiple different ways. And you don’t have enough funding compared to your peers, certainly in the AI world, to make a big PR push or something like that, or a marketing push. None of them are going to drop millions of dollars in a Super Bowl ad.

But open source is a way where if their thing, their technology, is good enough, it could be at least evaluated on its own merit for a period of time for anybody who bothers to spend the time to kick the tires a little bit. And certainly, their climbing the benchmark chart, that really helps with getting noticed. You can talk about benchmarking as a way to get around all this sort of stuff that’s maybe not super honest. But at the end of the day,

this is one of the only vectors in which any Chinese tech can really have a presence abroad when frankly the domestic economy or the domestic ecosystem for consuming software technology is still really, really bad and not a great way to make money. So they have to go through this route. So open source has become not just a nice-to-have, I think, but a necessary strategy, which is probably why you see all the Chinese tech gravitate towards open source and really to play it well, to play nice, to build on top of each other.

Their vibe to each other is super friendly. There’s like no obvious hatred or rivalry between any of the AI labs. All that is part of this image that has helped them push themselves abroad in a way that can be recognized despite where they’re coming from.

Kyle Chan (51:15)
Yeah, yeah, yeah. I like that characterization. And now we’re sort of in the stage where earlier open source might have been sort of a newer, nascent, emerging social trend community and then became more central. Now we are in the age of open source frenzy in China, it seems like, like really wholehearted embrace of things like OpenClaw. And I was just wondering, what do you make of all that?

What does that say right now about the state of AI and the state of open source in China?

Kevin Xu (51:48)
Yeah, you know, it’s so, it’s kind of mind-boggling how quickly OpenClaw, which is, for those of you who haven’t heard about it, it’s an open source AI agent platform, right? That was released, I’m going to say a few weeks ago, two to three weeks ago, maybe a month ago, I can’t quite remember, but it’s not a very long time. It climbed the chart on GitHub faster than any project ever. I think it already exceeded a star count

over Linux of all projects within just a few weeks that it has been alive. So it’s almost a new Linux if you want to think about it. And it’s an open source tool. So one thing that’s worth mentioning is that any open source tool, doesn’t matter how popular it is, it’s inherently kind of difficult to set up and use and operate if you’re not somewhat technical or feel comfortable kind of fooling around with some technical details, which is not for everybody.

Kyle Chan (52:19)
Yeah.

Kevin Xu (52:45)
And a very common way to make open source easier to use and to also make money off of it is to offer kind of cloud-hosted solutions, right? Kind of zero setup required. Your granny can do it. Your uncle can do it. And then you just start using it. And the funny thing is that every single AI lab from China and also the hyperscalers up to this point have now offered their own hosted solution of OpenClaw on their own platform

to boost adoption, even to municipal governments. I think Shenzhen and Wuxi have issued directives and policies about how great open source — or not open source, but OpenClaw in particular — is, and we as a city government want to incentivize, subsidize, and help people get on board with this new thing that came out of frankly nowhere on the internet. And the more interesting question to me is that this playbook has been well played

by American hyperscalers too. They really started that, right? AWS was a huge cloud platform of fully hosted open source things, whether it’s database or containers or all this sort of stuff, for a long time and also without giving back for a long time. So AWS’s reputation in open source is very checkered in its past, right? But they kind of played that playbook really, really well. Every single Chinese vendor has internalized that playbook.

And the big question to me right now is why are the American vendors so slow to offering hosted OpenClaw on AWS or Azure or whatever, when all the Chinese vendors have been almost overnight at this point. But yeah, the adoption of OpenClaw is super interesting. Again, it shouldn’t be surprising if you go through the commercialization history, the VC-funded startups that have happened in China,

how they’ve learned to commercialize open source in the cloud era. So now that we are in the AI era, which is still 80% cloud, if you think about it, all the AI stuff that you consume has to ping a cloud server somewhere before you get the response back. So the playbook is actually very, very similar, how they’ve kind of turned on that playbook just much faster than any other country, including the United States.

Kyle Chan (55:02)
Yeah. And it’s interesting because there’s clearly a lot of demand in the U.S. and in China for this. Like you see these crowds of people getting together to all set up their OpenClaw on the used MacBook mini that they got. And so there’s clearly demand, but yeah, somehow you just have this — I mean, I don’t know, maybe this also feeds back into like, if you’re already open-source-forward in China as an AI lab, for example, this kind of fits in better.

And if you’re, I don’t know, like a big public tech company in the U.S., you might be more cautious because, I don’t know, there are sort of various business risks along the way of doing this wrong. I mean, one can speculate.

Kevin Xu (55:47)
Yeah, yeah, I think there’s certainly a little bit of that too, right? Obviously we focus on why China open source is a thing. I think it’s worth a podcast and somebody to talk about why closed source in the U.S. is becoming a thing or why is open source AI in U.S. less of a thing? You know, open source did not start in China. It certainly started in the U.S. on the campus of MIT of all places, right? But yeah, we are shutting ourselves off here in the United States from the goodness of open source.

Kyle Chan (56:04)
Yeah, yeah.

Kevin Xu (56:16)
Is it because of sky-high valuation? Is it just because we have to make more money? I mean, all these AI startups in China also have venture funding, so they have to make money at some point too. So I don’t know if that incentive is necessarily in conflict with each other, but that’s something that’s worth honestly thinking about a little bit harder as far as the overall conversation of where U.S.-China AI is going.

Kyle Chan (56:38)
Yeah, yeah, yeah. Well, on that question, I was wondering what you thought about sort of the future trajectory of open source in China and whether you think that this will continue, especially for the AI labs, or whether you think that they’re going to eventually converge with the U.S. and switch to more proprietary models. I mean, we see like OpenAI, Anthropic, they are spending like ungodly sums on compute.

But they are also making billions of dollars, or maybe even tens of billions of dollars, in revenue from subscriptions, APIs. So yeah, do you think that this is something that Chinese AI labs are going to eventually have to switch to in order to monetize? Or can they, or will they want to keep pursuing open source as sort of their main strategy?

Kevin Xu (57:24)
Yeah, I think that is probably the trillion-dollar question, right, for the entire industry. My current feeling is that because, like we mentioned, the 14th Five-Year Plan really, I think, enshrined open source at the top strategic level. Even the most recent Two Sessions speech by Premier Li Qiang gave open source a shout-out in his work report speech. So it’s very much at the very top level. I can only surmise or assume that

open source will be prominently featured in the 15th Five-Year Planning cycle, which we haven’t seen the official document yet, but that’s my current expectation, which means that a lot of these labs will continue to open source, probably because it aligns with government incentives now, right, or government strategy in ways that the government did not play a huge role five, six years ago at that level. Now it is, and that’s

always a double-edged sword, quite frankly, but it is where it is and they have to do it. But I also think there will be probably three different tiers of open-sourciness that comes out of China. Tier one is DeepSeek, which is going to be an N-of-one example. They will, I believe, always open source to the max because they actually don’t have a

strong incentive to profit-maximize in their own setup. It’s very unique that they get their funding from, whether it’s a hedge fund that’s still making money or just whatever revenue they generate from their API and their chatbot to keep the lights on. And that’s more or less sufficient. They don’t, they’re not accountable to any outside investors or the public market or things like that, right? They will continue to do so. I think a middle tier will be maybe the Alibabas of the world,

where they’ve open-sourced a lot up to this point, but their pressure to monetize is much higher because of the fact that they have a huge cloud business that they have to run, they have to grow. One company that we did not mention up to this point is ByteDance, which is interestingly the least open source of all the AI labs. They have hardly released any open source model that’s worthy of any mention. Their closed-source model, Seed, Seedance,

it’s very, very good, right? And Doubao, their app, is very, very popular, but all that is closed source. So they’re not actually going along with this trend. They’re going against this open source trend to do their own thing. So they’re going to be in that world of closed or half-open-source type situation. And then you’re going to have, I think, some of these independent labs like Zhipu, MiniMax, and Kimi, who again, coming back to what I was talking about, open source as a vector to go overseas.

Kyle Chan (1:00:16)
Yeah.

Kevin Xu (1:00:16)
I

believe they have to continue on this path for maybe much longer than their cap table could even withstand. But it’s very important and it’s also working really, really well. If you look at OpenClaw’s most-used model right now — not OpenClaw in China, OpenClaw period globally —

Kyle Chan (1:00:25)
Yeah, yeah.

Kevin Xu (1:00:39)
I think MiniMax is like number three right now, like Kimi is like number four, like both of them are in the top five as far as if you’re just a random OpenClaw user anywhere in the world, which model do you end up running your OpenClaw on? They’re up there, right? So I think it’s working right now, so they have to keep pushing. But I would say three tiers of open-sourciness along the spectrum of open versus closed is what I’m seeing.

Kyle Chan (1:01:04)
Yeah, yeah, yeah, yeah, especially these smaller startups. I mean, in China, they’re really riding this wave, especially for OpenClaw and really gaining a lot of popularity outside of China because of that. Yeah. Yeah. Is there anything that we missed that you wanted to kind of highlight or any other sort of trends you want to call out? Or if not, worries.

Kevin Xu (1:01:29)
Yeah, no, I think if we were to wrap up, one thing I wanted to just mention for anyone who was listening is that open source is this global movement, right? And it’s probably one of the few things that are actually keeping global collaboration together when everything else in technology is becoming more nationalistic, right? More zero-sum versus positive-sum. And there are even movements and energy in different policymaking chambers to

close off open source, whether it’s from China or from elsewhere in general, to shut off open source, right? Which I think is probably shooting yourself in the foot. It doesn’t matter which country rolls out that strategy. One thing that we haven’t talked about as much, but I think is worth emphasizing here, is just that without open source, there is no technical pipeline for your own country, right? We need to emphasize that open source is at the root of any technical education

Kyle Chan (1:02:19)
Hmm.

Kevin Xu (1:02:27)
initiative. And if we talk about sovereign AI right now, how do you exert sovereignty and control on AI for country X or country Y? If you don’t have enough capable, technically capable people within your own country, there’s no way you can figure it out, right? You can’t have sovereignty and control without having good people to really exert control over technology. And that starts with learning about technology. You cannot learn technology without open source.

So I hope whoever’s listening that’s kind of on that vein, we can have a conversation about this. Open source is pro-innovation, it’s pro-education, and it’s actually pro-competition. It’s anti-monopolistic by nature. So all of these stuff has to come together. And the China story is, quite frankly, just one interesting iteration of the global movement of open source that has been, again, happening for three or four decades at this point.

Kyle Chan (1:03:19)
Yeah, this is amazing. Wow. Yeah. It’s like, it’s not just the particular projects that come out of the open source movement and all that. You know, it’s this broader, deeper effect that it has on what you can do with code and what you can do to foster this sort of technical capability at a national scale globally. So there’s a lot more to the story. Well, this has really been fantastic. I will definitely link to the piece

that you just wrote. I will link to Interconnected. Are there other ways that people can follow you and your work?

Kevin Xu (1:03:58)
I mean, those links are great. I’m on Twitter probably more than I should, so you can follow me on Kevin SXU is my Twitter handle. Obviously I run my own fund as well. The website is pretty boring, but you can go to interconnectedcapital.com if that side of what I do is what you’re interested in. So thanks again for having me, Kyle. Really appreciate it.

Kyle Chan (1:04:22)
Yeah, this is awesome. Thanks so much, Kevin. So just to wrap up, if you like this episode, please rate and subscribe on YouTube, Spotify, or Apple Podcasts. You can find episode transcripts and more information on the High Capacity newsletter at highcapacity.org. I’m your host, Kyle Chan. Thanks for joining and see you next time.

Watch or listen to the High Capacity podcast on:

• YouTube

• Apple Podcasts

• Spotify

In this episode, I speak with Afra Wang, writer and observer of China’s tech culture as well as broader cultural trends in China and the US. Afra is the author of the Concurrent newsletter and has published work in Wired Magazine and Asterisk Magazine.

Follow Afra’s work:

• Concurrent newsletter

• Afra’s professional website

• Twitter / X

• CyberPink podcast (in Chinese)

• Wired: “You’ve Never Heard of China’s Greatest Sci-Fi Novel”

• Asterisk: “The China Tech Canon”

Transcript

Kyle (00:00)

Welcome to the High Capacity Podcast. I’m your host, Kyle Chan, a fellow at Brookings. I’m thrilled to be joined today by my guest, Afra Wang, an absolutely brilliant writer focused on China’s tech culture and broader cultural trends in China and the US. She’s the author of the amazing Concurrent newsletter and host of the popular CyberPink podcast, which is in Chinese. She’s currently joining from London, where she’s also a fellow with Gov.ai.

Welcome, Afra, and thanks for coming on the show.

Afra (00:31)

Thank you for having me, Kyle.

Kyle (00:33)

I want to start by asking a big question: what is Chinese tech culture? In Silicon Valley, we hear about hoodies, tech bros, and billionaire founders with big egos who love to move fast and break things. In China, from an American perspective, we hear a bit about the 996 work culture (996工作制)—working from 9 a.m. to 9 p.m., six days a week—and this relentless, bring-your-mattress-to-work culture. But beyond that, what is Chinese tech culture actually like? Is there even such a thing? How does it compare to Silicon Valley? Are there tech bros (科技男) in China? Let’s start there.

Afra (01:05)

Yes, there are a lot of tech bros in China. If there’s one similarity between Silicon Valley and China, it’s the density of tech bros, although they are “bro-y” in a very different manner. We can talk about that later. This is a great, broad question. When you ask what China’s tech culture is, I can break it down through founder archetypes, generations, and regions. There’s a Shenzhen (深圳) founder versus a Beijing (北京) founder—they have very different personalities and backgrounds.

In general, there are some patterns I can map out. The first one is engineer realism. Honestly, Kyle, your newsletter, High Capacity, is a reflection of this engineer realism, where all sorts of industries converge to produce very tangible products. This “product first, shipping first, philosophy later” virtue is something Chinese tech culture appreciates. There’s intense focus on execution, shipping, iteration, and scalability. Anyone constantly talking about philosophy or posting on social media would be slightly looked down upon by Chinese tech people because they value doing rather than talking. There’s a strong sense of wushi (务实) instead of wuxu (务虚). Wushi means minding the substance, and wuxu means minding the philosophical or intangibles.

Another aspect is the hypercompetitive mindset. I was reading Kai-Fu Lee’s (李开复) 2018 book, AI Superpowers (AI·未来). He used Wang Xing (王兴), Meituan’s (美团) founder, as an example to educate Silicon Valley readers: “Look at people like Wang Xing. They are no less smart than you, yet they are fighting 100 times more brutally.” If Silicon Valley founders are driven by ideals and changing the world, Kai-Fu Lee noted that Chinese founders don’t necessarily have this. If we trace back the history of the internet, pivotal moments are named by wars. For example, the “Hundred Regiments Battle” (Baituan Dazhan / 百团大战). That was the moment when Meituan, Didi (滴滴), and other mobile-first applications were trying to take control of the market and kill each other.

Afra (04:42)

Right now, we’re in another Baituan Dazhan with AI. It’s the Spring Festival (春节), a moment when everyone is on their smartphones sending messages. This is a pivotal time for AI companies to send subsidies and promotional messages during the Spring Festival Gala (春晚) to take control of the market. There’s a mindset in these subsidy wars: “If I dare to lose two arms, I can win over someone who only dares to lose one arm.” It’s a self-sacrificing mindset where they are willing to double down, burn resources, and sacrifice short-term interests to gain long-term longevity. It’s really brutal if you read the history and watch the current competition.

Another interesting aspect is how path-dependent Chinese tech culture is on the gaokao (高考), the Chinese college entrance examination. Just as a Silicon Valley founder is essentially replicating the college application “telling your story well” culture, the Chinese tech culture runs on a science experimental class (理科实验班) culture combined with extreme meritocracy.

Afra (06:30)

The gaokao is essentially a resource allocation system. People excelling in this system—those who get into the top 1%—spend the rest of their lives trying to replicate their success. They try to extract the magic potion that let them win the gaokao and drink it over and over again when facing obstacles while expanding their companies. That DNA really pushes the founder forward. One interesting feature is that many Chinese tech companies are extremely good at winning benchmarks and excelling on different AI leaderboards. This is such a gaokao mentality.

Afra (07:27)

The gaokao has a very clear reward and scoring system. If you follow this clarity of what gets rewarded and what doesn’t, you can win. People tend to apply this clearly defined metric to everything in their life. That’s why a lot of Chinese tech people are blamed for lacking humanities sensibilities, philosophical frameworks, or just being criticized as poor public speakers. When I listened to Yan Junjie (闫俊杰), the MiniMax founder, on Luo Yonghao’s (罗永浩) podcast—a three-hour podcast, much like Lex Fridman’s—it was painful. He mumbled a lot and didn’t know how to draw an interesting framework to describe the situation MiniMax was going through. He lacks the emotional depth to tell a beautiful founder’s story. He just sounds like anyone from my semi-elite public high school’s experimental class.

Afra (09:04)

Next is the generational divide. I would argue that China has four different generations with very distinct cultures because they are products of their time. In the Chinese movie ecosystem, directors are categorized into generations—Jia Zhangke (贾樟柯) is the fifth generation, Bi Gan (毕赣) is the sixth, Zhang Yimou (张艺谋) is the third. Each generation has a distinct feature and theme. This is a very applicable framework for tech founders as well.

Afra (09:56)

The first is the Ren Zhengfei (任正非) generation. They spent their formative years during the Cultural Revolution (文化大革命) when China started to reform and open up (改革开放). In the 90s, they dared to “go into the sea” (xiahai / 下海). Xiahai was previously a derogatory term, but Ren Zhengfei represents the generation willing to strip away socialist idealism and become capitalists. He founded Huawei (华为) in 1987 after leaving the People’s Liberation Army (解放军). This generation viewed business as warfare and experienced real poverty and struggle, like famine and the Anti-Rightist Campaign (反右运动). Huawei’s corporate discipline is semi-military style, often referring to the “Huawei Army” (华为军团) or the “wolf culture” (狼性文化). Ren Zhengfei’s worldview, shaped similarly to Xi Jinping’s (习近平), draws on Maoist vocabulary—viewing things as struggles, revolutions, and fights. However, Ren Zhengfei is also deeply westernized at his core. Huawei is actually more internationalized than Tencent (腾讯) in many aspects, like the number of foreigners they hire and operating based on IBM’s organizational system.

Afra (13:05)

The next is the Jack Ma (马云) generation. They matured in the early 2000s when China entered the WTO and global capital flooded in. Jack Ma built Alibaba (阿里巴巴) during the honeymoon phase between the US and China. There was a beautiful synergy between globalization, an economically booming China, and a booming internet culture. Jack Ma absorbed a lot of Silicon Valley optimism, even while remaining deeply attached to Hangzhou (杭州) and drawing on local traditions. He performed like a rock star on stage, talking about grand ideas and limitless growth, much like Silicon Valley founders. His generation represents the most romantic, globalized, and innovative period of Chinese internet history. People like Jack Ma, Pony Ma (马化腾) from Tencent, Robin Li (李彦宏) from Baidu (百度), Zhou Hongyi (周鸿祎) from Qihoo 360 (奇虎360), and Lei Jun (雷军) from Xiaomi (小米) emerged from this generation. They brought an openness to the world and built China’s digital infrastructure.

Afra (15:00)

Moving towards the 2010s, the Zhang Yiming (张一鸣) generation emerged. This was the decade of the mobile internet, when the majority of Chinese people experienced the internet for the first time via smartphones. The Zhang Yiming generation saw the huge opportunity of 1.4 billion people accessing the internet. ByteDance (字节跳动) is emblematic of this generation. They are incredibly good at refining recommendation algorithms and building mobile apps. Wang Xing at Meituan and Cheng Wei (程维) at Didi are from the same generation. They are practical; they stopped worshipping or directly copying Silicon Valley blueprints. Facing a vast Chinese mobile market where Western experiences weren’t replicable, they figured out methodologies that Silicon Valley eventually started to borrow.

Afra (17:10)

Now we’re in the fourth generation—the AI and robotics generation of the 2020s. They were forged in the crucible of US-China technology rivalry. Growing up with the reality of pushbacks and sanctions that began around 2018, they realized they had to grow out of China’s own tech strengths and advantages. They are doing the most impressive things right now in semiconductors, AI, autonomous systems, and new energy. Think of Yu Kai (余凯) from Horizon Robotics (地平线), Yang Zhilin (杨植麟) from Moonshot AI (月之暗面), Chen Tianshi (陈天石) from Cambricon (寒武纪—the Nvidia of China), SenseTime (商汤科技), Zhipu AI (智谱AI), Wang Xingxing (王兴兴) from Unitree Robotics (宇树科技), DJI (大疆), and Bambu Lab (拓竹). They are smaller, “wolfpack-ish” leaders in niche industries, especially hard tech. Many of them, like Yang Zhilin, who was educated at Carnegie Mellon, are bilingual and highly internationalized, yet they choose to return to China to start their companies. They have more confidence in themselves and a stronger domestic tech ecosystem to build unique companies combining the best of both worlds.

