Podcast: China's semiconductor industrial policy
Is China building the next NVIDIA? Featuring industry expert Paul Triolo
Watch or listen to the High Capacity podcast on:
China is racing to build its own complete semiconductor industry, from AI chips and fabs to chipmaking tools like EUV lithography machines. I'm joined by industry expert Paul Triolo.
00:28 China’s Semiconductor Self-Reliance Push
11:50 How Huawei Became China’s Chip Champion
17:37 Huawei, SMIC, and the Fight for Fab Capacity
24:14 Can China Build Its Own EUV Lithography Machines?
36:55 DeepSeek, Huawei, and China’s AI Chip Feedback Loop
53:53 Why Chinese Tech Giants Are Designing Their Own Chips
Links:
American Affairs: “Innovation Under Pressure: China’s Semiconductor Industry at a Crossroads”
Paul Triolo on X: @pstAsiatech
Paul Triolo’s Substack: AIStackDecrypted
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, Paul Triolo, a partner at DGA-Albright Stonebridge Group and one of the absolute best experts on China’s semiconductor industry and tech landscape more generally. Welcome, Paul, and thanks for coming on the show.
Paul Triolo (00:20)
Great to be here, Kyle. I’m also a big fan of your work on industrial policy in China.
Kyle Chan (00:25)
Thanks. I want to dive right in. China has been making a huge push to develop its own semiconductor industry across the entire supply chain, from lithography machines to etching and deposition, and even high-bandwidth memory. You had a fantastic piece recently in *American Affairs* looking at China’s efforts to do this and the major players involved. You also recently traveled to China to look at the semiconductor industry there.
Could we start with you laying out what you see as China’s broader strategy for building its domestic semiconductor industry? Who are the major players, and which aspects of the supply chain are they focused on at this point?
Paul Triolo (01:11)
Great question, and a complicated one. Stepping back a little, I addressed this to some degree in the *American Affairs* article you mentioned. That was the fourth article I have written focused on how China’s semiconductor industrial policy has evolved over the last five years. Initially, this started before some of the bigger U.S. export controls were shaping the industry. But it is important to note that China has long had the idea of reducing dependence on foreign sources of semiconductors. No country can import, for example, 100 percent, 90 percent, or 80 percent of its semiconductors, which China is still doing.
Going back through government programs, I think I outlined this pretty well in the first paper I wrote on this. There was an evolution from early state-backed efforts, including the 863 Program, and then the 909 Program, which involved a government-backed fab in Shanghai that did not work very well. Too much government involvement did not work. All of this eventually led up to 2014, with the National IC Investment Fund and the government’s effort to inject more market forces into the sector. The idea was to have the fund invest in private sector companies that were investable and let them take a more market-driven approach.
Then, of course, there was Made in China 2025, where the government’s role was defined as setting targets, although those targets were wildly unrealistic across sectors if you look at the Made in China 2025 data. Another thing to note is that it was not until the 14th Five-Year Plan that self-reliance really became part of the vocabulary. Previously, Chinese industrial planners viewed the development of China’s semiconductor industry as part of the global industry. This is a very globalized industry. No country jumps into it and says, “I am going to do this all myself.”
Up until the 14th Five-Year Plan, the goal of China’s semiconductor industrial policy was not total self-reliance. It was to reduce dependence and figure out areas where Chinese companies could be competitive. Then, when you get to the 14th Five-Year Plan, and then the 15th Five-Year Plan, you see much stronger language around self-reliance. After the 14th Five-Year Plan, and arguably because of U.S. export controls, the policy shifts to: now we have to tackle more. Eventually, China has to tackle the whole supply chain.
Up until that point, China was focused on building fabs and capacity in mature-node semiconductors, and its fabs would have moved on to more advanced semiconductors over time. But starting in 2018 and 2019, when the U.S. first imposed controls via the Wassenaar Arrangement on advanced lithography, that sent a signal to the Chinese industry. I think that is why, in the 14th Five-Year Plan, you get a stronger emphasis on self-reliance, which really manifests itself in the plan’s section on self-reliance.
That is the important backdrop. How does a country like China think about reinventing big chunks of the semiconductor supply chain? Up until 2021 or 2022, the focus was very much on front-end manufacturing. Chinese fabs like SMIC, Hua Hong, Grace, and others relied on the best Western tools available. One thing I frequently push back on is the idea that Beijing always intended to kick out foreign suppliers and dominate the whole thing. No. I have talked to many Chinese semiconductor industry people about this. As long as companies could buy the most advanced tools, they were not going to use domestic tools just because the government said, “go domestic.”
One of the themes of the *American Affairs* article is that export controls have forced collaboration between fabs and toolmakers. Subsequent Chinese iterations of the National IC Fund have focused, unsurprisingly, on semiconductors. Finally, if you look at the controls the U.S. put in place in December 2024, those were across the supply chain: materials, substrates, and other inputs much farther back in the chain. Some of the companies ended up on the Entity List in December. China is now confronted with the decision to recreate big parts of the existing supply chain around tools, materials, and inputs to fuel a system that still relies on Western lithography, including ASML deep ultraviolet lithography, for many of its more advanced nodes.