To summarize briefly: we know BAT (Baidu, Alibaba, Tencent / 百度、阿里巴巴、腾讯), but there’s a newer acronym called TMD—Toutiao (头条 - ByteDance), Meituan, and Didi. These TMD companies represent a generational shift toward being more innovative, AI-focused, and tailored to specific consumer needs.

Kyle (19:36)

That is fantastic. I could just type up what you said, print it out, and it would be a really awesome piece right there. It’s so interesting to see the story of China’s tech development through these generations of founders—to see the evolution in their focus, especially the shift towards hard tech and technical founders who were at the top of their class.

I was wondering if you could talk about their position in society. Are Chinese tech founders worshipped like they are in the US, the way people look up to Elon Musk or Sam Altman? And how do they navigate domestic and international politics, balancing the need to be patriotic enough for Beijing while avoiding US sanctions?

Afra (21:17)

The crackdown on Jack Ma changed Chinese tech culture profoundly. Before, celebrity tech figures like Jack Ma would publicly criticize the Chinese banking system and try to cultivate a group of charismatic, “philosopher-king” entrepreneurs through Hupan University (湖畔大学) to change society both technologically and ideologically. After the purge of Jack Ma, that stopped. The state essentially sees the Jack Ma class as a threat, and taming them was a process of shaji jinghou (杀鸡儆猴)—killing the chicken to scare the monkeys. Tech founders learned from this and became much more low-key. They barely tweet or speak up anymore. I constantly go back to the archives of Wang Xing’s Fanfou (饭否) or Zhang Yiming’s Weibo (微博) to remember what it was like when tech founders could lead discussions.

Afra (23:30)

Right now, entrepreneurs and the state have an interesting equilibrium. It’s a constant push and pull where entrepreneurs are told they are useful servants of national development, but they’re also trying to gain global influence. It’s complicated and divided by generations and industries. For example, consumer AI founders are more willing to move their headquarters to Singapore or the US to sever ties with the Chinese state and get Silicon Valley funding. Conversely, founders in semiconductors or robotics find it more beneficial to stay in China, enjoy the supply chain and state subsidies, and be celebrated as national heroes for boosting tech sovereignty. For instance, Wang Xingxing from Unitree Robotics was featured on the Chinese Spring Festival Gala for the third time. He is worshipped as a national hero. It really depends on the industry and generation.

Kyle (25:07)

That’s so interesting. When you align with the national program and don’t stick your neck out too much—maybe by not appearing on every podcast talking about AGI—you can find a sweet spot where you do well without incurring the wrath of the state.

I want to pivot to a piece you wrote recently that really took off: the “China Tech Canon” for Asterisk Magazine. Could you explain what the China tech canon is? What books influence China’s tech community? Does it overlap with the Silicon Valley canon, and how does it feed into their different worldviews and attitudes toward technology?

Afra (26:33)

That’s a great question. Part of my obsession is understanding the different cognitive frameworks operating in Silicon Valley versus China. Last year, I read articles about the Silicon Valley canon initiated by people like Patrick Collison. When I looked at the book list, I realized many of them were actually China tech canons as well—books recommended by Lei Jun, Zhang Xiaolong (张小龙—the father of WeChat / 微信), and Wang Xing. If people are so fond of understanding the ideas driving Silicon Valley founders, what is the Chinese counterpart?

Afra (27:50)

The first story that comes to mind is about Lei Jun. He frequently tells the media that he started dreaming of becoming a founder as a college student at Wuhan University (武汉大学) in 1987 after reading a book called Fire in the Valley (硅谷之火). It’s about the 1970s homebrew hacker culture that took over Silicon Valley and birthed companies like Microsoft, Apple, and IBM. As a 21-year-old, he fell into an excited insomnia, running around the school playground because the book fueled his imagination of what he could build. Lei Jun is famously known for mimicking Steve Jobs by wearing black turtlenecks, earning the nickname “Lei Busi” (雷布斯)—a combination of his name and Steve Jobs’ Chinese name, Qiao Busi (乔布斯).

Afra (30:00)

Stories like this are everywhere among Chinese founders. Many fell in love with technology and entrepreneurship because they were inspired by Silicon Valley. There is a huge overlap because Chinese founders voraciously consume Silicon Valley’s “holy bibles.” For example, Wang Xing loves Peter Thiel’s Zero to One. He constantly asks his employees Thiel’s contrarian question: “What important truth do very few people agree with you on?” Wang Xing’s drive to make Meituan a gigantic super-app is rooted in Thiel’s belief in monopoly over competition. Zhang Xiaolong loves Kevin Kelly’s Out of Control, a 90s book arguing that technology has its own will and trajectory. He required all WeChat product managers to read it. Ben Horowitz’s The Hard Thing About Hard Things, Ray Dalio’s Principles, James Collins’ Built to Last, and The Lean Startup are widely read in China. Zero to One sold more copies in China than in the West. Elon Musk’s 2013 biography sold phenomenally well. In 2014, Beijing subway Line 4—the route taken by many tech workers—was plastered with ads for Musk’s book, The Iron Man of Silicon Valley (硅谷钢铁侠). That was right when China started its serious endeavor into EVs, back when BYD (比亚迪) was still viewed as a joke. Reading that book inspired many to jump into the nascent EV industry to become China’s Iron Man.

Afra (35:10)

However, there is an asymmetry. While Chinese founders consume Silicon Valley texts, Silicon Valley doesn’t do the reverse, creating a cognitive gap. I would argue there are other canons less visible to Silicon Valley, which I categorize as the “Red Canon” (红色经典) and “Gray Canon” (灰色经典).

The Red Canon includes Maoist texts. Aside from parroting Silicon Valley ideals, Chinese founders often use pragmatic, Maoist strategies as tactical manuals. There’s a culture of reading Mao’s classic essays from the founding period of the Communist Party. Mao is sometimes seen as the greatest entrepreneur in Chinese history, forging a party and its spirit much like building a company culture. Earlier founders like Ren Zhengfei constantly use Maoist language. Furthermore, conquering the rural, trickle-down market (xiachen shichang / 下沉市场), as Pinduoduo (拼多多) did, is seen through the Maoist strategy of “encircling the cities from the rural areas” (农村包围城市). It’s about massive grassroots mobilization.

Afra (40:00)

Next is the Gray Canon. Chinese tech founders frequently mention reading Confucian (儒家), Daoist (道家), or Legalist (法家) classics when experiencing spiritual crises or dilemmas. They turn to ancient Asian texts, like Wang Yangming’s (王阳明) Philosophy of the Heart (心学) or the Dao De Jing (道德经), for strength and self-help. In the West, Gary Tan might advise founders to read Carl Jung or go to therapy. In China, while mental health is emphasized, many turn to texts like the Analects of Confucius (论语)—a brilliant guide on dignity, sense-making, and values. Just as Silicon Valley founders reference ancient Greek or Roman philosophy, Chinese founders return to their ancient philosophers to understand statecraft, bureaucracy, honor, and society.

Afra (44:00)

Lastly, there are mythological and sci-fi readings. Just as Silicon Valley founders are obsessed with Tolkien and Asimov, Chinese tech founders love Jin Yong (金庸) and Liu Cixin (刘慈欣). Jack Ma is a massive fan of Jin Yong’s wuxia (武侠 - martial arts) novels. In the early 2000s, he created a tech conference called “Swordsmanship Competition by the West Lake” (西湖论剑). Jack Ma repeatedly said that without Jin Yong, Alibaba wouldn’t exist. He sees the tech world as a jianghu (江湖 - martial arts world). Alibaba employees use nicknames, and Jack Ma’s nickname, Feng Qingyang (风清扬), is taken from a legendary swordsman in a Jin Yong novel. He even named his office “Peach Blossom Island” (桃花岛) after a location owned by a brilliant, righteous, yet arrogant master, Huang Yaoshi (黄药师), in the novels.

Another essential text is Liu Cixin’s The Three-Body Problem (三体). It’s so influential that management theories are named after concepts from the book, like the “Dark Forest Theory” (黑暗森林法则) or the “chain of suspicion” (猜疑链). It provides a perfect cultural reference pool to describe strategic business scenarios. Together, Jin Yong and Liu Cixin give Chinese technologists an imaginative toolkit, making the ecosystem more vibrant and humanistic.

Kyle (50:14)

It’s super interesting how these texts form the foundation for a common language and approach to tackling both pragmatic business strategies and deeper introspection.

Related to this, I was wondering if you could talk about another text you wrote about for Wired magazine: Morning Star of Lingao (临高启明). It’s a fascinating, crowdsourced sci-fi novel online with millions of words contributed. In a nutshell, a group of time travelers goes back to the Ming dynasty (明朝) to help China industrialize before the West. What is the significance of this text, and what does it say about Chinese attitudes toward technology and industry?

Afra (51:46)

That piece was a historical probe. I wanted to take the audience back to around 2006 to 2010 to understand the fabric of Chinese internet discussions back then. Unlike today, people predominantly talked about politics. Morning Star of Lingao was an experiment in political imagination. I recently interviewed one of the core writers, Ma Qianzu (马前卒). He participated in this collective science fiction writing because he was dissatisfied with China’s hard and soft power at the time, as well as its leadership. He believed that if you created a vacuum world and let people who understood the highest form of productive forces rule, you could build a better political structure and ideal society. Ma Qianzu was a civil engineer, and he injected his worldview into the story. The writers—many of whom literally built modern China’s infrastructure—believed that building things and industrialization have the ultimate righteousness.

Afra (54:40)

These writers, who were hardcore Marxists, formed a loosely connected intellectual group called the “Industrial Party” (工业党), a term coined by the nationalist scholar Wang Xiaodong (王小东). This was in opposition to the “Sentimental Party” (情怀党), who sounded more like today’s leftist intellectuals promoting degrowth ideologies—arguing that rapid development should pause to address the harms it caused. The Industrial Party organized intellectual rebuttals against them.

Kyle (56:17)

Just to clarify, they were not a political party, but rather a group of people, right?

Afra (56:21)

No, they were just a group—like “Silicon Valley AI Twitter.” There’s only one effective political party in China, as we all know. The core ideal of the Industrial Party, which relates to your piece on “industrial maximalism,” is that the greatest source of strength and strategic asset a nation can have is its industrial capacity. Industry isn’t just factories; it’s the tacit, embodied knowledge embedded in the organizational system. Wang Xiaodong famously said, “Let the American people sing and dance for us. Our country’s greatest strength lies in how much copper you melt and how much iron you forge.”

We are now living in a world where that prophecy came true. China has built a comprehensive industrial system, producing nearly all of the United Nations’ listed manufacturing categories. Industrial maximalists believe this complete supply chain is China’s equivalent to America’s dollar hegemony or the Middle East’s energy reserves. Lingao crystallizes this collective unconsciousness and worship of national manufacturing strength.

Kyle (1:00:18)

I’ve been toying with the phrase “the revenge of the real world.” It wasn’t so long ago that Marc Andreessen said “software is eating the world.” And he was right—we saw the rise of American tech giants and digital platforms acting almost like nation-states. But now, with supply chain vulnerabilities, trade wars, critical minerals, and semiconductor export controls, the real world has come back with a vengeance. The Industrial Party would argue this is where China will thrive because they invested in capabilities that the US let wither away. It fuels anxieties in the West about trade and job loss, but this idea is so dominant now.

Afra (1:01:47)

Yes, I totally agree. There’s a status game in China where working in manufacturing, semiconductors, or robotics garners more respect from the state. When Xi Jinping invites tech founders for talks, the hardware people like Ren Zhengfei sit in the middle, while the software and AI people like Pony Ma sit at the edge.

I vividly remember speaking to my cousin’s husband in Xi’an (西安) last year. He works for a prestigious state-owned steel company. When I mentioned AI, he shrugged it off as a gimmick compared to the massive scale of steel processing he handles. He viewed AI as something for the “fancy little circles” in Beijing and Shanghai, while the real heroes are the pillars of Chinese industrial power. Hardware workers are genuinely worshipped.

Kyle (1:04:21)

That’s super interesting. It brings us full circle back to Marxist materialism—material power as the ultimate foundation.

Afra (1:04:29)

Exactly. For example, Zhang Yiming and ByteDance might be driving the best video generation technologies right now, but they aren’t deeply respected by the broader Chinese tech circle or trusted by the state. ByteDance is sometimes seen as a distraction, and Zhang Yiming is rarely invited to meaningful party conferences.

It’s an interesting asymmetry. Silicon Valley used to focus only on its software counterparts and overlooked China’s hardware strengths. But now, seeing the push to reindustrialize America, Silicon Valley is trying to catch up. Reading pieces from a16z’s “American Dynamism” feels like they are trying to reverse engineer China’s advantages in areas like BYD’s electronic stack, grid capacity, and solar cell manufacturing. They are feeling the revenge of reality and starting to catch up.

Kyle (1:06:38)

Right. It’s like a hall of mirrors. Chinese tech entrepreneurs admired Silicon Valley greats, and now Silicon Valley is borrowing pages from the playbooks of Chinese tech companies.

We could talk for hours, Afra. You are one of my favorite writers on these topics, and it’s a huge honor to have you on the podcast. How can listeners follow your work?

Afra (1:07:31)

The go-to place is my Substack newsletter, Concurrent. I named it that because I see China and Silicon Valley as two innovation engines concurrently creating interesting futures. I also freelance for some publications, and you can follow me on Twitter at @Afrazhaowang.

Kyle (1:08:08)

I will definitely include links in the show notes. I highly recommend Concurrent and will link to the pieces we mentioned. Thank you so much, Afra, for an amazing conversation.

Afra (1:08:22)

Of course, Kyle. I’m also a big fan of your newsletter, High Capacity. You pick up on the overlapping Chinese tech realities with a lot of clarity and authenticity. It’s easy to fall into clickbaity “China is winning everything” tropes or hawkish tropes, but you maintain an authentic voice without being lured into either camp. I really respect that.

Kyle (1:09:11)

Thank you. That’s very nice of you to say. We’ll wrap up here. If you liked this episode, please rate and subscribe on YouTube, Spotify, or Apple Podcasts. You can find episode transcripts and more information on the High Capacity newsletter at high-capacity.com. I’m your host, Kyle Chan. Thanks for joining, and see you next time.

Watch or listen to the High Capacity podcast on:

• YouTube

• Apple Podcasts

• Spotify

In this episode, I speak with Lei Xing, former editor-in-chief of China Automotive News and co-host of the China EVs & More podcast with previous guest Tu Le. Lei Xing is a world expert on China’s EV and autonomous driving space.

Key takeaways:

• China’s AV industry is supported by central and local government policy. Different Chinese cities have taken different approaches. Wuhan is more proactive in expanding robotaxis while Beijing is more conservative.

• China’s smart driving players are taking different approaches. Some are moving up the ladder from “driver assistance” systems to L4 autonomy. Some are starting with robotaxis and then moving into private passenger cars. The LiDAR vs. camera-only debate is ongoing.

• Many Chinese companies are taking a full-stack approach to autonomy. They’re developing their own semiconductor chips and their own AI models.

• But the industry is also taking a mix-and-match approach. Many Chinese EV companies are using Nvidia chips and partnering with US mobility companies like Uber and Lyft. Western automakers are also relying on Chinese smart driving components.

• China is moving fast on autonomous delivery and commercial trucking. Autonomous delivery vehicles have become commonplace in parts of China.

Follow Lei Xing:

• Twitter / X

• LinkedIn

• YouTube

• China EVs & More podcast

Transcript

Kyle (00:00)
Welcome to the High Capacity Podcast. I’m your host, Kyle Chan, a fellow at Brookings. I’m thrilled to be joined today by my guest, Lei Xing, a world-class expert on Chinese electric vehicles and autonomous driving. He’s the former editor-in-chief of China Automotive Review and a co-host of the podcast China EVs and More with Tu Le, who was previously on the show. In that episode with Tu Le, we looked at China’s overall EV landscape. This time with Lei Xing, we’ll zero in on China’s push into autonomous vehicles and robotaxis. We’ll look at how China’s autonomous driving technology is evolving, how the Chinese government is supporting the industry, and how Chinese companies are scaling up robotaxi fleets around the world. Welcome, Lei Xing, and thank you for coming on the show.

Lei Xing (00:52)
Thank you for having me, Kyle. It’s nice to meet you through this platform and talk about AVs.

Kyle (01:02)
Absolutely. It’s a hot topic.

To begin, I’d like to focus on making sense of the technology. There are a lot of terms thrown around—smart driving, autonomous driving, ADAS, robotaxis, different levels of autonomy: L2, L2+, L4. For those who don’t follow the industry as closely as you do, what do these terms mean, and how do they differ in China?

Lei Xing (01:40)
This question alone could take hours, because the terminology is part of the narrative. “Smart driving” is definitely a Chinese term. In Chinese it’s zhijia (智驾). In China, “smart driving” usually points to ADAS or L2++—driver assistance—meaning you have to pay attention at all times, and the driver is still responsible in case of accidents.

“Smart driving” also makes it more appealing as a marketing term. Zhijia sounds better than “advanced driver assistance system,” which is jargon. But in practice, a lot of it is Level 2—sometimes called Level 2+, Level 2++—and “point to point,” which is where Tesla’s FSD is today. And Tesla still says in the fine print that you have to watch the road and pay attention.

The major difference comes when you go to L3 and L4, when responsibility shifts toward the OEM or the company—like a Waymo. In China, most recently, two vehicle models were approved to pilot Level 3 operation. Put simply, Level 3 is when you can take your eyes off. You can do other things. And that’s when, if the feature is commercialized, responsibility transitions to the OEM.

Level 4 used to mean “driverless,” like Waymo, Cruise, Baidu, Pony.ai, WeRide—we’ll talk about those. Historically, it wasn’t something a private consumer could buy, but you could use it as ride-hailing, either on their own network or through platforms like Uber or Lyft.

In recent months, there’s been talk about consumer Level 3 and Level 4, meaning you could actually buy a vehicle with Level 4 capability.

Kyle (04:52)
So these are like the robotaxis?

Lei Xing (05:10)
Eventually you can put that vehicle into a fleet, which is what Tesla is trying to do. FSD Supervised is essentially L2++. When you go from supervised to unsupervised, then it’s Level 4.

So now there are two tracks. Chinese companies like Xpeng are a good example. They’re pushing Level 4 consumer capability—what they’ll call “Robo.” They have Pro, Ultra, and then next is Robo, meaning Level 4 capability you can buy. At the same time, they’re developing robotaxis for fleets that you cannot buy.

That’s a recent and interesting change in direction for what’s possible in autonomy.

Kyle (06:12)
Yeah, the space is changing quickly. You have players that began with Level 2 “smart driving” features on consumer vehicles and are moving up. You also have companies that were mainly robotaxi-focused from the beginning. We’ll come back to that.

Before we dive into all of that, could you talk about Chinese government policies to support the industry—regulation, central versus local policy—and how they’re balancing safety and progress?

Lei Xing (06:41)
When you talk about China, you have to talk about roadmaps and overarching policy. I’d point to “Made in China 2025,” announced many years ago. There was a roadmap for EVs. If we bundle EVs, smart EVs, and autonomous vehicles together, in China it’s called the “intelligent connected vehicle” (ICV). ICVs encompass EVs, AVs, and everything else.

So you have that overarching roadmap—China wants to lead in these verticals: EVs, AVs. Then in recent years you’ve had a series of policies on safety. For example, we saw cybersecurity and data transfer rules announced this week: what could happen if a company like Tesla wants to use data collected in China outside China. That’s not currently allowed on both sides, but rules are coming into place that create pathways if you meet specific requirements.

I also mentioned the Level 3 pilot operation permits for the first two models. That’s on the national level. Locally, in tier-one cities like Beijing, Shanghai, Guangzhou, and Shenzhen, there are local pushes to pilot robotaxi operations. Pony.ai, for example—one of the big three robotaxi companies—is commercially operating in all four tier-one cities. They’re the only one currently, and that’s because of local policies supporting rollout.

Take Beijing as an example: permits are phased. First you test on roads with safety drivers. Once you reach milestones and accumulate experience and data, you move to the next permit—then you can have a safety driver as a passenger. Then you remove the safety driver completely. Baidu has achieved 100% driverless operations in the cities where they operate, but they had to go through testing and pilot operations before offering paid rides.

These local policies specify stages and thresholds for moving to the next step. So national and local policies work in tandem.

You also have requirements on the vehicles themselves. Years ago China was already piloting vehicle-to-road connectivity. Some cities—like Wuxi near Shanghai—have been test bases for vehicle-road communications. It’s a staged approach: communications, vehicles, testing, pilot operations—and now scaling. The policies make these developments possible. The overarching roadmap and local execution working together is important in China, and that’s true for many industries.

Kyle (11:52)
Yeah, definitely. This is a big theme for the whole podcast. And some of these things depend on the city—Wuhan versus Shenzhen versus Beijing. Beijing might be more restrictive for certain reasons.

Lei Xing (12:15)
Right. Wuhan is more lax—more willing to bring AVs into the city and run them. Beijing is the political capital, so it’s more restrictive. When companies like Baidu and Pony.ai get permits there, it’s tougher and more limited geographically—mostly certain areas like Yizhuang in southeast Beijing. They’ve also started operating at Daxing Airport and one of the train stations.

Wuhan is much broader—pretty much the whole city. So there are regional differences even within China—policies and support vary. It’s similar to the U.S.: Arizona is more permissive than other states, which is why Waymo runs in Phoenix and why Uber did a lot of piloting there in the past.

Kyle (13:41)
That’s really interesting. On robotaxis: what’s the experience like for an average urban resident in a major Chinese city? How do they summon one? What’s it like to enter? Is there a safety driver or not? Can you change your destination mid-ride? Does it feel like a Waymo, or is it different?

Lei Xing (14:15)
Yes to all of that. The latest robotaxi I took in China was in September, in Guangzhou, with WeRide. They’re running the GXR, a Geely van-style robotaxi.

You can hail these in several ways: through their own app, through WeChat mini programs, and through maps like Baidu Maps or Amap—because they’re integrated. In China, it’s very seamless.

The ride experience varies by vehicle type—comfort level, whether the driving is aggressive or jerky. I’ve ridden multiple platforms in China and in the U.S. For me, when I get into a robotaxi, I expect it to work. I don’t get nervous. It feels like another vehicle: you get in, you ride, you arrive.

If I had to rank comfort, Baidu’s latest RT6—minivan style—is very comfortable. Pony.ai runs a lot of Lexus RXs, and they’ve updated to newer generations. Baidu used to have cheaper vehicles that weren’t as comfortable.

Globally, I’d still call Waymo a gold standard. I’ve taken many rides in Jaguar I-Paces, which are quite comfortable. The best description is that they’re boring—which is good. They just work, and you don’t feel much.

In China, platforms like WeRide, Pony.ai, and Baidu have gotten to that point. But in terms of coverage and scaling rapidly, people still watch Waymo. We just saw their latest financing round and scaling plans. They want to be global.

But AVs are still regional. Chinese companies are arguably more global right now because they’re going into the Middle East and Europe. Waymo is moving toward places like London and Tokyo.

For consumers, the appeal is privacy—no other person in the vehicle—and then affordability, wait time, and ease of hailing. For robotaxi companies, the big questions are scaling and breaking even. Most now have fleets over 1,000 vehicles. By the end of this decade, many are targeting six digits—hundreds of thousands.

Kyle (18:42)
Things are moving fast. We’re just getting started on global expansion—robotaxi services coming out of the U.S. and China. A lot of them are using mixed strategies. Uber is partnering to offer WeRide in some cities. Waymo uses Zeekr EVs, or is starting to use them more. That could help drive down costs, especially outside the U.S. where Chinese EVs are heavily tariffed. So there’s an interesting mix of technologies—it’s not purely one tech stack versus another.

Lei Xing (19:29)
The simple debate is vision-only versus LiDAR. You have Tesla’s route with FSD, eventually robotaxi, versus most others using LiDAR. The jury is still out. There are proponents on both sides.

Kyle (19:32)
So Tesla uses cameras only, and the others use LiDAR?

Lei Xing (19:54)
That’s why we look not only at robotaxi companies, but the whole supply chain. And these companies have different strategies: focus purely on Level 4, or branch into Level 2++ “smart driving.”

Strategically, WeRide and Pony.ai differ. WeRide has moved into Level 2++ along with Bosch, while Pony.ai is still focused on pure Level 4 robotaxis. Then you have Momenta, which started with Level 4 but is now embedded in many Chinese and foreign brands in China for point-to-point Level 2++. At the same time, they’re working with Uber to launch robotaxis in Munich, and with Mercedes to launch both Level 2+ in China and robotaxis outside China.

Beyond that, there are mobility platforms: outside China, Uber and Lyft; inside China, DiDi, CaoCao, OnTime, T3, Meituan, and others. Hello Bike is moving into robotaxis. CaoCao Mobility—backed by Geely—has a goal of deploying 100,000 robotaxis.

DiDi has a joint venture with GAC Aion and recently produced a dedicated robotaxi. DiDi is a bit late. And then there are chip companies, compute, and other suppliers.

Even Lenovo is involved. They have a vehicle computing unit providing a domain controller with Nvidia Thor for the WeRide GXR robotaxis.

Kyle (22:39)
You mentioned LiDAR. Going back to the camera versus LiDAR debate: LiDAR used to be much more expensive, which is why Tesla leaned toward cameras only. Waymo has used LiDAR and radar, then refined its sensor suite over time. But LiDAR costs have dropped dramatically, and Chinese players seem to be leading. Can you say more?

Lei Xing (23:23)
There are “big threes” in each vertical. For robotaxis, there are big three companies; for LiDAR, there are big three companies too. The main leaders are Hesai, RoboSense, and another major player. Hesai and RoboSense are the biggest rivals. They’ve put LiDAR on vehicles at scale and reduced cost into the hundreds of dollars, compared to thousands just a few years ago.

At CES three years ago, the industry talked about the potential for a million LiDAR units annually. Now those two players alone are already delivering a million LiDAR units annually.

Part of it is that LiDAR is a selling point. Early adopters like NIO and Li Auto put roof-mounted “showerhead” LiDARs on vehicles. Now, with consumer Level 3 and Level 4 discussions—and robotaxi scaling—the addressable market is much larger.

Hesai, for example, doubled production capacity from 2 million to 4 million LiDAR units per year because Level 3 and Level 4 vehicles may require multiple LiDARs—top and sides. And foreign automakers and Chinese EV brands will standardize LiDAR on lower-priced vehicles because costs have dropped.

Kyle (26:09)
Right.

Lei Xing (26:09)
They can keep prices steady while adding LiDAR as a feature. Watch BYD this year—LiDAR will be a big part of their strategy to offset pressure. When you’re number one, it’s the most dangerous position. One of BYD’s tactics this year is pushing better features in their God’s Eye ADAS and standardizing LiDAR on cheaper vehicles.

Kyle (27:11)
That’s interesting, because some people think of LiDAR as a premium feature you only put on luxury models. But you’re describing economies of scale from standardizing across more models—even affordable ones.

BYD offering its God’s Eye smart driving system on vehicles below $20,000 is a capability people might expect only in higher-end cars.

Lei Xing (27:55)
Right. It’s not a premium thing anymore. It started as premium hardware, but it’s commoditized in China.

Kyle (28:18)
And there’s almost a floor in terms of safety, especially in difficult weather.

Lei Xing (28:32)
One policy we didn’t mention is compulsory AEB kicking in January 2028. Automatic emergency braking will be required on all vehicles in China—not only passenger vehicles, but also some light commercial vehicles. Companies like Huawei and Li Auto have said they can do AEB up to 120 kilometers per hour, and to do that, LiDAR is required, because at night you need long-range detection—300 to 400 meters. So policies like this expand the addressable market.

Kyle (28:42)
What is AEB?

Lei Xing (28:56)
Automatic emergency braking.

Kyle (29:29)
You mentioned Huawei. What is Huawei doing in this space? They seem to be everywhere, but they’re not a conventional EV or AV company.

Lei Xing (29:45)
Huawei officially entered the automotive sector at the April 2019 Shanghai Auto Show. We all know why: restrictions pushed them to expand into other verticals, and autos became one.

Huawei isn’t an OEM. They’re more like Xiaomi or Apple, but without making their own cars. Through the Harmony Intelligent Mobility Alliance, they work with state-owned automakers. They’re active across ideation and design of premium vehicles, and deep into hardware and software: components, their own LiDAR, HarmonyOS, e-motors, charging—many branches supporting the auto industry.

In China, “Huawei” pulls consumers. Their consumer products—phones, foldables—created a strong premium tech association. Vehicles like Aito and Luxeed are designed with features and quality tuned to Chinese consumer demand, and in many ways they’ve outpaced what foreign premium brands offer.

There’s also a national sentiment factor. Huawei positions itself as an integrator and facilitator helping state-owned automakers. Adding Huawei is like giving them wings—Huawei is the “Red Bull” for these automakers. Sales have taken off.

Kyle (32:42)
It’s the Huawei factor. The premium association, and integration with Huawei’s ecosystem—phones and other devices—adds something beyond the vehicle itself.

Lei Xing (32:56)
Take Aito M9: it’s priced above 500,000 RMB. For every ten vehicles sold in China above that price point, seven are Aito M9s. The value is hard to pass up. A Porsche at that price point isn’t really an option in China—Porsches are usually over a million RMB. Many buyers are upgrading, and Huawei delivers.

Kyle (34:15)
You had a great chart showing Porsche struggling in China. For a long time, it was the premium brand for successful businesspeople, but now market share is shifting toward Chinese brands—even in premium segments. It’s not just a race to the bottom.

Lei Xing (34:44)
And bringing it back to today’s topic: the ADAS L2++ race is intense, especially for point-to-point capability. We can put Huawei in the top tier right now for Level 2++ point-to-point.

Kyle (34:58)
I wanted to ask about AI and autonomous driving—models powering these systems and how approaches are shifting. In the U.S., we’ve seen a shift from hard-coded modular systems to end-to-end training. Do you see something similar in China? What’s happening on the model side?

Lei Xing (35:54)
Yes. Over the last one to two years, things have changed dramatically. Previously it was more rule-based. Now the buzzwords are VLA, LLM.

Xpeng is pushing out its second-generation VLA later this year. They want to reach Tesla FSD’s level—FSD Supervised, version 14-point-something. In retrospect, Tesla is still the gold standard for many Chinese companies: vision-only and a strong model.

Xpeng has even ditched LiDAR for its point-to-point Level 2++. They want to move to vision-only. Xpeng is often seen as China’s Tesla—they try to follow Tesla’s direction.

VLA—vision, language, action—and more generally “input in, output out,” end-to-end. For that, companies are moving into their own SoCs so they can optimize compute and capability rather than relying entirely on third parties.

NIO has its model, Li Auto has its approach, Huawei has its model. I experienced Li Auto’s system in September in Beijing. The moment you leave the parking lot, you press the button and it goes. During the trip, you can say, “At the next intersection, turn right,” even if the navigation was set differently, and it responds and adjusts.

This is what Jensen Huang talked about at CES—thinking and reasoning models. In China, Xpeng, Li Auto, NIO, Huawei all have their own models, but the direction is end-to-end. If vision-only works, the jury is still out. For true Level 4 robotaxi, LiDAR may still be needed.

We’re also seeing integration with broader models—DeepSeek, Doubao, and others—being incorporated into experiences. But Tesla is still a major benchmark.

Kyle (40:49)
It’s interesting to see convergence: adding reasoning and language capabilities, not just reacting to traffic patterns, but interpreting what people might do and responding appropriately.

You mentioned SoCs—self-developed chips. Can you say more about how much Chinese EVs and AVs rely on Nvidia chips versus domestic suppliers?

Lei Xing (41:50)
Still a lot of reliance on Nvidia. Thor and earlier chips are in many Chinese EVs, and it’s a selling point. Qualcomm is also moving into autonomous driving; previously their chips were mainly for smart cockpits. Domestic chip companies include Horizon Robotics and Black Sesame.

The future is single-chip integration—one chip that can handle both cockpit and driver assistance. Xpeng’s Turing chip is an example of that direction.

WeRide’s GXR robotaxi uses Lenovo vehicle computing with Nvidia Thor in the domain controller. Many vehicles on sale have dual Orin-X chips.

Some companies pick different strategies. Li Auto, for example, uses Qualcomm heavily for both smart driving and cockpit. Others think more about cost. There’s also geopolitical risk considerations—like access to high-end chips—and de-risking. Xiaomi is developing its own SoC too, extending from phones into EVs.

At the same time, it’s not the end of globalization. Capability and “make vs. buy” decisions differ across companies.

Kyle (45:37)
So many are pursuing full-stack approaches—self-developed chips and models—while still integrating with suppliers.

Lei Xing (45:58)
And this is where Mercedes and Volkswagen come in. They can’t build everything in-house, so they buy. Mercedes is working with Chinese partners and with Momenta. Volkswagen is working with Xpeng for EV architecture. Volkswagen is also the first foreign customer of Xpeng’s SoC—publicly announced—at least in the Chinese market.

This is the contrast: it used to be Chinese companies depending on foreign firms, and now it’s flipped—at least in China. That’s part of why China is a global frontier in this space.

Kyle (46:58)
Yeah, it’s flipped—especially in EVs and autonomous driving, where Chinese technology is sought after, from CATL’s batteries to Momenta to EV startups like Xpeng.

Looking ahead: with geopolitics, do you see a fragmented global market for robotaxi and smart driving tech? And what are the chances we’d see Chinese AV players in North America?

Lei Xing (48:10)
As we speak, Tesla is testifying with the Senate Commerce Committee on self-driving regulations. Yesterday, the CEO of NADA said their association would support blocking Chinese auto manufacturers from entering the U.S. They won’t control what individual dealers do, but that’s the stance.

For robotaxis, being truly global is difficult. Chinese “big three” are going into Europe and the Middle East. Mobility companies like Uber and Lyft aren’t in China, because China has DiDi, CaoCao, OnTime, T3, Meituan, and many others.

So it’s more regional, but there are neutral locations where multiple players can operate. The U.S. is not one of those right now. We can’t buy NIO, Xpeng, Li Auto, BYD—at least not yet. Chinese AV companies have tested in California and Silicon Valley, but there’s more scrutiny now.

Even LiDAR companies have U.S. presence: RoboSense has an office near Detroit, Hesai is based in Silicon Valley. Baidu had a U.S. office years ago; in 2018 they were showing ride-hailing robotaxi concepts, but those efforts didn’t go far.

The U.S. is a tough frontier. But there are workarounds. Waymo uses Zeekr vehicles—essentially the base vehicle—and then retrofits sensors in the U.S. So you can ask whether something could happen the other way around. I’m not sure, but it’s possible in some form.

In the near term, companies expand where operating conditions are favorable—dry weather, supportive policy, and funding. That’s part of why the Middle East is important. Europe too, in certain regions. WeRide and Pony.ai are already launching commercial services in Abu Dhabi and Dubai.

Can Chinese robotaxis operate in the U.S.? I can’t see that in the near future, just as Waymo is not in China. For the foreseeable future, the home market is the biggest market, but everyone wants overseas expansion to be called “global.”

We also have more players now. It used to be Cruise, but Cruise is no longer around. Now there are new partnerships and players—Plus there’s Zoox coming into play. In China, more local players keep entering too. The space is still expanding, and the end game is still murky.

Kyle (53:36)
One last question: commercial and delivery autonomous vehicles. We’ve talked about passenger vehicles, but there’s also trucking and last-mile delivery. What’s happening there?

Lei Xing (54:04)
This is another area where China is far along. One company is UIC—its CEO is a friend of mine. Their company, Cargobot, works on platooning for autonomous trucking. The founder was an engineer in the U.S. at Delphi. They’ve been operating platooning in Inner Mongolia—delivering goods and coal. In a platoon, you might have one human driver, but the other trucks don’t have drivers.

For last-mile delivery, there are companies like Neolix and Zelos—two major players for small autonomous delivery vehicles. They were at CES. People are often surprised by how many are already running in China—tens of thousands. They’re already doing real work.

When I go back to Beijing, I stay in Shunyi district. In that area, SF Express, Meituan, and others have autonomous delivery vehicles integrated into daily life. They’re just part of the flow with bikes and cars—no one is shocked by them anymore. In Yizhuang, where JD Logistics has its headquarters, these vehicles run around for last-mile delivery—including groceries.

In the true “last centimeter,” you still have human couriers delivering to apartments, especially for fast food, but they often take bags from these autonomous vehicles.

UIC’s niche is moving luggage at airports—pickup and drop-off. They’re already in Hong Kong, and I think in Qatar, and other airports. These are companies that are only about ten years old. Like EV startups, many are 10 or 11 years old. Their niches differ, so they don’t have to compete head-on. They find a niche, scale, and grow. And there are many more beyond the ones I’ve mentioned.

Kyle (57:59)
There are so many startups. I’d love to go to an airport and have a robot take my luggage—especially when you’re traveling with kids and don’t have an extra hand.

This has been awesome. If people want to learn more about you and your work, where should they go?

Lei Xing (58:45)
I’m on several platforms. I co-host China EVs and More. I have a Medium channel, and I do some YouTube—videos from events and my thoughts. You can find me on YouTube, Medium, LinkedIn. My handle is leixing77.

And one thing I want to correct: I don’t consider myself an expert. I consider myself an enthusiast.

Kyle (59:26)
You’re being too modest.

Lei Xing (59:28)
Because there’s so much to learn, and things are moving fast. The verticals—EV, AV, mobility, robots—at the end, it’s all one thing: automobile, auto, autonomous, mobile. I’m still learning myself. I’m glad I’ve covered the industry for so long, and there’s still so much to learn.

The people you meet along the way—Tu, you, founders of these companies—some I know personally. It’s fascinating to be part of this revolution. And it’s great we can communicate and educate through these platforms. We try to be fair.

Kyle (1:00:30)
Yeah. In our lifetimes, it’s very cool to watch.

Lei Xing (1:00:46)
We don’t want to be sensational. We can’t only say positive things—we have to be critical. Bad things happen too. And it’s interesting to watch, to be on this journey.

Kyle (1:00:50)
Yeah, exactly. We’re all learning together. I’ll include links to where people can follow you. Thank you so much, Lei Xing, for an awesome conversation.

Lei Xing (1:01:23)
Thank you for having me.

Kyle (1:01:25)
If you liked this episode, please rate and subscribe on YouTube, Spotify, or Apple Podcasts. You can find episode transcripts and more information on the High Capacity newsletter at high-capacity.com. I’m your host, Kyle Chan. Thanks for joining, and see you next time.

Watch or listen to the High Capacity podcast on:

• YouTube

• Apple Podcasts

• Spotify

In this episode, I speak with Grace Shao, founder of the AI Proem newsletter and host of the Differentiated Understanding podcast. Grace is an expert on China’s AI and tech industry, with years of experience working with Chinese tech companies like Alibaba, Lenovo, and Kuaishou.

Key takeaways:

• Chinese AI companies are moving quickly into AI agents. Beyond mere chatbots, these AI assistants can do useful, real-world tasks such as buying concert tickets or ordering food delivery.

• China’s big tech firms are pursuing different AI strategies. Alibaba is integrating agents across its closed ecosystem of services. Tencent is integrating AI into its popular WeChat platform. ByteDance is pushing toward AI-native interfaces, including phones.

• China’s AI startups are highly capital-constrained. The recent IPO valuations for Zhipu and MiniMax were only around $6-8 billion. This forces them to pursue different strategies, such as low-cost models or specific industry verticals.

• Countries like Singapore are a geopolitical gateway for AI. A number of Chinese AI companies are using Singapore as a launching pad for their global strategies. The case of Manus, recently acquired by Meta, is a striking case study.

• One area to watch is AI hardware and devices. China is moving fast on AI wearables, such as smart glasses, and supplying robotics components for the rest of the world.

AI Proem pieces by Grace Shao:

• China and the US are Running a Different AI Race

• AI Strategy Convergence: US and China Are Meeting in the Middle

• ‘Mens et Manus’, the story of Manus. Its brains and hands.

Transcript

Kyle (00:00)
Welcome to the High Capacity Podcast. I’m your host, Kyle Chan, a fellow at Brookings. I’m thrilled to be joined today by my guest, Grace Shao, founder of the fantastic AI Proem newsletter.

Grace is based in Hong Kong and one of the sharpest analysts of China’s AI and tech sector out there, looking not just at the tech itself, but at what’s happening in terms of industry strategy. She spent years covering China’s tech, businesses, and economy as a reporter, and then worked and consulted for some of the largest tech companies like Alibaba, PayPal, Lenovo, and Kuaishou. She also has an awesome podcast of her own called Differentiated Understanding on AI and tech in China, which I highly recommend. Welcome, Grace, and thank you for coming on the show.

Grace Shao (00:47)
Hi Kyle, thank you so much for having me today. And you were just on my show, so people should give it a listen. That was a great episode. I’m very flattered, and again, thank you for having me.

Kyle (00:58)
I’m glad we can return the favor to each other. So just to start off, I was wondering if you could tell our listeners a bit about you, your background, your experience with China’s tech sector, and what inspired you to start AI Proem.