At the same time, and I laid this out in the paper in quite a lot of detail, Huawei has been given the broader task of overseeing development of something like advanced lithography or extreme ultraviolet lithography. That is a separate effort, and only Huawei could probably do that in China. It is an example of how the Chinese government, adjusting policy again, is comfortable with a national champion. Usually it is a state-owned enterprise, as in the aerospace industry. But Huawei was chosen for obvious reasons: it has a lot of experience in advanced semiconductor design. It is overseeing this new era that coexists with the existing system of key players like SMIC and Hua Hong, along with toolmakers like NAURA, AMEC, and Piotech.
If you look at the sales of these toolmakers, they are going through the roof because they are taking over supply of advanced tools in areas like etch and deposition that were previously dominated almost entirely by U.S., Japanese, and Dutch companies. The game now is how all of these pieces come together to eventually enable more advanced-node manufacturing using more advanced lithography, and how long the existing base of tools can be stretched. Chinese companies are now restricted from buying more immersion lithography gear, particularly from ASML, so they are stretching their existing base. They are doing things that nobody has done before, including quadruple patterning and more, to push those machines beyond their intended capacity and get down to something like 7 nanometers and potentially some layers of 5 nanometers, but not beyond that.
We can talk about why that matters for end-use controls. They are struggling because yields are difficult, particularly for advanced chips like GPUs. It is a little easier for smartphone ASICs, but much harder for GPUs. The whole industry is undergoing a tremendous amount of domestic development.
If you go to SEMICON China, for example, I was there last year and talked to a lot of people this year, it is amazing how many domestic companies now show up across the supply chain. At the higher end, you have Huawei, SMIC, Hua Hong, and the toolmakers. SiCarrier is another major player, a Huawei-affiliated company, although Huawei denies that affiliation. SiCarrier was put on the Entity List as a Huawei-affiliated company. Then you have other toolmakers like NAURA, Piotech, and many smaller players. They are all important.
But this is tricky because no country has ever tried to do this before. Evaluating how quickly certain barriers can be overcome, such as advanced lithography, is very hard. ASML took 30 years and tons of PhDs to do that. I try to estimate where this is going in the paper. It is a very dynamic industry, with lots of money being poured in by the government, the private sector, and other investors. CXMT and YMTC are expected to do big IPOs, which will generate a lot of capex to upgrade their facilities. The Chinese government has evolved its policy very flexibly in the face of different challenges over the last 20 years. It has had many failures. Government-run programs and parts of the investment fund were failures. But it is a very different ballgame than it was 10 years ago.
Kyle Chan (11:16)
This is fantastic. I was wondering if you could pull on this Huawei strand a bit. The Huawei story is especially interesting, not just because Huawei has become one of the key leaders in this space, leading a broader ecosystem-level effort, but also because its story seems to capture the Chinese semiconductor story in miniature. I do not think Huawei’s original goal was necessarily to become this lead actor across an entire industry. It seems like it was pushed into that role.
Paul Triolo (11:52)
No, not at all. That is a great question. You are right. I visited Huawei in 2019, just before the pandemic and just before Huawei was put on the Entity List. They took me to their showroom and proudly showed me all the semiconductors they were designing. Huawei had realized that semiconductor design would be useful across its product lines: mobile communications, cloud, servers, and other applications. It decided to eventually design the semiconductors for the verticals it was in.
Huawei invested a lot in HiSilicon. Its secret weapon is still its 10,000 or more engineers. Semiconductor design is very hard to get into. It is expensive and requires a lot of effort and a long-term commitment of resources to iterate designs at the highest level. Huawei did that with HiSilicon. They were designing ASICs that were being manufactured at TSMC in Taiwan when that was still possible. That was the path they had chosen.
I think Huawei probably later regretted that it had not gone down the Samsung path. Samsung has it all: design, fabs, memory, logic, and devices. It is a really integrated company. Huawei came into semiconductors from the telecom side. It had been highly focused on telecom equipment, and as long as it could access TSMC and other Western fabs, it thought there was no reason to get into the fab business, which would be problematic and expensive.
In a similar way, the U.S. forced Huawei to develop a mobile operating system, HarmonyOS, because of controls on Google Mobile Services after the Entity Listing. Huawei and the Chinese government realized that, if Huawei was going to overcome the problems of going beyond design and actually manufacture semiconductors, it would have to jump into that field whole hog.
Huawei has many other companies involved in this endeavor. I think there are something like 2,000 companies under the Huawei umbrella, based on reporting and people I have talked to. It is not just Huawei. But because of Huawei’s central role and its ability to design semiconductors, it was chosen by the NDRC to be the player. But you are right: they were reluctant. They were reluctant to build HarmonyOS. I talked to one of their senior people in 2019, and they were asking, why redesign an operating system when we can just use Android? It is expensive and unprofitable. You do not make money from the operating system. You sell the phone and things on top of the operating system, but not the operating system itself.