Grace Shao (01:16)
Yeah, I think most people don’t know this about me, but I actually studied finance and first interned at a hedge fund covering China’s TMT sector. And that’s what really brought me to Asia first. I came to Hong Kong in 2014 and really just love the energy here. But while I was working in that role, I realized I really wasn’t good at crunching through numbers at all. And truly I was a storyteller at heart. I think I really always wanted to kind of combine my interest in business and storytelling. So I went to grad school for business journalism, and that kind of gave me the first kind of step into China.

I worked in Beijing for three years and then I moved around to Shanghai. I later on covered China’s tech sector, the APAC tech sector out of Singapore, and now I’m based in Hong Kong. And like you mentioned, over the years—nearly a decade—I’ve gone from analyzing the tech sector as an investor to being a reporter covering the tech sector, following the latest trends, to actually being in the companies, working on their international positioning, their crisis management, working closely with the PR and IR teams, and then later on joining a consultancy and advising them from an external lens.

So today I think I bring—hopefully—a business analytical lens and a storytelling nature to my analysis at AI Proem, and it’s something that’s entertaining and insightful to readers.

On why I really started this: honestly, I think it was just personal. Like I mentioned, I’ve always loved storytelling and business, and I want to tell the business story. But I think in the last few years, there’s just been a lot of geopolitical clout around covering China. And I really wanted to create something that just focused on the businesses themselves, the founders, the entrepreneurs, the really exciting stuff that’s happening on the ground, and remove a lot of the noise in the background. And we all know what’s happening. It is what it is, right? And I want it to bridge that information gap and help people—from, I don’t like to say both sides of the world, but really, in this global environment—understand each other a bit better.

Kyle (03:24)
Yeah, it’s been amazing following your work. And yeah, I love the way that you really go behind the stories and give that concrete narrative there—like, what are they trying to do? What are their goals? Who are the personalities? There’s so much drama in this space. And when you just see it purely as the U.S. versus China in this AI race, it becomes too abstract. There’s a lot of wonderful details along the way. Speaking of which, I wanted to ask you about agentic AI. This is such a big topic these days. I mean, 2025 was supposed to be, in many ways, the year of agents. It was talked about by a lot of the major tech and AI leaders in the U.S. And maybe there was more caution about rolling out agentic capabilities—like on the iPhone or elsewhere—at scale, at least in the U.S. It seems like China is going more heavily into AI agents. It looks like a number of different tech companies are experimenting, trying out different strategies. And I was just wondering if you could maybe talk about what you’re seeing, what’s interesting to you, and who is up to what.

Grace Shao (04:40)
Yeah, for sure. I think I also drank the Kool-Aid. Like I wrote in my recent piece, my husband and I spent last Saturday night just sitting together, tweaking our own agents on Claude Code, and I thought it was so cool what it could do. I think there was a lot of hype around vibe coding over the last two years, but for someone who has no technical background, I really didn’t find a lot of the tools that intuitive until Claude Code. So for sure, I think that was a pivotal moment, and every lab in China is experimenting with something like Claude Code.

But it’s not really just an agent in that format. And I think to your point, a lot of the big tech companies are doing different things and trying to tap into their existing ecosystems.

So on a Claude Code–like product, we can see that MiniMax—one of the four tigers, and it just went public in Hong Kong earlier this year—just launched its own version last week. The agent is called MiniMax Agent, and it is free for use. It really kind of embodies the idea of an AI-native workspace. It has that vision similar to Claude Code. However, I think a lot of people are saying even though the experience might be a bit glitchier than Claude Code, the competitive advantage right now is that MiniMax M2 API pricing is at about 8% of Claude Sonnet 4.5. So there’s that obvious price differentiator for them.

And in terms of the big tech: last year, I think around November or December, Alibaba launched a Qwen app. Previously their consumer app was called Tongyi—in Chinese, Tongyi Qianwen. It just means like a thousand questions are all answered. But we all know the model is called Qwen, and last year they actually rebranded their consumer app to Qwen. So I think it was a really strong signal that, look, the consumer can now access our most frontier model. That was one on branding.

But what do they do? Well, they really tap into their ecosystem. They open up their whole series of apps. And I think for people outside of China, sometimes they don’t realize Alibaba is not just a platform to purchase, say, cheap clothing. It actually has a series of applications. For example, Fliggy—Fliggy.com—is essentially like Ticketmaster plus Booking.com. There’s an app called Gaode, which is their navigation tool like Google Maps plus Uber—you can order ride-hailing. Ele.me, its delivery app, is kind of like DoorDash, and it has so many more. And then obviously Taobao is its flagship application that lets you go into Tmall, which lets you purchase foreign goods into China, or Taobao, which is where you get the cheaper goods, the white-label goods.

So essentially what the agent looks like is a shopping agent. You go to the app Qwen, and automatically it will help you find, search, and complete transactions—because even Alipay, the payment system, is within its own ecosystem. So this is really quite different from what we saw with OpenAI, where it opened up many apps and they’re trying to integrate the shopping feature into OpenAI, because all of those—whether it’s payment, transaction, or the activity itself—has to be done on a third-party platform. But within Alibaba, the data is within its own ecosystem, so the suggestions, the personalization, it is incredible. However, it’s still in early stages, but I just think it shows that China’s agents are looking quite diverse, and they’re not just a workspace agent right now.

Kyle (08:29)
Yeah, no, that’s so interesting. I mean, especially the case of Alibaba—given the breadth and depth of their existing services—once you start to integrate all that together, you kind of build this broader data, user application services ecosystem that is pretty hard to replicate. And I’m thinking in the U.S., Google—Gemini—integrating that with personalized Gemini with Gmail. But those are also digital services, versus Alibaba really reaching out to the real world. You can order food, you can order real products, you can get real things done, buy tickets to concerts—all through this AI app interface.

Grace Shao (09:22)
Yeah, I think for sure. I think Google has a lot of advantage once Gemini is embedded thoroughly into the work productivity tools. And I kind of see it as: if you have to categorize it, Alibaba’s tools are maybe purely consumer applications, whereas a Google agent potentially would be a lot more prosumer-focused. And in this case, Tencent—the rumor is during CNY they will launch a personal assistant in Yuanbao, which is their AI app, consumer app.

And I think in many ways, Google’s feel might look more like Tencent’s feel, because essentially you have the data and the workflow within what you do when you email people, message people, arrange calls, book calendars. It might look a bit more similar on that front. But I think it’s also the issue of: how much does a company allow the walled gardens to come down between business units? And that’s a regulation issue and a compliance issue as well, which we won’t go into details on.

But Alibaba has a culture of a very top-down driven management style, so they were able to really quickly rally everyone—or at least, in some ways, force everyone—to open up their interfaces to Qwen. I think Tencent, some are saying, might be struggling a bit more with that because the company has always operated a bit more in silos from each business unit. They would work together less closely.

Kyle (11:01)
Yeah, that’s so interesting. With Tencent—being the company behind China’s perhaps most powerful app, WeChat—it was always interesting to me that that didn’t give Tencent an insurmountable lead over everyone else, right? You have this huge user base, huge distribution platform, but they in some ways seem like they’re catching up for some of these, especially now looking at AI agents. And yeah, part of it is strategy and part of it is execution and implementation.

And there was something else you had written about, which is this idea about AI apps being platforms, or AI being the platform itself, versus AI being the operating system. We’re used to iOS for our iPhones or Windows when we think about operating systems. But I was wondering if you could say more about what this idea is about AI itself being the operating system and the interface.

Grace Shao (12:14)
Yeah, I think it all ties to what we were just talking about in terms of how AI can now complete tasks. So I’ll use Tencent as an example. It’s a perfect example. And to answer your previous question about why Tencent felt like they were falling behind and how they’re catching up now: Tencent was one of the first to integrate DeepSeek into its WeChat platform. In the beginning, it was a strategy decision, a cultural reason—because they’re always kind of slow to follow, but they’re very good when they execute—and also a technical challenge, because their LLMs have just been one of the weakest amongst BAT, with B being ByteDance here.

This is relevant to your comment about the platform because I think they really saw WeChat as an interface—the entry point into a new iOS essentially, or a new operating system for mini programs, for deploying AI eventually.

Last year—no, the year before—they first integrated DeepSeek into WeChat. It was a pretty bold move because at that point Alibaba and ByteDance were still doubling down on their proprietary models and had all the guards up and didn’t allow multi-model access through their platforms. The thinking back then was: Tencent has distribution through WeChat, and Yuanbao was kind of average in performance. So they weren’t focusing on pushing Yuanbao to the frontier level. Instead, they said, if we can offer DeepSeek—which was the best, most cost-efficient model in China at that point—through our platform, then maybe we can win the AI race by capturing users within the 1.4 billion MAU on the platform, right?

However, I think the idea of using Yuanbao only as an operating system—or an entry point for an operating system—didn’t really work, because if you’re still relying on other people’s models, it’s not the most native to your functionality. And every month when a new model came out—whether it was ByteDance or Alibaba—it would steal the thunder again, and people would move to the best model, because people want the best capability. And it wasn’t just gimmicky; there was significant growth in intelligence every month in these models.

So there was that. And I think there’s definitely been a change in their mentality to still double down on using WeChat as an entry point.

So last year in December, Tencent hired OpenAI’s Yao Shunyu, which is one of their senior researchers, to lead a whole new model in Tencent. And he said at a summit in Beijing that it’s not about humans being replaced by AI; it’s about people who know how to use tools replacing those who don’t. And instead of obsessing over model parameters, it’s more meaningful at this stage in China to teach people how to use Qwen, Kimi, Zhipu, and other tools effectively.

So what we’re seeing right now is they’re doubling down on model expansion and capabilities, but they’re still trying to use WeChat as an interface to capture users and funnel that into the mini-programs, which will then allow them to complete the agentic tasks we just talked about—booking trips, itineraries, booking restaurants, whatever. And similarly, I think ByteDance rolled something like this out.

ByteDance took a different approach. Even though they’re still thinking about iOS, they pushed out the ZTE JV phone last year. They emphasized: look, we’re not going to create our own phones—we’re not going to manufacture our own phones. Instead, we’re offering this essentially new AI-native operating system to the phone. So they’re taking it one level further than what WeChat is doing. WeChat is still: we’re using existing iOS or Android, but you go to WeChat to use WeChat as an AI OS. For ByteDance, we’re creating an AI OS, and you use a phone where the phone itself is AI-natively integrated and can complete tasks within your hardware.

So I think there’s a move towards this idea that in the future, AI will be called on through a new operating system, and potentially it might not be iOS from Apple or Android that we know today.

Kyle (16:24)
Yeah, yeah, that’s really fascinating. I mean, this move with ByteDance, which is the company behind TikTok, with this Doubao phone really made a lot of waves. And I think the phone was sold out really fast. It offers agentic capabilities that a lot of people expected Apple to roll out at some point. Even Google Android phones don’t offer these kinds of capabilities, right?

Kyle (17:16)
You still need to manually click your app, go through, scroll down, input whatever you want to input. And here—I haven’t tried one myself—but watching the videos, it’s kind of magical to see your phone run on autopilot and accomplish tasks for you, seemingly across apps as well, as if a human were right in front of you operating your smartphone.

More broadly on ByteDance: what do you see as their approach with Doubao, which I believe is the most popular AI app in China, which is maybe surprising to outside observers who might be more familiar with DeepSeek, for example. Why is Doubao so popular, and what is ByteDance’s broader strategy with AI?

Grace Shao (18:09)
Yeah, I think definitely DeepSeek has stolen the headlines because of its model capabilities, but it’s not really focused on the consumer market. I think Liang Wenfeng has openly shown he’s not that interested in capturing the consumer market or trying to profit from that. He is an AGI-pilled guy inside out—he’s really trying to chase frontier research.

But Doubao—people don’t realize, first of all, they spent a lot of money on marketing in the beginning. They went into the consumer market earlier than Qwen. Alibaba only really pushed into the consumer market quite recently. Tencent integrated DeepSeek into their platform quite early, and their models were lagging in the beginning, so that put Doubao in a very good position.

For a lot of people, they might not realize Doubao still came quite late, even though it went consumer early. The Doubao model lab came really late—it was officially launched in August 2023. This is already right after OpenAI redefined expectations globally. It’s already after Baidu, and even Alibaba, had made two or three iterations of their models—people already knew about it. But it was very fast to go 2C immediately.

And to your point, it’s probably the most popular consumer app right now in terms of MAU. I think the latest numbers say it reached 300 million MAU and about 100 million DAU, so they’re really leading. In comparison, I think Alibaba’s Qwen app hits about 100 million MAU at this point. ByteDance’s AI assistant also reaches a scale that is really threatening legacy traffic modes.

There’s also cultural context here to show how successful they are. It was just announced this week that ByteDance Doubao secured an exclusive CCTV Chinese New Year Gala AI partnership. And that’s a clear signal that Beijing is supportive of them, because the CCTV Chinese New Year Gala is probably the most-watched TV show for Chinese people globally, including the diaspora. I think it reaches about 700 to 800 million people globally, and it’s completely vetted by the government.

Kyle (20:15)
Yeah.

Grace Shao (20:36)
They’ve been quite successful in courting everyone. However, because they’re a private company, they’ve remained a lot less vocal about their strategy—at least they don’t have to publicly disclose what they’re doing every quarter, unlike Alibaba and Tencent.

But their entertainment distribution is second to none. In the West, we know TikTok. In China, we have Douyin, and users can go through Douyin and access Doubao directly for chat, creative writing, and AI-generated content. At the end of the day, ByteDance’s mindshare is extremely high in China. So I think although people in the West don’t know so much about Doubao, they know ByteDance. And sooner or later, I believe ByteDance will win in this consumer race in terms of an AI-empowered creative tool or entertainment tool. In many ways I see it similar to Meta in terms of distribution, reach, and potential. It’s just a matter of when they push out all the AI functions through their existing platforms, and then people will get hooked.

Kyle (22:05)
Yeah, no, that’s so interesting. In some ways it makes sense that each of these companies is trying to play to their strengths—leveraging not just competing to have the best model, but integrating into their ecosystems, whether it’s social media–based or e-commerce. Did you want to add something?

Grace Shao (22:13)
Mm-hmm. Exactly. No, I was just saying: exactly to your point, when you look at Chinese big tech companies—ByteDance, Tencent, Alibaba—they’re all trying to tap into the ecosystem. That’s the easiest way into the consumer market. And ByteDance, because of their existing products such as Douyin, TikTok, Jinri Toutiao—their strength is multimodality. Their strength is not helping you send a message—that might be Tencent’s strength eventually. So their agents will be focused on exactly what they’re good at. And I think that’s where the real breakthrough we’ll see this year.

Kyle (23:11)
Yeah, super interesting. Related to this, I was wondering if you could talk a bit about some of the other players beyond BAT—beyond Tencent, Alibaba, and ByteDance. Some startups like Zhipu, MiniMax, Moonshot—some are listing on public markets and raising money. It’s interesting to see how they each reveal slightly different strategies. Any thoughts on seeing this unfold? It’s been quite an enthusiastic time for the Chinese stock market for these new IPOs.

Grace Shao (23:56)
Yeah. So the four tigers—let’s start with that. The four LLM tigers are MiniMax, Moonshot, Zhipu, and Baichuan. Essentially these four labs all started anywhere between 2022 to 2023. They’ve been in this space for a while, so they weren’t just followers after the ChatGPT moment—they’ve been around as well.

In the beginning, they all focused on creating the best frontier LLM. However, now quite a few of them are pivoting or focusing on verticals, and I think that’s a very pragmatic business decision. For example, MiniMax and Zhipu just went public this year—early this year—on the Hong Kong Stock Exchange, both at around six to eight billion dollars, I believe. These are crazy small numbers compared to what we’re seeing in the U.S. in terms of AI lab valuations right now. But there’s a lot of capital constraint for these companies. We all know training and inference costs a lot, and it’s not a game for the poor or scrappy, unfortunately. That’s why we’re seeing big tech leading in model training and deployment.

Kyle (25:15)
Yeah.

Grace Shao (25:23)
And then DeepSeek obviously has its own funding from its hedge fund business. Because of that, you can see MiniMax and Zhipu trying to sell the global story. They’re both saying they’ll have frontier models, and they’ll be cheaper APIs compared to global peers—hint, hint, American labs. And the joke is someone at Zhipu came on my podcast and said: if Anthropic sells for $200, we sell for 20 bucks. That’s the game they’re playing.

They’re thinking of deploying to the Global South. That’s their strategy. They’re thinking of providing model-as-a-service to SOEs across Southeast Asia, and to clients who are more cost-constrained—mid-size, small-size firms. That’s the niche they’re going after.

And in terms of Moonshot: it’s still a private company, but it’s currently the most well-known lab out of the four because of its Kimi models. It’s been pushing frontier research, and recently fundraised within China. Its founder, Yang Zilin, is very focused on AGI. And he’s a charismatic and interesting person, from reading about him and talking to people in the industry. He named his company Moonshot because he said what they’re trying to do is like landing a rocket on the moon. This is what they’re chasing. They’re not going to get distracted by going to consumers or selling to enterprise—they just want the best research.

This is important to note because I think in 2024, Kimi—its consumer app—was at one point one of the most popular in China, and it was competing for first consumer mindshare with Doubao. And I think in the end, Kimi decided we’re not going to put our resources here—Doubao, you can take it; you have more money for marketing and consumer branding—but we’re going to focus our energy back on research.

Baichuan is the one that’s kind of shied away from everything now and is focused on vertical AI, serving the healthcare industry and working with healthcare institutions to find solutions in that sector.

So there are quite a few labs in China, obviously, but the really relevant ones in the global race are maybe the four we mentioned, and then the big tech.

Kyle (28:14)
Yeah, that’s great. It’s so interesting to see the diversity of strategies—how they’re trying to come up with AGI and push benchmarks to the limit, versus consumer-facing applications.

To wrap up this section, I was wondering if you could speak about your view looking across the broader Chinese AI landscape. How do you assess the approaches these companies are taking, and how would you compare that to what we see in the U.S.? And how has that changed over time? You’ve written fantastic pieces about whether they’re running different races or converging to some extent. So I was wondering if you could share some insights.

Grace Shao (29:17)
Yeah, I think my observation was that—first on the startups—they kind of have to go vertical and niche because, to our earlier point about distribution ecosystems, they simply cannot win users from scratch compared to Alibaba, ByteDance, and Tencent’s existing reach. And no one in China has that number-one seat that ChatGPT has really taken up in the U.S., in that sense.

In terms of strategy and how it’s evolved: quite a few of these startups talked about how prior to the ChatGPT moment—November 2022—no one in China took their labs or companies that seriously. It was extremely hard to get funding. And I think that’s because the VC ecosystem here is pragmatic—they want to see a potential product, something feasible and sustainable before they give money.

So when ChatGPT came out, all of a sudden it was easier to fundraise. And after that moment, big tech realized this wasn’t something to be taken lightly—it would be the next playbook or tool or evolution within the internet space. So it pushed both labs and big tech to chase frontier research.

Then when DeepSeek shocked the world with V3 and R1 over Christmas and CNY, it showed there’s a clearly cost-effective way to push forward research. There was sheer excitement and a frenzy—even companies talking about integrating AI into robotic vacuums, AI tutors—it got very crazy for a little bit. But then it tempered down.

And the idea of quick diffusion of AI into the real economy also met challenges. When your technology isn’t good enough, you can’t keep users. The moat isn’t there. Something better will come out, disrupt behavior, and users will go to the new product.

Kyle (31:28)
Right.

Grace Shao (31:29)
So over time, that excitement tempered, and labs and frontier research labs pivoted back and said: okay, we need to develop two things in parallel. One is deployment and diffusion, which Chinese companies were really good at in the beginning. The second is frontier models and synchronization.

And again, if you look at Yao Shunyu joining Tencent, you can see that mentality switch. In the beginning it was: let’s bring in whatever model we can that’s best into our ecosystem, as long as users can access it. Now they’re thinking: how do we create the best model for our users, AI-native to our ecosystem and functionalities, and find that functional adjacency?

On the other hand, in the U.S., in 2024 and early 2025, there was a lot of obsession with chasing AGI, and that was it. There wasn’t much conversation about how AI integrates into the real economy. There was conversation about AI disrupting the labor force, but less about how AI creates new jobs or new value. And that’s shifted.

I think there’s now much more focus and balance among regulators, builders, policymakers, and investors on how to balance technological advancement and diffusion in the most cost-effective way. It might sound less exciting, but in many ways it’s the most sustainable way forward.

Kyle (33:35)
Yeah. It’s fascinating to see different ways of talking about the same technology. And regular people’s attitudes toward AI are interesting too—you have the Silicon Valley bubble, and maybe a tech bubble in China, versus your regular Douyin user or your regular Gmail/Google Search user who sees ChatGPT as a kind of super search, and maybe a personal therapist on the side. To me it’s fascinating to see how it’s being used, and the strategies around that.

I was wondering now if we could shift to Meta and Manus and this really high-profile deal. There are so many things you can read into it. But what happened there? What was the announcement? Why does it matter so much? And what do you make of it all?

Grace Shao (34:58)
Yeah, I think to start with some context: Manus is an agentic AI company—a tool—and it was founded in Beijing. It had a few iterations of branding and product, but we’ll focus on what happened with Meta.

In 2024, it moved its headquarters from Beijing to Singapore. And from Singapore, that’s when they first received their major U.S. capital injection from Benchmark. That was a big deal because it was Benchmark’s first China AI deal, and also one of the highest-profile Chinese AI companies receiving U.S. capital in this sensitive period with the geopolitical backdrop.

Because it received U.S. capital, it had to restructure. It trimmed its Shanghai research team, its Beijing people, and relocated a majority of its co-founders and leaders to Singapore.

And then just recently—a couple weeks ago, maybe two months ago—Meta announced that it bought out Manus for roughly $2 billion.

It was a big deal because a lot of people thought there weren’t exits for Chinese companies going abroad because of AI sensitivity. However, this was a first high-profile one.

I think it’s also a big deal because over the years we’ve seen Singapore benefit in this geopolitical era, if you want to put it that way. In the late 2010s, we saw expats and Western companies move their Hong Kong APAC headquarters to Singapore to soften China exposure. Then there was an influx of Chinese entrepreneurs setting up family offices or offshore accounts in Singapore during the early 2020s regulatory probe in China.

So Singapore opened its arms to money, talent, and vibrancy. Language-wise and proximity-wise, it’s a natural step. Singapore has a Chinese-majority population and they speak Mandarin, while the official language is still English. So whether you’re doing business with the West or with Chinese counterparts, it’s an easy place for companies to stop.

We’ve seen companies do what’s called China-shedding: a Chinese-born company moves to Singapore and softens or severs ties to China, saying they’re now Singaporean, to remove themselves from geopolitical sensitivities or restrictions.

People are watching if this will continue. Soon after the deal was announced, Chinese regulators said they would investigate. People are analyzing it and saying it’s fair because the R&D was done in mainland. If you take that IP outside China, there are restrictions and it may need to be investigated.

So people are focusing on what it means for entrepreneurs who want to build in China but potentially exit outside China—or whether future Chinese AI companies must find exits in China, like what we saw with MiniMax and Zhipu going to the Hong Kong Stock Exchange (or even Shanghai). In many ways, $2 billion isn’t huge, but it’s very symbolic.

Kyle (39:03)
Right. Yeah.

Grace Shao (39:28)
So I think a lot of people are watching to see what comes out of this, especially whether the investigation goes through and whether the deal can successfully go through with Meta.

Kyle (39:28)
Mm. Yeah, yeah. It’s been at the intersection of so many geopolitical and technology issues. And it’s amazing to me on the Meta side that Meta feels like they’re trying to catch up. There’s a lot of excitement around the latest Gemini models or now Claude Code. And to get back to the frontier and be back in the race, it was interesting they looked to a company with origins back in China, and this viral agentic startup that had something Meta didn’t have on its own.

So this acquisition was a story of a technology born in China that moved to Singapore, that maybe is going to the U.S.—a “to be continued” story. And they’re not the only ones who’ve done this Singapore strategy, right? What are some other companies that have tried this move to Singapore or used Singapore as a hub?

Grace Shao (40:56)
Yeah. I think—I’ll say this carefully—lots of AI companies are trying to do so, but they don’t want to be talked about. They don’t want to be flagged.

But exactly to what you’re saying: it shows that even in this globalized world, there’s no such thing as really decoupling. If the technology is good and talent is mobile, how could you stop transfer of information or transfer of ideas? And Meta, if anything, is known to have a majority of engineers who are Chinese ethnic, so there might even be personal relationships between engineers and founders.

So it’s hard to say we’re just going to decouple. I think it’s impossible and not conducive to anyone.

And companies were doing this before. We talked about the early 2020s China-shedding move, and the most famous case study was Sequoia. Sequoia had a huge operation in China and used to just be Sequoia. They spun off their China business and renamed it Hongshan, which is a literal translation of “Sequoia” into Chinese. They separated the businesses, and the Sequoia team either dissipated, were let go, or some moved to Singapore.

ByteDance is also notorious for this. When TikTok was being investigated in the U.S., they moved key figures within the TikTok org chart to Singapore, hired a Singaporean CEO, built a massive ByteDance headquarters there, and repositioned TikTok as a Singaporean company.

Singapore has long been crucial for big tech companies and investment funds globally. Meta and Google have APAC headquarters in Singapore. ByteDance, Alibaba’s Lazada, and Tencent have APAC headquarters there too. It’s a very talent-concentrated place. Even though the market is small, it’s been a gateway to fragmented Southeast Asian markets. And in more recent years, it’s been a gateway for going into China—or going out of China to go global, what they call going to the West.

Kyle (43:42)
Yeah, yeah. It’s ironic that Singapore’s role as a gateway between China and the rest of the world—and Asia and the rest of the world—has been around for a long time. And now with something so new like AI, cutting-edge technology, Singapore is again playing that role.

Some of these long-established factors—its position, geopolitical positioning, institutions, access to capital—keep it playing that role. A lot of other countries look at Singapore and wish they could replicate that navigating, especially now with growing tensions with the U.S. and China.

So do you think this will be a viable pathway going forward? Do you think we’ll see more companies doing this Singapore strategy—or moving elsewhere? Or is it too hard to tell?

Grace Shao (45:14)
I would say I’m not a policy expert, so I don’t really know how to predict what governments would do. At the end of the day, we all live under restrictions that governments establish for business.

But Singapore is interesting. I recently went to Singapore and thought about it. It’s a country of only 60 years of establishment. People think of it as a prosperous city-state, but it’s always been pragmatic for survival and prosperity.

In many ways it’s very Chinese: predominantly Chinese ethnic, early prime ministers were Chinese, they believed in education and talent schemes, and working closely with Chinese relatives from afar. There were many Chinese migrant workers in the early days. However, it’s also very Western: it embraced free trade, democratic values, and things that in the West we think of as Western characteristics.

So it’s always played that role and courted both sides—if you must say that they’re against each other. They’re not against each other, but in terms of differences, Singapore sat in the middle. And today’s young Singaporeans mostly speak English and their ethnic language proficiently. They have a pragmatic sense that for a country so small—by population or geography—they always had to play the game. People like to think they’re Switzerland.

But Kevin Xu wrote a great article saying Singapore is not Switzerland. Singapore is not neutral. Singapore is pragmatic. They will choose what is beneficial to their own people.

And we saw recently at Davos—getting a bit further away—but what Carney said was that these middle countries might need to think about what is good for their own citizens. That might mean choosing certain countries’ cheaper or more efficient goods, but also leaning into certain countries’ ideals, values, and political systems. These are decisions of government leaders and policymakers.

I don’t cover that space, but I think Carney’s speech shed light on the idea that Singapore has benefited from almost all the happenings, and they made sure what they do is best for their people.

Kyle (48:25)
Right. Singapore is a special case of a broader phenomenon: how do all the other countries in the world that are not China or the United States navigate this period? To the extent they take a pragmatic approach, they’re thinking about their own consumers, citizens, and workers. I don’t know if there’s such an ideological bend to how they operate. Increasingly, it seems like pragmatism is the new ideology, as it were.

Grace Shao (49:06)
Yeah. Sorry—just one comment on Singapore. People don’t realize they’re not mindless, like “give me money, I’ll take it.” They’re extremely mindful in planning, and I think it’s admirable.

For example, we talked about bringing talent from China and from the U.S. and Europe. But there are schemes where—at least when I was at CNBC—there was a 6-to-1 ratio: for every foreigner you hire, you need to guarantee a job for a local. There are housing schemes where young newlyweds get subsidies to purchase a house. There are huge tariffs on imported cars to ensure streets aren’t too congested and roads aren’t too polluted. There are harsh policies, but they ultimately benefit people—the citizens, the average lao baixing who live there.

Kyle (50:06)
Yeah, yeah. That Singapore model has been enduring for quite a while now.

Now looking forward: what trends are you following closely now in AI in China, China’s internet ecosystem? What are you paying attention to for 2026 and beyond this year? What do you think everyone else should be paying more attention to?

Grace Shao (50:39)
Yeah, I think we touched on agents at the very beginning. I think 2026 is really when we’ll see agents come to life. In 2025, a lot of people talked about agents. I tried a few agentic tools, but as a layman I didn’t feel the ability to take away mental load or workload. But with what we’re seeing now—especially with Claude Code’s newly launched product—that has shifted how I see agents.

I think we’re going to see how different companies build out agentic tools within their ecosystems—Google with Gemini, Tencent with their Yuanbao app, Alibaba continuing to push Qwen integration within their ecosystem. That’s what I would focus on.

Another thing is interface. We talked about the potential of an AI OS disrupting the old OS system. Honestly, if Apple pushes something out, they’d still probably win because we’re all slaves to our iPhones. I cannot live without my iPhone anyway.

Kyle (51:47)
Right. Yeah, same.

Grace Shao (51:51)
Yeah, but if they can sort out the AI side, I think we’ll still stick with it. But eventually we’ll see a change in how we interact with AI—whether it’s a new OS system or something as simple as interacting a lot more through voice.

I was inspired recently in Singapore at an event where the Grab CTO—sorry, CPO, chief product officer—said people think AI and wearables are gimmicky because for people like you and I, we’re knowledge workers and sit at a desk. We don’t really need a wearable—our hands are free. But what he said was: with Meta glasses, especially if they can be cheaper, they can enable laborers—for example, Grab delivery people—to navigate roads better if they’re wearing glasses. If it frees up their hands, it could be safer and more intuitive, because their work is oriented through their eyes—roads, storefronts, apartment complexes, addresses. So potentially we can think about how we interact with AI leaving our phones and devices as we know them today.

Kyle (53:24)
Yeah, that’s so interesting because I often think: who benefits from all this? Who gains more, and who loses out? Some intuitions get flipped—for example, worries that AI might displace white-collar workers faster than blue-collar workers.

But with wearables, who benefits might not just be people like us, the laptop class. It could be very beneficial for people out in the world—food delivery, construction workers, people working on physical sites who need to engage with the physical world in a totally different way.

Grace Shao (54:18)
Exactly.

Kyle (54:21)
And wearables also seem to play to China’s strengths—consumer electronics, hardware supply chains. And I think Meta’s glasses and some other U.S. wearables are made in China, or a lot of the components come from China. So that’s another interesting aspect.

If it’s purely the models, maybe Google with TPUs is hard to catch up to. But if it’s deploying in the real world, it’s more varied in terms of who stands to gain.

Well, this has been really fantastic, Grace. You are such an amazing person on this topic and related areas. I’m so glad we got a chance to finally have this chat. Any last thoughts you want to add before we wrap up?

Grace Shao (55:39)
Yeah, I think just—when you were talking about hardware—I recently met a company that’s essentially an accelerator. They help European companies and even South Asian companies find hardware supply chains in China, and they’re creating white-labeled robots. It’s crazy.

Back then, we’d think: okay, this shirt or these sneakers are made in China, then you put a Nike logo on or a Zara logo on, and it becomes American or Spanish. But now it’s funny—there’s a huge business in this unsexy area people aren’t looking at, which is the supply chain of robots. They’re producing arms and legs and engines and machinery, shipping it out to a European country, and then a company there assembles it, puts a logo on, redesigns it slightly aesthetically, and it becomes essentially a German or French robot. It’s very interesting—there’s a lot of that happening as well.

And I just want to say: we spent a lot of time today talking about the fusion of AI into the real economy. As someone wrapped up in this, I’m very excited about what it can do. But I hope AI safety people can come up with more standardized international standards on how to regulate this—especially with wearables. I’m very, very scared about what that means for the future of our children—their safety, their privacy, and how that will interact with them down the line. If anything, I hope internationally people can collaborate more on standards and regulations, even if they don’t see eye to eye on politics and other things. This will be very important for the future of humanity, I believe. But yes—thank you so much for having me on your show today.

Kyle (57:24)
Yeah, yeah. Yeah, I completely agree. That’s a really great point. Thank you so much, Grace. This was such an awesome conversation. If people want to learn more about you and your work, where should they go?

Grace Shao (58:03)
Well, please check out AI Proem on Substack. That’s AI, P-R-O-E-M. And people ask what that means—it’s a very nerdy way of saying “preface.” Or check out the podcast Kyle was just on as a guest, called Differentiated Understanding, on YouTube, Spotify, or Apple. If you ever want to reach out to me, feel free to ping me on any of these platforms.

Kyle (58:13)
Fantastic. All right—thank you again, Grace, for coming on the show. If you like this episode, please rate and subscribe to the show on YouTube, Spotify, or Apple Podcasts. You can find episode transcripts and more information on the High Capacity Newsletter at high-capacity.com. And I’ll include links to AI Proem and some select pieces that Grace has published recently. I’m your host, Kyle Chan. Thanks for joining, and see you next time.

Watch or listen to the High Capacity podcast on:

• YouTube

• Apple Podcasts

• Spotify

In this episode, I speak with Tu Le, founder and Managing Director of Sino Auto Insights. He’s one of the best-known experts on China’s EV industry with years of experience working in Detroit, Silicon Valley, and China.

Key takeaways:

• Chinese EV makers are transforming auto manufacturing. They’re taking a page from consumer electronics and compressing timelines for launching new models.

• Geely is planning to enter the US market. Despite US tariff barriers and security restrictions, Chinese EV makers still want to reach the US auto market, likely through investments in the US.

• Western automakers are using Chinese EV tech. Through partnerships and investment deals, some legacy automakers are leveraging Chinese EV platforms and battery technology.

• Chinese EVs are not just cheap; they’re packed with features. Megawatt charging, in-car fridges and theaters, drone selfies, karaoke systems.

• China is pursuing intelligent driving on multiple fronts. Robotaxi services like WeRide and Pony.ai are launching in cities around the world. Meanwhile, smart driving systems for Chinese EVs are moving up the autonomy scale.

Links:

• Sino Auto Insights and the SAI Newsletter

• China EVs & More podcast

• At The Wheel podcast

Transcript

Kyle (00:00)
Welcome to the High Capacity Podcast. I’m your host, Kyle Chan, a fellow at Brookings. I’m thrilled to be joined today by my guest, Tu Le, one of the absolute best experts on China’s EV and automotive industry. He spent years working in Detroit and in Silicon Valley, including at Apple on their supply chains. Then he later moved to China, where he spent over a decade working with Ford and a number of Chinese startups. Now he’s the founder and managing director of Sino Auto Insights, a global tech consultancy focused on transportation and mobility. He’s also the co-host of the podcast China EVs and More with Lei Xing and the author of the SAI Weekly newsletter—both of which I highly recommend. Tu, thank you so much for coming on the show.

Tu Le (00:54)
Kyle, thanks for having me.

Kyle (00:56)
So I thought we would start with the technology. There’s this view out there that Chinese EVs are getting popular because they’re cheap—that affordability is the main thing that gives them an edge. But we both know they’re actually packed full of technology and cool features. A lot of stuff you won’t find elsewhere, like megawatt charging, fridges, or even theater systems built into the cars themselves. I was wondering if you could call out a few interesting features—maybe even things that you saw recently at CES—stuff that would help our American audience realize what they’re missing.

Tu Le (01:42)
Well, you stole my thunder a little bit, because I was going to mention a few of those things already. Companies like BYD have drones on top of their vehicles to do selfies while you’re driving. Also karaoke—very unique to China. An anecdote: a couple of years ago, Tesla sold a microphone for their vehicles in China—nowhere else—and it sold out within, I think, a week.

But on the unique technology: battery swapping is a major, major thing, and it’s successful in China. People kind of poo-poo it around the rest of the world, but there are a couple of reasons it works in China. Number one is the density of most of the cities and the size of the vehicle market. It might not make sense in parts of Europe where there are 200,000, 300,000, 400,000 people, but in a city like Shanghai, where there are 25 million people, or Beijing, 22–23 million, if there are enough NIOs, then it makes sense to have battery swapping.

And then the second thing you mentioned: ultra-fast charging. Last year—in April, right after the auto show—BYD had a little gathering and they showed off mega charging, the ultra-fast mega charging. There was a marketing term they used in Chinese, but it effectively translates to “as fast as gas.”

And what’s important here is that—well, let’s also talk about intelligent driving. In the United States, we think of FSD. Some people might think of Super Cruise and BlueCruise, but what other options are out there for intelligent driving? No one else is really offering it from a brand-name standpoint. But if we look at the China market, XPeng has XNGP, NIO has NOP.

Kyle (03:21)
Yeah, yeah.

Tu Le (03:35)
Lei and I—two years ago—drove an XPeng G9 from Beijing to Shenzhen. Ninety percent of the miles were using their L2-plus system. And then last year we took a NIO ET5 and we only battery swapped, and we used their L2-plus system for about 90% of those miles. Then BYD last year drops a bomb: their system is actually called God’s Eye, and they’re offering it as standard for 20 of their 22 models. So it’s not only that these vehicles have these great technologies—people talk about democratizing technology—but that’s what BYD just did. And that might have pulled intelligent driving forward for the rest of the world by 10 years, because BYD is in over a hundred countries. So we’re going to look back at this time.

Tu Le (05:02)
Even though there’s a huge price war in China and the Chinese government is talking about involution, I think in the automotive space globally, we’re going to look back at this time as some of the most innovative in the history of transportation.

Kyle (05:16)
It’s a really, really exciting time to see the whole industry transform. And you just came back from CES, and there were an unusual number of Chinese EVs represented there, showing off their latest wares—new capabilities, new features. I was wondering if you wanted to highlight any of those—anything you learned while you were out there.

Tu Le (05:40)
So there were two major OEMs featured at CES: Geely and Great Wall Motors. Now there’s another company called Dreame, which is like the Dyson of China. They also showed off—yes, like high-end intelligent vacuums. They had one of the biggest booths at CES; look up Dreame.

Kyle (05:58)
Vacuum makers.

Tu Le (06:10)
They showed off their electric vehicle. Yes—and it’s supposed to be an ultra-premium competitor to a Bugatti-type vehicle, but at a much cheaper price point. It just points to the ambitiousness of the Chinese. And the two main languages outside of English that I heard walking around all the convention halls were Korean and Mandarin.

And we talk about the politics and the relationship challenges between the United States and China, but that’s not stopping businesses from wanting to be players in the United States—or at least exposing U.S. and Western media to what’s out there from China. And I think that’s an important point to make, just because as it seems like things are getting more restrictive, you’re not seeing that at the ground level here in the United States.

And one thing that’s really important that you and I talked about offline: Ash Sutcliffe, who’s the head of international PR for Geely Holdings, had a quick chat with John McElroy—an Autoline journalist, an automotive journalist legend here in Michigan. They sat in a Lynk & Co vehicle, and Ash said Geely will be in the U.S. market within the next 36 months. So the countdown clock for the legacy automakers is now ticking.

Kyle (07:52)
That is a stunning announcement. We think about all the barriers that have been put up—whether it’s 100% tariffs from the U.S. on Chinese EVs, or the connected vehicle rules worried about data flowing through EVs. Despite all of that, it’s fascinating to see a company like Geely say, “We’re going to do it—we’re going to find a way.” And I’m curious to see how U.S. automakers react to that kind of throwing down the gauntlet.

Tu Le (08:33)
Well, if you think about Geely, they’re probably the most international, if you look at their portfolio of brands. They have Volvo, they have Polestar. They have a factory in South Carolina that’s currently underutilized. So this makes sense financially, because if I’m a Chinese OEM with uncertainty about entering the U.S. market—because the current administration is kind of back and forth a little bit on its policies—it would make sense that Geely has first-mover advantage because they already have a facility here. They don’t need to invest billions of dollars in building a greenfield factory or a brownfield or something like that.

All that being said: as a Michigander and someone who grew up working at the traditional automotive companies, I’m fearful for the U.S. automakers. But as a consumer who doesn’t like the direction of MSRPs in the United States tipping over to $50,000, we need more competition at the low end. And I think the Chinese will bring that.

Kyle (09:50)
Yeah. Speaking of some of these Chinese companies, I was wondering if you could take a step back and help us navigate the landscape of Chinese EV companies—all these different brands. People might have heard about some of them: BYD, Geely, some of the newer ones like XPeng and NIO. There’s this proliferation of brands. Now we even have Dreame coming on with their own EV, Xiaomi—which I remember when Xiaomi was an air purifier and smartphone maker, and now they make really, really cool EVs. So I was just wondering if you could give us the buckets of Chinese EV companies and automakers we should be following.

Tu Le (10:39)
Yeah, no problem. So in China there are state-owned enterprises, or SOEs. The requirement the Chinese government made 40 or 45 years ago was that if foreign automakers wanted to come to China, they needed to create a joint venture with some of these state-owned enterprises. Some of the strongest SOEs are SAIC (Shanghai Auto), GAC (Guangzhou Auto), and FAW (First Auto Works). They’re strong because they’re all located in different parts of China. Notable joint ventures are SAIC with Volkswagen, SAIC with General Motors, FAW with Volkswagen. And when we look at those guys, they do sell in the millions of units.

So historically there’s these state-owned enterprises, and technically there are only three independent automakers in China—loosely—because in their home markets, where they’re headquartered, they get a lot of support. BYD is in Shenzhen and gets a lot of support from the city. For example, if you go to Shenzhen right now, I believe every single bus is electric and most—if not all—have a BYD badge. The taxis in Shenzhen—we’re talking tens of thousands of them—are all BYD, all electric.

So BYD, Great Wall Motors, and Geely are technically independent. And if we think about those three, they’re also in the millions of units of sales. BYD is such a compelling story because less than five or six years ago they were at 700,000 units globally. Last year they hit 4.6 million units. What’s really impressive is not only did they have that demand, but they could build to that demand. They didn’t leave anything on the bone. They made sure they were able to accommodate it because they’re vertically integrated.

Tu Le (13:05)
They fab their own silicon, they make their own batteries. So they’re a unique animal. You’re seeing Tesla, Rivian—Ford, I was at a Ford press event yesterday—they said they’re bringing mostly everything in-house. So this is a trend toward control. They see that BYD is able to control a lot more because they do these things in-house. But just like Apple is a unique company, BYD has a formula that works.

With all that being said: Great Wall Motors, Geely, and BYD are in the millions. I would also put Chery into that mix because they have several million units of sales.

And then we get into the startups: the NIOs, the XPengs, the Li Autos. They’re in the hundreds of thousands of units of sales. They’re primarily electric or EREVs—extended-range electric vehicles. In the United States, they might call them REEVs, but in China they call them EREVs. And Li Auto—up until last year or two years ago—were only selling EREVs. Now they also sell EVs, the most famous or infamous probably being the Li Auto Mega, that wedge-looking MPV.

One of the tech companies you mentioned earlier, Xiaomi, is around 14 or 15 years old. I remember because in 2010 I was living in Beijing when they launched. Their automotive division is only a little over four years old. They’ve thrown a monkey wrench into the market because they have two products and they’ve hit over 400,000 units of sales in four years. To give you a quick comparison, XPeng and NIO, which are closer to 10 or 11 years old, are hitting the 400,000-unit mark with many more products in the market.

I’m not trying to oversimplify it, but those are the buckets I’d use. There are still brands launching—like you mentioned, Dreame with new vehicles—and that’s what makes the China market so ultra-competitive.

One other point: most cars—let’s say over 85%—that are purchased are under 300,000 to 350,000 RMB, which is less than $40,000–$45,000. These companies mostly live in the mass market. That’s what differentiates the China market. If we go to the premium side, the average buyer of a Mercedes or a Cadillac is 20 to 25 years younger than in Europe or the United States. Because of that age difference, they embrace technology and digital much more.

Kyle (16:31)
Yeah, this is fascinating. There’s such a huge range of models coming online—from ultra-luxury competitors to Bugatti or Lamborghini all the way down to the mass market. You have basically minivans for families—you have the full range. It’s whatever you can think of.

I want to come back to your comment about Xiaomi being so fast at going from zero cars ever made to suddenly becoming a major player with some of the most sought-after models like the SU7 and now the YU7. I wanted to read back a quote you said on your China EVs and More podcast: you said Chinese EV brands are acting more and more like consumer electronics. If they’re not seeing the sales they want with a model, they can go back and refresh it in 12 to 14 months. You said the automotive space is starting to mimic the technology space. Can you elaborate? What did you mean by cars becoming like consumer electronics?

Tu Le (18:03)
Let me qualify that first, because what I don’t like is the comparison a lot of media outlets make—calling it a “mobile phone on wheels”—because it’s not. Different use cases. Your mobile phone can’t run you over and kill you. That’s the biggest difference, and oversimplifying it misses the point.

Now, the complexity of the technology, the digital, the software—yes, we’re getting to that level. They’re changing the game. Traditionally in automotive, you would design and engineer a component—like a door handle—then build three, four, five prototypes, and test it that way. The Chinese are doing simulations. Toyota—best in class before the Chinese came in and changed the game—had product development cycles like 36 or 38 months, or something like 40 months. So clean sheet to Job One in about four years. The Chinese can do it in 12—let’s say 14 to 16—months.

There are a couple of reasons. The platform is pretty basic and universal, so they can shorten and extend it without creating a brand-new platform. And then the “top hat”—a lot of that is design. And the most important part of the top hat—the seats, everything—is going to be software and integration.

And I’m going to get on my soapbox, Kyle: I hate the term “software-defined vehicle.” As an Apple alum, how come we don’t say “software-defined phone” or “software-defined computer”? Because that’s not what it does. It enables great design. It’s the integration of hardware and software that creates the experience.

I’ll give you a quick example. We’re used to our Android phones and iPhones where the latency is effectively zero. If I touch the screen, something happens. A big challenge automakers had early days when they started becoming “software-defined” was as simple as touchscreen latency. You and I are used to immediate responsiveness, but when we go to our Volkswagen it might take half a second. By then, you’ve pressed it two more times—so when it catches up, you’re three layers deep instead of the one layer you wanted.

Those are the simple things. As an ex–Silicon Valley guy who worked at arguably the coolest hardware-software integration company in the world, those are the things the Chinese companies really lean into. One important note: Li Xiang (Li Auto), Li Bin (NIO), and He Xiaopeng (XPeng) all started as technology guys. They made their wealth from tech companies. Their mindset isn’t first “factory capacity” or “components and supply chains.” They looked at it through a technology lens first—then they learned how to manufacture. That’s the difference between legacies and some of these newer companies. And you could argue Tesla and Rivian a little bit too.

Kyle (22:02)
Yeah, that’s super helpful. In some ways, the term “software-defined vehicle” says a lot about the people who use that word. It’s like you assume a car doesn’t have software or it isn’t a priority, and then you add it after the fact. But with some of these Chinese EV companies, they’re born technology companies—they build cars, and they build these sophisticated machines that are responsive and can do a whole bunch of things beyond getting us from point A to point B—which they also do.

Tu Le (22:44)
And to hammer home how fast things move: if they launch a vehicle with features that aren’t being used or are buggy, they can send an OTA update within a week or two. They’re super responsive because they know through the data whether it’s being used properly or whether it’s buggy. Then they flash the software and you don’t have to go to the dealership.

Traditional automotive refreshes were annual—minor improvements—and then every five years you’d get a major refresh where the sheet metal changes. This is a game changer because maybe the vehicle physically stays the same, but the features, performance, and some of the technology are different just by flashing the software.

Kyle (23:50)
Yeah—so when you’re buying that car, you’re kind of buying the future potential capabilities that might come online later through these over-the-air updates.

Tu Le (24:01)
A specific example: NIO put in silicon that—anticipating a software update six or 12 months later—was kind of dormant. The performance was dormant. But through their product roadmap, they knew OTA update number three would allow them to use the performance of those chips at an optimal level. So they put those chips in without being able to fully utilize them at first. Think about the effect on residual value, right? Traditional analog vehicles depreciate, but if companies are aware enough of their product roadmap, they can increase the value of the vehicle at year three by offering new features that weren’t available when you bought it. These are new opportunities people are having a hard time getting their heads around.

Kyle (25:10)
Totally.

Next I wanted to ask you about manufacturing and changes in the auto industry. What’s going on in China, and how is this changing the rest of the auto industry outside?

Tu Le (25:30)
Let’s start at the 100,000-foot level and then get down to 20,000 feet. The China passenger vehicle market has been the number one market in the world since 2009—they took the crown from the United States. So let’s say 2025—I don’t think all the numbers are out, so I’m going to give estimates. Around 25 million units were sold in China, about 15.5 to 16 million in the United States, around 13 million in Europe, in the EU. Again, estimates.

Last year, John Elkann, chairman of Stellantis, said the China market could be the size of the U.S. and European market combined in 2025. I don’t think that happened, but it got close.

And the take rate of NEVs—China uses that acronym, new energy vehicles—basically BEVs, PHEVs, and fuel cell. Fuel cell is a rounding error. So NEVs are basically battery electric and plug-in hybrids. That take rate was around 50% last year. So one out of every two vehicles in China was an NEV. If we divide 25 million by two, the number of EVs sold in China is almost the size of the European market itself. That’s a scale thing. Manufacturing is all about scale—economies of scale—and leverage against suppliers to push costs down.

If you’re a company buying four of something and I’m a company buying one of something from the same supplier, who do you think has a cost advantage? People talk about labor rates being cheaper, but labor rates are actually going up. I was told engineers in Shanghai at some OEMs make about as much as somebody in Michigan. So it’s not some step function where labor is much cheaper. They make up for it with automation at the manufacturing sites. I went to F2, the factory where NIO builds most of their vehicles in Hefei, and parts of the factory were dark because it was pretty much all automated—only the last part, final assembly, had people walking around.

I think people in the West still imagine a million people on an assembly line putting parts together. I’d argue, as someone living in the Midwest where the UAW is very strong, there’s more manual labor in U.S. factories than there is in China.

And then let’s address the elephant in the room: overcapacity. I lived in China from 2009 to 2022, moved back three years ago, and day one people were like, “Consolidation is right around the corner.” Consolidation has been talked about forever across different sectors. What people should understand is that state-owned enterprises—their job is to keep people employed. It’s not actually to be profitable. If profitability happens, great.

How do provincial leaders, city leaders get promoted? Increasing tax revenue and employment. And I’d love your input, because I think at the central government level they do see the concerns and the risks with overcapacity, and they communicate it to the provincial or local government level—but the incentives aren’t aligned, and it creates challenges.

Kyle (30:30)
Yeah—you get this kind of coordination problem. At a national level, you want a robust automotive industry that’s cutting-edge and competitive, and you might not want everyone replicating and producing at such large volumes. But at the local level, you have provincial and municipal governments thinking about their factories, their workers, and the revenue. And they’re like, “We don’t want to be the ones to give it up. Maybe someone else can do it, but not us.” It’s an interesting multi-level game theory issue.

Tu Le (31:18)
And it pushes against the consolidation needs. The Chinese government recognizes there needs to be less competition because the competition is so fierce that no one’s profitable. There are a handful of companies that—from an accounting standpoint—you could say are profitable, but in real dollars, everybody’s kind of struggling.

Kyle (31:21)
So where do you think this is heading—looking into 2026 and the years ahead? What’s going to happen with the overcapacity problem in China, and how might it spill over to the rest of the world?

Tu Le (32:01)
It already has. Automakers that manufacture in China—GM exports, Ford exports, Tesla exports. So the China market is not just an export hub for Chinese brands; it’s an export hub for automotive brands, full stop.

One narrative that’s inaccurate is that it’s only Chinese brands exporting—that’s not the case. So the tariffs in the United States for Chinese-exported vehicles also affect U.S. legacies, because GM co-owns a brand called Wuling. They rebadge some products as Chevys and ship them to Latin America and South America. Ford booked $900 million of profit out of the China market last year, and their vehicle sales there are very low—they did a lot of that profit through exporting.

There are 65 to 80 million vehicle sales globally annually. If we add EU, U.S., and China, that’s over 50%, closer to 60% of total volume. Those three markets—especially China—dictate the pace. Recently, legacy automakers who were making money hand over fist with ICE vehicles manufactured alongside their JV partners have been eaten away at market share because BYD can build you a PHEV or BEV at the same price with better features—features that attract the young Chinese consumer. So the severity of overcapacity is on the ICE side right now, not as much on the EV side.

But when we look at EVs: not only are the Chinese OEMs the top brands, but on the battery side—which is the largest portion of cost—two Chinese companies have over 50% market share: CATL and BYD. So there’s a lot of capacity on the battery side as well.

Exports have never been higher, but we’re seeing Mexico, Brazil, the United States, Canada create protectionism through tariffs. Everyone’s playing defense. I believe tariffs only hurt consumers in the markets where they’re launched because I want the most innovative products at the lowest price—and you need competition for that.

In 2026, there probably does need to be a reconciliation of capacity. The pressure release valve used to be exports, but now, because of protectionism, exports will flatten out or even reduce a bit. The two Chinese companies I’d call out as different from everyone else are Chery and BYD because they have significant manufacturing outside of China, whereas everyone else mostly ships from China. But if I build capacity in Spain or Poland, it doesn’t help with my overcapacity in China—that’s still an issue. So with protectionism in Europe, the U.S., Canada, there’s going to be more pressure to export to Africa and Southeast Asia where protectionism isn’t as strong and there are no national brands to take share from.

But Southeast Asia is complex—many countries, languages, currencies. It’s hard to add up 15 countries’ market share to equal what you can do in the United States alone.

Kyle (37:09)
So the U.S. market is still crucial.

Tu Le (37:16)
Very much. It’s just difficult to enter Southeast Asia compared to the United States—although the U.S. has political and cost considerations.

Kyle (37:33)
Speaking of what happens to the global auto industry as we see more and more Chinese EVs enter overseas markets: what can existing automakers outside of China do to take advantage of Chinese EV technology—through partnerships? There are high-profile examples like Volkswagen working with XPeng or Ford licensing battery technology from CATL. So it doesn’t have to be a brutal bloodbath where Chinese EVs crush everyone, like so many headlines frame it. Are there ways the industry is reconfiguring where others can take advantage of these technologies and manufacturing processes coming out of China?

Tu Le (38:39)
You brought up one partnership—it’s more of an investment. Volkswagen invested and acquired about a 5% share in XPeng. And this year, in 2026, will be the beginning of that joint venture because Volkswagen will launch XPeng-IP electric vehicles and plug-in hybrids into the China market. Now they’re using the last version of the platform, which is kind of weird because XPeng has a more mature platform on their current vehicles, but it has global implications.

If we look at Europeans, the politics seem less. Stellantis is also heavily invested in a Chinese partner, Leapmotor—19% owned by Stellantis—and they have a sales distribution agreement to help Leapmotor in Europe. These are practical partnerships and opportunities to thrive together.

You could also look at it as Europeans waving the white flag: “I can’t get to the level of software development or user experience design I need without a Chinese partner.” In the U.S., it’s too political—there’s been backlash about the Ford-CATL partnership. But it’s practical. CATL is the largest battery manufacturer, around 37% market share, and they’re experts in LFP manufacturing.

There are different chemistries, but LFP is dominant for Chinese EVs. The only way Western automakers compete on price is if they use LFP too, because before that they were using a more expensive chemistry—NCM/NMC—nickel and cobalt, much more expensive. Longer range, but if we’re trying to make sub-$40,000, sub-$35,000 cars, we need LFP.

But it’s still too political in the United States to announce something like that. Kyle, we know the Chinese want to enter—why don’t we do a joint venture requirement with the Chinese? They did that to us 40, 45 years ago. And a smaller and smaller group still claims they were “forced” to do this or that. Nobody forced Volkswagen to enter China. Nobody forced GM to enter. They signed those joint venture agreements.

Kyle (42:06)
And they made a lot of money along the way. I remember looking at the data for Volkswagen—the year of Dieselgate in Germany, when they got hit really hard. But their sales and profits in China basically carried them through that period. Similar story with GM after the financial crisis.

Tu Le (42:33)
Yeah. I’m not trying to take sides—I’m just trying to keep score and be objective. That’s the honest truth. They’ve been in the market 35, 40 years. For a while, they were booking two, three billion dollars of profit from the China market. And I think a little bit of it is they got busy counting their money, and then the Chinese kind of…

And so, IP theft—at least in the automotive space, if there was IP theft on ICE powertrains, why did Chinese brands need to pivot to EVs? Because they knew they were never going to catch up to Western automakers on powertrain design and engineering. Then the Chinese government took a step back in 2009 and incentivized manufacturing of batteries and EVs. That’s long-term planning.

Kyle (43:55)
Definitely. I totally agree. I think the old days of hoping American cars would just be wanted the world over and you could coast are long over.

Before we get to smart driving and robotaxis, I wanted to ask about hybrids—whether this is a transition strategy for Chinese EV makers and maybe the global auto industry. Hybrids seem popular in China, Europe, the U.S., and much of the world where EV charging infrastructure isn’t as developed. What’s your view on the role of hybrids? Is this a continuing market, or a temporary phase in the broader industry evolution?

Tu Le (45:11)
When we think of hybrid, Westerners think of Toyota—the Prius. We think of it as a few miles of non-gas range. But in China, the plug-in hybrids being launched now might get 150 to 200 miles of range. And there are hybrids, and then there are EREVs. They’re different, but similar.

An EREV technically doesn’t ever have the gas engine power the wheels—it only recharges the battery. Purists will say that’s not clean, and I don’t want to debate it. To me, it’s a win—if I’m not using gas, it’s a win. Practically, I looked at a Cherokee 4xe when I repatriated to the U.S. as a daily driver because I have kids. It got less than 25 miles, and I was like, “Okay, that’s basically a mild hybrid.”

The important thing: dense cities in China, dense cities in a lot of Asia—orders of magnitude different than what people outside New York City think of as dense. San Francisco proper is 700,000. Detroit proper, 700,000. In Beijing, that’s probably the size of my neighborhood.

Kyle (47:03)
Yeah—that’s like Haidian District or something.

Tu Le (47:08)
So hybrid use is really practical in the United States where people are truck-heavy and there are consumers who need a clean energy vehicle with longer effective range. If someone puts a heavy load into their truck, an estimated 300-mile battery range might go down to 150–170 in ideal conditions. In winter, it might go down to 100, 90. So there are practical applications where a hybrid makes sense.

If the effective range is 700, and a heavy load drops it to 300 or 400, that’s still a win. You’re seeing Dodge, you’re seeing Ford. Ford is moving from the battery-electric F-150 Lightning strategy toward a PHEV approach.

So yes, there’s a bridge—until battery swapping comes along, more charging infrastructure comes along, faster charging comes along. All these things are happening. The one binary thing I want to point out: the world is moving to clean energy vehicles, full stop. Whether the United States wants to recognize that or not, it will make our companies less competitive. The question becomes: do we want a profitable Ford that only sells an F-150 in the United States, or a profitable Ford that’s a global company?

Kyle (49:18)
Yeah.

Tu Le (49:19)
I think it’s important to recognize we’re prideful as Americans, and to say another country is doing something better is a difficult pill to swallow. But my first job, Kyle, was in a factory for GM. I’ve driven a lot of miles in China behind the wheel of a lot of Chinese cars. Americans would buy many of them. And hopefully that scares but motivates Western companies—because that’s the key. It can’t just scare you; it needs to motivate you to do better.

Kyle (49:55)
Yeah, definitely.

The last topic I want to ask you about is related to the global expansion of Chinese EVs: the global expansion of Chinese robotaxis. You have services like WeRide, Baidu’s Apollo Go, and Pony.ai. They’re operating across quite a number of cities globally now. And at the same time, I just saw news from CES that Waymo was launching their new autonomous vehicle built on a Zeekr EV—imported from China and retrofitted with sensors and LiDAR. It’s an interesting moment where the EV story and the robotaxi story are happening at the same time. Where do you think this is all going?

Tu Le (51:21)
I would say 2025 was almost like the second coming-out party for robotaxis, because Waymo started marketing quite aggressively that they’re going to enter 20, 30, 40 cities in 2026 and 2027. The important thing is they’re entering London, New York, Detroit—four-season cities with snow and rain. That implies a level of confidence in their technology, because historically up through 2024 or 2025, most launches were in warm-weather, one-season or two-season cities—Arizona, Austin, San Francisco, LA, which doesn’t see snow. So this tells you there’s more confidence in capabilities overall.

Kyle (52:17)
Yeah—it doesn’t say anything about the driver, right?

Tu Le (52:20)
Right. And to your point, let me note: WeRide and Pony are publicly traded in the United States. So whether they enter the U.S. or not, there are a lot of Western investment dollars in these companies.

Waymo is, to me, still the undisputed champion and leader. The Tesla fans will try to throw in FSD, but I don’t see that many Cybercabs out on the road yet. They can talk all they want—they need to show me they’re launching. If we look at numbers, there are over a thousand Waymo vehicles on the roads—I think closer to 1,200 to 1,400. Baidu in China has about a thousand, and WeRide and Pony have a few hundred each. That also gives you an indication of capital.

Rewind three years ago: it was all about data, and the number of pilots and robotaxis created a distinct advantage. I don’t want to get into vision-only versus LiDAR, but it’s an interesting debate.

Long tail: autonomy is happening in commercial and then passenger. What will happen is there will be specific geofenced areas and use cases where Level 4 is available. I was at a presser for Ford yesterday—they said they’ll have Level 3 by 2028. GM said they’ll have Level 3 by 2028. The difference between Level 2 and Level 3—without turning this into autonomy 101—is that at Level 2, we’re responsible and liable as drivers. At Level 3, the liability shifts to the OEM.

And when we talk Level 3, we’re not talking “press a button and it drives me through the city.” Teslas do that, but their lawyers don’t allow them to call it Level 3. And they’re not as good as a lot of people think—although FSD is amazing, and I’m not trying to take anything away from it.

Tu Le (55:15)
What gives some Chinese companies an advantage is industrial policy—there’s a wide embrace nationally of robotaxi companies and pilots. In Europe, it’s starting to happen more and more, because like you said, the Chinese are entering Europe. The London market might be the coolest for robotaxis in the short term because Wayve is there, Waymo is going to be there, and Baidu. So you’ve got a British, American, and Chinese player.

In the Middle East—why? Because they’re trying to be friendly to Chinese tech and U.S. tech. Waymo’s there, and WeRide’s there. And operationally, it’s a desert—sunny all the time—so challenges are lower than Detroit, New York, or Boston.

Waymo is not in China. Baidu is not in the United States. Long term, they probably won’t be, in any significant way. That means Europe is kind of a bystander—with the exception of Wayve—because there are no European robotaxi companies of significance that I can think of. Commercial trucking has a couple of players, but robotaxis—Europe has no horse in the race.

Zoox has a pilot in Las Vegas. I tried, but the wait was 45 minutes and I was too busy. Nuro and Lucid also announced a partnership with Uber. So there are other players showing up in the U.S.

There’s also a second level of robotaxi companies—Deeproute, QCraft—that are players. But the convergence between robotaxi and intelligent driving is happening in China much more than in the West. A company to look out for is Momenta—launched seven years ago. Toyota’s an investor, GM is an investor. The current Buick Electra with the L2-plus system in China is using Momenta. GM invested, I think, $400 million in Momenta.

We’re starting to see the fruits of that labor. Over the last 18 months, Momenta has announced partnerships with a ton of Western OEMs. So when we talk robotaxis, that’s kind of the final frontier, but also look at intelligent driving—NIO has a version, XPeng has a version, Huawei has a stack used by a set of brands. We’ll start to see that bleed into other regions.

For autonomy to be widely adopted, we have to create awareness and then build trust. The Chinese companies are doing that because they’re offering intelligent driving in so many vehicles.

Kyle (58:56)
Yeah. This makes me want to fly out to London or Abu Dhabi and try them all in one place, on the same roads—and then we can all meet up for a drink. You take your Waymo and I’ll take my Apollo or my WeRide, and we’ll meet up afterward. Everything is changing very quickly, and it’s a very exciting time.

Well, that’s all I’ve got from my end. Thank you, Tu, for a fantastic conversation. I’ve learned so much from you. If others want to learn more about you and your work, where should they go?

Tu Le (59:42)
As you mentioned earlier, first of all, thank you for having me. I’m a big fan of your newsletter as well. We talked about this offline—I think you and I aren’t trying to take sides; we’re just trying to lay out how things are working. And hopefully governments and private enterprises look at us as reliable information and data points.

You can find my newsletter at SAI Weekly, or sinoautoinsights.substack.com. I actually just launched another podcast—I co-host two podcasts: China EVs and More, and another one called At the Wheel with Joe White, who is a Pulitzer Prize–winning journalist. That one is more focused on North America and Europe than China.

Good luck on this podcast—I know it’s going to be super successful. You’re going to have super interesting people on it. And I hope for us to collaborate again in the future, because I think you’re doing awesome work as well.

Kyle (1:00:46)
Thank you so much, Tu. That’s very kind of you to say. I absolutely look forward to chatting more in the future. And thank you—thank you for coming on the show.

Tu Le (1:00:56)
Right on, man. I appreciate you having me and inviting me. We have to figure out when we’re going to meet up in what city, and which robotaxis we’re going to take to go grab that drink, right? Maybe we can get a few more people and then race to the bar—race to the same bar. That’d be super fun.

Kyle (1:01:05)
I love it. All right—I’ll just wrap up and say: if you like this episode, please rate and subscribe to the show on YouTube, Spotify, or Apple Podcasts. You can find episode transcripts and more information on the High Capacity newsletter at high-capacity.com.

I’m your host, Kyle Chan. Thanks for joining, and see you next time.

END

I’m excited to launch the new High Capacity podcast, focused on technology in China. The podcast will feature in-depth interviews with experts on what’s happening in China with AI, robotics, EVs, clean tech, biotech, and more.

Watch or listen to the High Capacity podcast on:

• YouTube

• Apple Podcasts

• Spotify

In this episode, I speak with Georg Stieler, Asia Managing Director and Head of Robotics and Automation at Stieler, a global technology consulting firm. He’s a leading expert on robotics in China with over a decade of experience based in Shanghai.

Key takeaways:

• Know-how for robotics applications is now flowing from China to the West. Previously, Western and Japanese firms brought automation expertise to China. But now this flow of knowledge has flipped, especially EVs and batteries.

• China’s robotics market is also flipping. Made in China 2025 was a huge market opportunity for foreign robotics firms. But now they’re being challenged by Chinese players who are moving into global markets.

• Chinese robotics firms won in newer and simpler segments first like “cobots” (collaborative robots) and mobile robots, then moved up the industry ladder. The pandemic gave Chinese firms an opportunity to gain market share.

• The EV and electronics industries caused a step-change in the pace of production. Chinese manufacturers have accelerated development cycles, driving an acceleration in the global auto industry.

• Humanoid robots are part of China’s push into “physical AI.” China has an edge in robotics supply chains. A crucial factor will be the development of vision-language-action (VLA) models that power the “brains” of robots.

Follow Georg Stieler on Twitter / X and LinkedIn

Georg Stieler’s recent publications:

• The Robot Report: China experiences physical AI surge — and how the U.S. should respond

• The Robot Report: iREX 2025: From programmed to perceptive

• The Robot Report: CIIF 2024 shows major robotics trends in China

Transcript

Kyle (00:00)
Welcome to the High Capacity Podcast. I’m your host, Kyle Chan, a fellow at Brookings. And I’m thrilled to be joined today by my guest, Georg Stieler, a leading expert in robotics with extensive industry experience working on robotics in China. He is the Asia Managing Director and Head of Robotics and Automation at Stieler, a global technology consulting firm headquartered in Germany with branches in China and Switzerland. He has spent over 10 years based in Shanghai and traveling all across China to work for some of the leading international robotics companies. Welcome, Georg, and thank you for coming on the show.

Georg (00:39)
Hi Kyle, thank you very much for the invitation, happy to be here.

Kyle (00:43)
Great, so I thought it would be useful to start just with your background. If you wouldn’t mind sharing your experience working in the industry, your experience working in China—you’ve had a long career, really in the details, on the ground, working with clients, working in factories, seeing how this is playing out in real time. So yeah, I just thought, if you wouldn’t mind sharing some of your background in this area.

Georg (01:11)
Yeah. So my first encounter with China for a longer period was actually as a student at Tongji University in 2007 for an exchange year. And after graduation, I first worked as assistant to the CEO of a German renewable energy company. And then after doing that for almost two years, my father asked me if I wanted to do something in China. So I came back to Shanghai in 2010—first on a project basis—and this was so much fun being back that we decided to formally set up a branch office there and moved there permanently in 2011.

And yeah, I mean, our company has traditionally had a strong client base in everything that goes into mechanical engineering—things where Germany has been traditionally good at. And so we started relatively early to work with ABB Robotics. That was in 2012, the first project with them. And they wanted to develop the Chinese market more proactively. That meant we identified the 300 leading Chinese automotive component manufacturers, identified the people in charge for robotics and automation, and asked them what they were using already, what they were planning to purchase, and generated enough leads for ABB to keep their sales team busy for a year.

And we saw then that there was a wave of investment coming for robots at that time. And then we analyzed the whole industrial automation ecosystem in China very extensively. We wrote a large study that was almost 250 pages, and then we were very well placed within the industry to also win other clients.

So we continued working with ABB, helped them, for example, to specify the requirements for a new robot for the electronics industry. But we also worked with KUKA very closely for eight years. We advised Universal Robots and their mother company, Teradyne, on a China strategy during the pandemic.

And also, yeah, I worked for smaller companies in the periphery, like Schunk—they do robot grippers. And yeah, very exciting strategic questions: what are high-value applications for collaborative robots? How do robot applications change when the vehicle industry turns from internal combustion engines to electric?

And yeah, we also took the other perspective and helped companies like Bosch or Mercedes-Benz—or even Coca-Cola—to identify solution providers in the automation field and evaluate them.

Yeah, and then we also worked for American companies in this field who wanted to scout the Chinese market: an autonomous mobile robot company from Pittsburgh, which we then ended up supporting with their market entry in Europe, and also an AI startup from Palo Alto called Drishti. We also scouted China first with them, and then I helped them with their market entry in Europe. Then they were finally sold to Apple in 2023.

So we worked for a lot of different companies in this field and took multiple angles as advisors—and saw how China, from being a market that was lagging technologically, is now a leader in many fields.

Kyle (04:52)
Yeah, no, that’s fascinating. And it’s incredible—the range of different industries, types of applications, types of robots that you’ve worked on, types of clients, both on the supplier side and on the manufacturing side. And yeah, I was just kind of wondering if you might elaborate more on various trends that you’ve seen over time in your experience working in China. So, you know, whether it’s also the shift in the types of questions that you’re getting, the types of projects that clients are asking you for help with, or even the shift in the mix of types of clients.

And then also, I mean, feel free to touch on any of these things or pick and choose, but you mentioned, for example, the shift from internal combustion engine auto manufacturing to electric vehicles. And that must’ve been sort of a huge moment in terms of retooling, rethinking, supply chains, automation processes. So yeah, if you wouldn’t mind sharing some of these longer-term trends that you’ve seen, given your experience working in this space for years now.

Georg (05:58)
Yeah, one of the major changes is now that application know-how is flowing more westwards. Ten years ago, we had, of course, the large companies, but also some quite small and medium-sized automation companies from Europe—Germany, Switzerland, Italy—who wanted to sell into Made in China 2025. They saw this automation tsunami coming and they had these advantages of decades of more experience in automation.

So we also worked for four years—we ran the China branch for a small company with 60 employees. But what—

Kyle (06:37)
And that was a time when—oh, sorry—that was a time when Made in China 2025 for some companies was seen as an opportunity rather than purely a competitive challenge.

Georg (06:49)
Yeah, I mean, for the first couple of years, it was a huge chance for companies to double their revenue. But many underestimated how quickly China caught up, actually. So for many, this optimism changed to fear within just one decade.

Kyle (07:03)
Right.

Georg (07:12)
It also went quicker than I expected.

Kyle (07:16)
Hmm. Yeah, no, that’s interesting. Okay—so sorry I interrupted you, but you were sort of—

Georg (07:20)
And yeah, so—I mean, we can then continue: if we look at the market share of Chinese robot manufacturers, six years ago that was still below 30%. And now we are at 57%. And we’re only counting the local manufacturers there. This doesn’t take into consideration that ABB, KUKA, now also Universal Robots, Yaskawa—that they all manufacture in China.

So the initial goal of Made in China 2025 was to have a localization rate of industrial robots of over 70% by 2025. And we’re probably there when we take the foreign robots manufactured on-site into consideration for this number.

And then, yeah, of course—

Kyle (07:54)
Mm.

Georg (08:19)
Yeah, you mentioned the electronics industry. This is the electric vehicle industry—this is a field where China is just the global benchmark now. You cannot build a competitive battery manufacturing plant in Europe without Chinese know-how these days. So this is clearly something where we have manufacturing know-how coming from China.

Interestingly, it’s the same phenomenon that we saw with German companies going abroad for 40 years: they brought their own suppliers and helped them to globalize. And now when CATL builds a factory or BYD builds a factory in Europe, they bring their own robot manufacturers and system integrators.

Kyle (09:08)
Yeah, yeah—no, this has been a huge change, not just within China, but for the global auto industry. Yeah, actually sort of building on that: is there a large transformation that needs to happen when you’re making that shift from, say, internal combustion engine to EV production? Or are we talking about just brand-new factories? To what extent is it retooling, reconfiguring existing auto plants? And to what extent are you installing completely new robots, completely new systems—obviously dealing with, say, the battery components and battery management systems and things like that. How much of that is very, very new versus modifications to existing production?

Georg (09:56)
Changes for the car body are not so much, but for the whole drivetrain—that’s a major change because you have completely different applications, you have different parts, you have different weights.

A major difference there is also that you have a new set of players. When you’re a European company and one of your major competitive advantages was strong relationships into, say, the German car industry for decades—where it’s very difficult to get listed as a supplier there—and suddenly they are losing market share, and you have some new Chinese players which you barely know, then you need to change a lot.

Kyle (10:33)
Yeah, yeah—it’s a technological change, supply chain change, deep transformation across the industry, including the emergence of—

Georg (10:40)
Yeah—prices, they expect much faster deliveries, they expect more flexibility, faster product turnovers.

Kyle (10:52)
Yeah. Yeah. And in general—

Georg (10:52)
And yeah, this often has not been understood.

Kyle (10:55)
Yeah, yeah, yeah. That’s fascinating. And in general, the automotive industry is the largest user of industrial robots—at least historically. And so it matters a lot for robotic suppliers, both international and Chinese, that there’s this huge shift in the auto industry itself. Is that right?

Georg (11:17)
Also in that regard, China changed something because I think it was 2016 or 2017 when the electronics industry was actually the more important sales market than the automotive industry. And that drive came from China—when companies like Foxconn, or their local competitors like Luxshare, started to deploy robots at a larger scale.

Kyle (11:44)
Yeah,

Georg (11:44)
And that again came with different requirements: shorter payback cycles. Correspondingly, you need to bring them into operation much faster. When you know in March how the next iPhone generation will look, and then you want to start mass production by mid-June, you need to be much faster than people were used to from the automotive industry, where a model would run for seven years—and maybe with a facelift after three and a half years—you had much more time to do things in a diligent way.

Kyle (12:19)
Yeah. Yeah, yeah. I mean, it makes sense in many ways that the automotive industry will take its time. You know, there’s a really strong emphasis on quality and safety control. So, you know, perhaps a longer testing period, a longer development period.

But then I do wonder—to your point—to what extent did China’s experience with electronics manufacturing, like with iPhones, feed into its emerging EV industry? Including faster timelines, maybe different ways of doing production, different ways of organizing supply chains, different ways of even doing design and R&D. Do you see an overlap there—capabilities for one industry spilling over into another?

Georg (13:11)
Yes, of course. So development cycles for new cars shortened dramatically with the emergence of Chinese players—easily by half or even less.

Kyle (13:22)
Yeah, yeah. No, it’s been fascinating to see now some of the non-Chinese automakers try to speed up their timelines, right? So there’s an industry-wide effect, it seems, in some of these cases. And some of that might involve even R&D in China.

Georg (13:39)
Yeah, definitely. The interesting thing is that the permanent emergence of new car models—to keep customers interested—also creates a permanent demand for new industrial equipment. So when we expected that actually the car capacities had been built now, the demand continued for a while.

Kyle (13:54)
I see.

Georg (14:03)
Because new models require new lines with new production equipment as well.

Kyle (14:08)
I see. So this is sort of like a step change in the whole industry—the demands for automation, demands for robotics. It wasn’t just like a wave of installations and then it was done. It was like the entire pace of the industry has stepped up, and the demand overall for the suppliers for industrial automation has increased.

Georg (14:21)
Yeah. In certain fields you definitely have some saturation, but the permanent need for new product differentiation—at least for now—has created a relatively permanent stream for new industrial equipment as well.

Kyle (14:49)
Yeah, that’s super interesting. And then just to take a step back—looking across the broader industrial automation and robotics landscape—where do you see China making the fastest progress? Where are there certain areas—whether it comes to the traditional six-axis arms or cobots—we can get into humanoids later on—but are there certain areas where you see Chinese players really entering into the market fairly actively and becoming more competitive? And are there areas too where it’s been a slower evolution and maybe the global incumbents are still the main players?

Georg (15:39)
Yeah, so where Chinese players made the first easy inroads were price-sensitive fields where the high-quality benchmarks did not matter so much. Like one of the key measures, of course, is precision and reliability—mean time between failures. And this is critical in particular in the automotive industry.

When you have a robot breaking down and the line stops on a Friday afternoon, that can easily cost you a couple of million US dollars. So they are very picky in who gets in there.

And in collaborative robots, and also in mobile robots, the technological backlog was not so much. Collaborative robots is a segment that is only 20 years old. So Chinese players now have a market share of over 90%. The same in mobile robots—we have some very strong local players in this field.

And then also, when we look at the high-quality industrial robots, we have a couple of players that are also attacking the automotive industry now. And that started within the pandemic, when some of the leading international robot manufacturers had some supply problems. Then suddenly companies like BYD gave Estun, for instance—the local market leader—a chance. And they showed that they were good enough. And they provide this industry now with solutions that used to be reserved for the international top four—like spot welding, for instance.

And we also see that companies like Xiaomi, who are an investor in a company called Rokae, are trying to bring them to a level where they can use them for welding applications.

Kyle (17:34)
I see. I see. So they’re kind of working together—the Chinese customers and the Chinese robotic suppliers are working together.

Yeah, yeah. So, I mean, this pandemic story is so fascinating to me because you have this external shock to the whole system. You have a temporary sort of almost like embargo on what can really go back and forth in and out of China—or it makes it much more difficult to bring outside components, service technicians into China.

Would you mind just walking through—from the perspective of a BYD or any large manufacturer in China—I assume previously they would prefer to use established foreign brands for the robotics because they’re reliable. They may be more expensive; if it costs you millions of dollars on the production line when you have a failure, then it might be worth it. But then this shift to the Chinese players—yeah, if you wouldn’t mind elaborating more on what the logic there was, what the rationale was, and how the pandemic reconfigured the industry.

Georg (18:51)
It was not necessarily the movement of people—that was more in other fields of mechanical engineering or plant building. These were semiconductor constraints that didn’t necessarily originate in China.

What we saw were different ways to handle quarantine obligations—so that ABB did it less successfully than, for instance, KUKA or FANUC. ABB’s plant during the Shanghai lockdown in 2022 was basically not running for six weeks in a row, and the others were somehow able to operate.

If you start with 3C—the electronics industry—or continue this to automotive: I mean, one of the things is of course the price, but the other thing is delivery speed. Yeah, so this is a hit or miss in many regards.

And yeah, BYD is also not in the top premium segment. So probably that’s one of the reasons why they were more open to give a local manufacturer a chance. And they still work very closely with, for instance, also with KUKA. So it’s not exclusively with Estun.

Kyle (19:44)
Right.

Kyle (20:06)
Mm-hmm. Yeah, no, this is really interesting. So it’s been sort of opportunistic areas where cobots are newer and maybe don’t require quite the level of precision as a heavy industrial robotic arm for the automotive industry. Maybe some of those areas had seen greater progress for some of the Chinese players. But now even in some of those more difficult-to-enter segments, we’re starting to see Chinese companies move up.

Georg (20:38)
Yeah, I mean, take semiconductor manufacturing, for instance. There you need to fulfill clean room requirements. And this has become a very important market for Siasun—a company which is state-owned and which I never saw very successful in high volumes.

Kyle (20:44)
Mm.

Georg (20:58)
Yeah. Since this is a very strategically important industry for the government, they managed to fulfill the quality requirements for this industry and are selling very well into this now.

Kyle (21:15)
Yeah. So you mentioned a couple of these Chinese industrial automation and robotics companies. And I was wondering if there are any ones that you would like to highlight. So you mentioned Estun, mentioned Siasun, you mentioned some of these cobot players—there’s Inovance, for example, which has been in the news. Yeah, I don’t know if there’s any ones you want to highlight in terms of what they’re able to do, areas where they’re making interesting progress—yeah, which ones are the better ones or the stronger ones, or perceived to be so. And also if any of them have started to branch out beyond the Chinese market and do exports even to Europe, that would also be interesting.

Georg (21:56)
Now, Estun is clearly to mention in this context. This past year—2025—they were, by volume, the number one. Quite an achievement.

I remember we were working with a German company that actually built their production lines eight years ago. They kept calling me from time to time and asked me, “What’s your expectation for the robotics market in half a year?” because our customer doesn’t make any planning—they just say they want to scale as fast as possible. Yeah, we were laughing at that time, but success showed that they were right.

And they were not a state-owned company. They were not a favorite child of the government in the beginning. They have a very charismatic and driven founder who also seems to have a very clear strategy, which includes ambitious internationalization. So they made a couple of very valuable acquisitions. They acquired a motion control company called Trio in Europe. They acquired CLOOS from Germany already six years ago—at that time, CLOOS’s revenue was larger than Estun’s revenue. But the acquisition of Trio gave them the motion control capabilities to catch up with the traditional incumbents, and CLOOS gave them an edge in welding technology.

And all this is coming into play now. And yeah, they are now expanding all over the world—not so much to the US, but a lot to emerging markets and to Europe. I would say in robotics, it’s so far the only Chinese player who does it right.

They hired some very capable people from FANUC Europe and other competitors, which probably cost a lot of money, but also brings their network and attracts other high-profile salespeople.

And they built a factory in Poland. From time to time, I hear that there were some countervailing duties on robots and automation products being discussed in Brussels. They tried to hedge this risk, and this factory in Poland is basically finished. So I think they can be quite successful globally.

Kyle (24:03)
Mm-hmm.

Georg (24:24)
Then Inovance—truly impressive company. I met two of their co-founders just three weeks ago in Germany because globalization is now the major strategic goal they have.

They started as a spin-off from Huawei about 22 years ago, when Huawei was not in a financially strong situation. It was a former team from an American automation company, but they clearly have the “wolf spirit” of Huawei.

And the company—since 2010—they grew every year by 28% on average. It’s now a $5 billion company. And robots are less than 3% of the total revenue.

Kyle (25:00)
Yeah. That’s incredible.

Georg (25:12)
So they can cover the whole industrial automation field and robots are just one part of it. But they have tremendous cross-selling opportunities with their large client base. And I think we can expect a lot from them also because they are known as the price butcher in the market.

So they already took a lot of market share away from Epson—a Japanese company that is the global market leader in that particular field. But except for the iPhone, I think they also got into the Apple supply chain very strongly now.

Kyle (25:45)
I see. Yeah, so—

Georg (25:46)
So these two—they are the major companies in the industrial robot field. And then we have Rokae—also an interesting company. They build half industrial robots, half collaborative robots with integrated force sensors. This is something which sets them apart from much of the competition.

One of their investors is Xiaomi, as I mentioned. So they’re very strong in the Xiaomi supply chain. And I was impressed by the quality of their products when I visited them in Shandong in the factory last fall.

And then, of course, in collaborative robots, we have a whole number of companies—Elite, Obo, Yaka, Ferino. Ferino is very interesting because they sell collaborative robots starting at $4,000. Despite the tensions between the US and China, they are extremely popular with startups that just want to have a manipulator to try out what they are programming.

Kyle (26:52)
Yeah, yeah, yeah. So this is really fascinating. Despite all the geopolitical tensions and things like that, when you look at the actual industries—when you look at robotics developers in the US—they like working with some of these Chinese suppliers and Chinese hardware platforms. So it’s sort of an interesting trend that continues, at least so far.

And yeah, I want to come back to this question about supply chains in a second. But I have to ask about humanoid robots. I think we can’t get this far in the discussion without talking about the videos of dancing robots, the robots doing backflips, robots now—humanoid robots potentially scaling up in production.

There are ambitious plans in China to do mass production of humanoids, to deploy them in actual factory settings doing real economic work. Yeah, I just wanted to kind of get your sense of what the humanoid space is like. What are your views on the technology, on what China’s doing, what the US is doing? So much is happening right now and it’s a very hot topic.

Georg (28:02)
Yeah, this humanoid craze started—it was clearly inspired by Elon Musk and his Optimus introduced in early 2023. And then the Chinese government wrote a paper with the goals for 2025 and 2027—what type of ecosystem for humanoids they want to build—and then the first wave of investments went into component manufacturers.

I always ask myself why, because there are still so many questions unsolved regarding the brains. And I see that in a certain way still. But we also see the investment pattern changing now. It’s going away a bit from the humanoid form factor—even though they’re still popular. I think Unitree will go public at some point in the foreseeable future.

But capital is gravitating a bit now to companies doing the vision-language-action models or a practical deployment of them. And then suddenly you have bimanual manipulation in the forefront. And it doesn’t matter so much if you have legs or if it’s standing on wheels.

So we saw other companies now becoming more interesting. A company called X Square, which was founded some two and a half years ago and now raised around $250 million. We have TAS—a company that doesn’t even have a proper website yet—founded in February, also raised around a quarter billion USD. And they just showed last week their first real demo: how a robot with two arms can put a thread through a needle. So quite impressive.

And then we have Agibot, which still has some humanoids, but which is also putting their vision-language-action model into the center. They showed something very interesting in November: at an electronics company, they could—with their vision-language-action model—reduce the deployment times when the robot was producing different stuff by 90%.

So this is where I see this whole thing in China going at the moment: all these companies—and in China we have around 200 companies that say they are working on some type of humanoid robot—but the more capable ones, they will show some real-world, commercially viable applications next year, which then will include some type of physical AI. It will be relatively limited.

Just last week, I was at Fourier in Shanghai and they were telling me a little bit about the use cases they are at the moment exploring and developing so that they can generate revenue from it next year. That’s what basically all of the companies that are bringing some smart features into their humanoids are working on.

Unitree is a bit an outlier because they have this strong emphasis on producing cheap hardware, which is also used as a platform by many other companies—by many research institutions—because it’s one of the few humanoid robots you can actually buy. Also here in Switzerland, in Zurich, there’s a company that uses the Unitree G1 as the platform for data collection and to make this robot autonomous and smart.

But when they have different requirements, they might also change to a different platform. Or maybe Unitree will finally build some industrial-grade robots—that might also happen. But at the moment, it might be the right way to go as long as the question of the brains is unsolved. You need to produce hardware, and then how do you do it when—and this is evolving all the time.

These prototypes are very expensive. And two weeks ago, I was in Tokyo at IREX, the largest Japanese—well, I think it’s the largest robot show in the world. And there was a company from Taiwan called Solomon—it’s a machine vision specialist. They took a Unitree G1, put proper hands with grippers and a camera on it, put an NVIDIA Jetson on the back, and then they trained it so that it could see five meters. And then you could ask it what to get on a table five meters away—whether it was a blue bottle, a red bottle, a cookie box—and it was still shaky. But it worked.

And you could watch on the screen—you could watch the thinking process, the reasoning process of the robot. Yeah, it was a primitive version of ChatGPT on legs. For me, it was kind of a moment of epiphany, because it’s still very imperfect and there are still many open questions.

Kyle (33:15)
I see. You can see it thinking. Yeah—thanks for sharing.

Georg (33:31)
It really gave you an idea where this whole thing is going.

Kyle (33:38)
Yeah, yeah. No, this is a fascinating time. And I like that you laid out the different directions this technology is going in, because I think there’s a tendency to focus on—like I mentioned earlier—the videos of the humanoid robots either coming out of China or maybe from Figure or Optimus in the US, and just the hardware from the outside.

But there’s so much happening underneath the surface. The vision-language-action models—the VLA models—are crucial here. And then there’s the hardware platform—like you mentioned, Unitree becoming kind of this de facto platform standard. Like I can order it on Walmart.com and Amazon.com here in the US—the Unitree G1 humanoid—something like in the $20,000 range.

And if you are a robotics lab in Silicon Valley or Carnegie Mellon or wherever, it makes a lot of sense that you don’t want to waste time trying to build all the hardware yourself, or spend a lot of money. So here’s a way to get that off the shelf. And then you can get to the exciting stuff like the vision-language-action models.

And then you also emphasized this push for real industrial applications: really trying to put—whether it’s humanoids, or humanoid-like systems that are wheeled, or dual-arm systems with grippers, whatever the exact form might take—this emphasis on doing real economic work. Whether it’s battery testing at CATL’s plant or EV manufacturing with, I think, UBTECH, which has a line of humanoids already working on some of this.

So there are different directions this is taking. And I think one can be entertained by some of the videos that we see, but there’s actually a lot happening underneath the surface.

And I want to come back to this question about data: how important is data in this whole process? What’s important when it comes to training these VLA models? Does China have an edge because of the large manufacturing base—the scale of factories—the push for direct rollout into the industrial space—in terms of creating this data flywheel, where those large volumes of data go back into training the VLA models, which then get improved over time through this feedback loop?

Georg (36:16)
Yeah, I think there are different approaches there. When we talk about setting up applications that work in a reliable way, Chinese companies might be a little bit ahead, but then they don’t spend so much effort on generalization. American companies have better access to training compute. It’s two different ways.

I think we might see—and yeah, we also saw—American companies say, “We don’t present our work at conferences or at trade shows because it doesn’t make sense. We need to deploy at least 10 people for that. And it just slows us down.”

Chinese companies—you could see them at the World AI Conference, at the World Robot Conference, the CIIF in Shanghai—they really show their vision-language-action models at the show and it works fairly well. It worked better than what I’ve seen from German companies in that field so far.

Kyle (37:22)
Interesting. Yeah, yeah. So maybe there’s a little bit of a parallel here in physical AI that we’re seeing with some of the LLM space, where you do have American companies making a lot of progress on the push for general-purpose robotics. So Physical Intelligence, Google DeepMind, Skild AI—they’re sort of banking on the AGI of robotics, maybe, or something like that, right? These general-purpose robotics models that don’t even need to have a specific form factor, whether it’s humanoid or quadruped or different kinds of robotic arms. They’re looking for something more generalized.

Whereas maybe some of the Chinese robotics firms and makers of VLA models—maybe they’re a little bit more specialized in getting those direct applications out more quickly.

Georg (38:14)
Yeah, definitely. They’re very specialized. It was also interesting to see how—last spring—some of these companies were still promoting themselves as a pure “brain” company. Over the last couple of months or weeks, all of them came out with their own type of dual-arm manipulators, mostly on wheels. They said, “To achieve what we want to do next year, we need to be, at least for now, in control of the hardware and have the proximity to the supply chain.”

Kyle (38:46)
Yeah, yeah. Speaking of supply chains, maybe I could ask you a bit about, you know, to what extent does China’s robotics industry have an edge when it comes to the components that go in—whether it’s humanoid or whether it’s more industrial robots. Also the growing relationship between, say, American or European robotics companies and Chinese suppliers for components.

Are there certain parts of the supply chain that China is especially strong in when it comes to robotics components? Are there certain aspects that China’s still dependent on, say, Japanese or German or Swiss or South Korean—say sensors or things like that? Yeah, how do you see the supply chain landscape in terms of components?

Georg (39:38)
When we talk about the brains of humanoids or quadrupeds, this is still dominated by NVIDIA. The training is also still NVIDIA. Even Huawei is trying to become a nexus in this field, and some companies say they are working with Huawei. I think Agibot and Leju Robots are partnering with them.

Overall, it’s not so easy to work with an emerging platform there just because of the network effects. So this is something where Chinese companies still prefer to work with international suppliers. They say they could change when absolutely necessary and it wouldn’t take years—it would be a matter of months. But for now, most of them don’t.

And I also don’t expect this to be a major choke point because China has some other choke points as well—like the origin of the whole semiconductor value chain.

Kyle (40:43)
Yeah, right—with the rare earths.

Georg (40:55)
Yeah, exactly. And then on the other hand, we have the robot bodies and joints. There was a report by HSBC earlier this year that showed the supply chain of Figure, where you have a lot of Chinese companies for harmonic reducers, joints, bearings, sensors, some structural parts.

The same for the Optimus: you have actuator assemblies, you have ball screws. And even in April, I think it was, when Elon Musk warned of the strong dependency on rare earths for magnets.

So we have some dependencies on the other side. And as you mentioned, there are also some Western suppliers that sit in China and purchase, say, motors or servos and then enrich them with jigs and software and then sell them on to the rest of the world. So there is more international collaboration than the political climate would let us assume.

Kyle (41:58)
Yeah, it’s fascinating because it’s like if you remove that geopolitical layer and you just look at the technology and the commercial relationships, it seems like there’s a natural pull—natural complementarities.

You have, you know, Jensen Huang announcing alongside some Chinese robotics companies the rollout of the NVIDIA Thor chips, which are very popular. And actually, NVIDIA chips are popular for electric vehicles in China as well.

In a way, people have been wondering: is this going to be a new choke point? But as you point out, it’s sort of a different technology than AI chips. There may be certain advantages of the CUDA ecosystem still, but there are different alternatives coming out of China as well in some of these spaces.

And at the same time, you have dependencies in the opposite direction—up and down the whole supply chain—all the way down to rare-earth magnets and going all the way up to actuators, bearings, structural components.

It’s one of these worlds where things would naturally want to become more integrated if it weren’t for some of these other tensions between the countries, for example.

But yeah, I wanted to build on what you were saying and just ask: do you see uncertainty on the horizon about these supply chains—about these mutual dependencies? I think probably some people would frame this as: this is a new risk for the US to be relying on Chinese components, or new risks for Europe.

Is this going to go the way of the semiconductor industry, where you have all these different efforts to limit flows? Or do you think this is a different beast, maybe partly because of the way that innovation is happening in this space? Yeah, I don’t know if you have a perspective on the direction of some of these trends.

Georg (44:01)
In semiconductors and toolchains, that’s clearly—we will have two different tech stacks in the world. I think this is a development that’s coming.

And from political and think tank circles, there’s a lot of push to lower the dependencies on China. That’s what I see in Germany in particular. But it strongly contradicts the economic reality.

Because if you have a robot startup and you want to develop something, what shall you do? Shall you just slow down and say, “We source everything from Europe now”? You will be put out of the game.

Kyle (44:47)
Right, right.

Georg (44:48)
Unless we create the same dynamics—an ecosystem that is as flexible and as cost efficient as China—it will be a competitive disadvantage. From that perspective, I think these are just facts, and I don’t see this emerging neither in Europe nor the US. Even with the strong tendencies to bring back some manufacturing to the US, it’s very difficult to catch up with China, who has done a formidable job there during the last 30 years.

Kyle (45:25)
Yeah, yeah. These are very, very useful points to keep in mind. Looking ahead—speaking of possible futures—I was just wondering broadly what kind of trends are you focused on? What areas are you really paying attention to? What do you think the industry might look like in the coming years, either globally or China-specific?

Areas that you see some big changes on the horizon—are there certain challenges when it comes to ramping up these VLA models, deploying humanoid robots at scale in actual industrial uses? What are some of the different aspects of this industry that you think people should be paying more attention to looking ahead?

Georg (46:10)
We need to differentiate there between the different segments. We talked about industrial robots—there we will also see the impact of physical AI. That was great to see in Tokyo.

Not all people agreed with me because some of the applications—say, what Yaskawa, the Japanese company, showed—the ROI of these applications is not entirely clear. But what was a first is that we see now applications that require on-the-fly adaptation, measurement of forces, visual input—so that industrial robots do jobs that they couldn’t do two years ago because they need real-time adaptation of force and of the trajectory and stuff.

And this is something which we will see more and more, which will again increase the number of applications you can do with a robot. And this doesn’t need the humanoid form factor at all. So we will also see physical AI in industrial.

But what we will also see in this field is a consolidation—that’s pretty clear. The Chinese industrial policy has been very successful in that regard that they built up some champions which are hyper-competitive, but we also have a large number of companies that are not. And I expect that many of them will disappear.

Last year, only a small share of Chinese industrial robot makers made a profit. It just shows the intense competition we have in this field.

System integrators—a couple of them showed some profits again towards the end of the year. But by mid-year, they were operating at cost or below just to survive.

Kyle (47:43)
Mmm. Yeah.

Georg (48:02)
These are pressures which we cannot ignore, and it’s also creating certain pushback from lobby groups outside of China because this competitive pressure is also showing in other markets.

Why did ABB decide to sell their industrial robot segment? Because it didn’t make the profits anymore compared to other more attractive markets. And this is certainly also caused by the cost pressure coming from China.

Humanoids and vision-language-action models—yeah, also there we will see a shakeout in 2026. We will see many companies rushing to do an IPO. But we will also see, I expect, some commercial breakthroughs.

It will be a mix of—it will not be the traditional industrial robot arm, but it might not be the robot running around the factory and behaving like a human. I think it will be more applications that require flexibility—autonomy to a limited extent. This will be the first things we will see.

For real humanoids, I think there are still many questions unsolved. But yeah, let’s see—companies want to show something.

Kyle (49:19)
Yeah, yeah. Now this is really interesting. I think you mentioned system integrators. I was wondering if you might say a word about what they are and about China’s movement in that space as well. Because I think for people outside of the industry, they might not realize that you can’t just put robots in the factory, press start, and have them start producing—stamping out car components. There is a huge sophisticated process for integrating, bringing everything online, creating this entire system.

And maybe historically, this area was dominated by a lot of foreign firms, and then to a certain extent now there’s the emergence of Chinese firms in this space. So I just wonder if you could say a little bit about: what is a system integrator, and what’s China doing in that area?

Georg (50:11)
Yeah, you described this already quite well. I mean, the robots don’t install themselves. Somebody needs to design a production line. This is not only the robots—it’s also other components. And then this needs to be designed well. This needs to be all programmed and put into operation.

Basically, the robot usually is only a quarter to a third of such a system integration project.

And when we started working in robotics in China 13 years ago, capable system integrators were still a major bottleneck. But in the meantime, we have companies in China that have 12,000, 15,000 people doing system integration, and they work in three shifts—day and night—when a production line needs to be set up and brought into operation.

We don’t have these types of companies at this size and with this agility neither in Europe nor the US. So this is another competitive edge China has.

Kyle (51:12)
Right, right—sort of this huge pool of technical capabilities, human capital in this space, that can roll out these. And the speed, yeah, yeah.

Georg (51:19)
Yeah, and also the speed. Also the speed. In China you can—and I’m not exaggerating—you can build a factory and take it into operation faster than you get the approval in Germany.

Kyle (51:36)
That’s incredible. Yeah, that’s—I think that will stick in people’s minds. And yeah, I think this is really an important point because I think when people think about building a factory very quickly, it’s not just obviously the actual structure of the factory. It’s everything inside—all the equipment—making sure these processes are working, that they’re functioning in an integrated fashion, that you have testing and monitoring capabilities already there, and that everything’s production-ready.

So this is really important. And to the extent that you can draw on this large talent pool of technicians who are able to—who have now a lot of experience in China, having done this many times.

And then also, to what extent are there third-party system integrators that go from company to company or factory to factory? And then to what extent do some of the larger Chinese manufacturing firms have their own—to what extent do they do it all in-house, essentially?

Georg (52:33)
It’s a mix. You have system integrators with certain specialties. And then we had companies that built these testing facilities—some for Apple, others for Samsung. Then they belonged to a company that was also an electronic contract manufacturer, but then they were sold again.

So it’s a constant decision of make or buy for many of these companies. But of course, also sometimes it involves some intellectual property, and some of the large electronic contract manufacturers decide to keep some sensitive stuff in-house. But when you need speed, sometimes you also hire outside firms. And then in other things like battery, you just have specialized companies.

Kyle (53:18)
I see. Right, right. Yeah, no, this is really interesting. I mean, it’s such a complex industry. I think you just begin to scratch the surface when you talk about some of these things and the different—not only technologies involved—but the different parts of the industry that are involved. And where China is moving in these spaces is changing very rapidly, whether you’re talking about cobots or humanoids or system integrators, or the industries themselves changing so dramatically.

So I think these are all really, really fascinating spaces to watch. Someone like yourself is in the perfect position to help everyone else make sense of all this.

And I just want to highlight: you have a number of really excellent articles out, including a recent one in The Robot Report titled “China experiences physical AI surge and how the U.S. should respond.” And I just want to mention that you have a lot of great writing in the public domain. And I’d highly recommend people go out and read your work. I’ll put up links to some of your recent articles.

And in general, if people want to find out more about you, your firm, your work—how should they follow you?

Georg (54:31)
Follow me on X—that’s my name, Georg Stieler. They can connect on LinkedIn, visit our website, which is stieler.ch, and yeah.

Kyle (54:46)
That’s great. Yeah, I’ll be sure to include links to all of those as well. And I’ll include links in the show notes. And I’ll include more information on the High Capacity website—so that’s high-capacity.com—which is also the home of the High Capacity Substack that I run.

And I just want to thank you, Georg. This is a fantastic conversation. You are an incredible expert in the space, and it’s really a privilege to be able to get your insights on what’s happening in the whole industry—and then specifically what’s happening in China. I think this is a space that a lot of us are going to be following very, very closely.

And yeah—thank you for joining. And that wraps up today’s episode. So please subscribe and rate the show on Spotify, YouTube, Apple Podcasts—wherever you get your podcasts. And I’m your host, Kyle Chan. Thanks for watching or listening, and see you next time.

Georg (55:43)
Thanks, Kyle.