For Huawei, that was a huge investment. It had been a hardware company focused on antennas, mobile base stations, and transport networks. Then it had to reinvent itself as a software company with HarmonyOS and build a developer base. That is still an ongoing challenge. Many other Chinese smartphone companies that are not subject to export controls still use Android and compete with Huawei in smartphones.
Huawei was dragged reluctantly into HarmonyOS and then reluctantly into semiconductor manufacturing. But there was really nobody else in China that could lead such an effort, with enough engineering experience and a culture of getting things done. Ren Zhengfei instilled that culture in Huawei. It is a very tight ship and very committed to what it does.
Huawei was the logical choice, but if it had its own preference, I think it would love to be making smartphones, designing semiconductors, manufacturing at TSMC, and continuing to compete with Apple and other telecom companies. But now, in 2026, Huawei has an open-source operating system for mobile devices and is leading a major effort to develop advanced lithography in China. It is a very interesting evolution.
Kyle Chan (17:03)
How do you see Huawei’s role in relation to SMIC, which is still the main foundry in China? Some of Huawei’s chips are fabbed with SMIC, but there seems to be a possible shift toward Huawei’s own facilities or even toward Hua Hong. How do you see these different pieces interacting, given that Huawei has taken a bigger role but there are still multiple major players?
Paul Triolo (17:48)
It is complicated because Huawei’s move into this space made other players unhappy. Huawei carries so much weight, and the government is favoring Huawei. There has been a lot of tension within the system as Huawei has moved into this space. Huawei and its associated companies have also been poaching engineers from other players, and that has generated resentment.
From a technology point of view, the problem has been that at SMIC, the vast majority of limited capacity at 7 nanometers has been taken up by Huawei. We are talking here about the Ascend series and Kirin processors. That has also upset a lot of people at SMIC, because SMIC has many customers it could service, even at 28 nanometers or 14 nanometers. But it has been told, more or less by the government, that because of the need to crank out more AI hardware for competition with the U.S., Huawei gets priority. The Ascend chips are not really GPUs; they are NPUs. But the point is that there is tension.
The need to offload some of that capacity to a facility Huawei controls is where we are. But that is tricky, because Huawei still has to have DUV lithography. At one level, Huawei is repurposing DUV systems, just like SMIC is doing, and trying to set up parallel capacity to take pressure off SMIC. Then SMIC could be freed up to make real GPUs for startups in China like Biren, Moore Threads, MetaX, and Iluvatar CoreX, which have good designs but nowhere to build them because they cannot get enough capacity at SMIC.
When I was in China a couple of weeks ago, I heard stories where one domestic GPU maker had originally been guaranteed a certain amount of wafers at SMIC, and then another big hyperscaler came in and said it would pay more for those wafers. That is the kind of thing that happens when capacity at SMIC is limited. There is a fight over who gets it. The government can play referee, but it is a complicated issue.
Part of the game is for Huawei to develop its own capacity. But that is hard to do. Huawei has a lot of experience helping SMIC, so I think by the end of this year we will probably see some of that capacity come online. Then there is the bigger EUV game. You probably saw the Fanny Potkin story in Reuters in December, which talked about a secret EUV lab in Dongguan. I spent a lot of time in the paper trying to peel that onion. That is a tricky issue. I talked to a lot of people about it. Nobody has seen any EUV light anywhere, so whatever is going on there is under very tight wraps, as the Reuters story pointed out.
But there may be more progress than people think. If they have had that lab up for a year, they could have made a lot of progress. The challenge of EUV is the end-to-end solution. It is not just about the light source, optics, or other individual pieces. Can you produce a commercially viable light envelope that allows you to scale up production? That is very hard. When you get to EUV, you need very good etch. Many other parts of the production line have to be very good: etch, deposition, metrology, computational lithography. There are a lot of complications.
A company or country trying to do all of those pieces faces many challenges. But they are doing it in an environment where some of this is known. There are a lot of people who know how to do parts of it. Still, making advanced semiconductors is one of the hardest things humans do collectively. It is much harder than rocket science. Rocket science is hard, but producing an advanced GPU involves an unbelievable number of technologies, touches, and complexities, down to the chemistry of an etch machine.
China has a lot of great engineers and experience, but again, it is doing something unprecedented. In the West, companies specialize in one part of the chain, and that is all they do. They do it very, very well. That is the challenge when you try to recreate it: you are going up against decades of experience in a very complicated industry. If anybody can do it, Huawei can do it. But it is a great question.
Kyle Chan (23:29)
One of the things I find most interesting about studying the semiconductor industry is that you can dig into one part of the supply chain, and behind that is another massive supply chain. For EUV machines, I believe ASML has thousands of suppliers. Some of those components are so precise and so rare and difficult to produce. People often cite the optics and ASML’s partnership with Zeiss, but you can go down to all of these different levels. Recreating the EUV machine, given all of those parts, is one thing. Recreating the broader supply chain behind it is a whole other level. Then you still have to put it all together and make it work at scale.
Paul Triolo (24:05)
Exactly. There are these key technology barriers. I do not want to call them choke points, but they are barriers and obstacles: the light source, the optics, dedicated chucks, overlay, and many other technologies that have to come together. ASML is more of a systems integrator. It works with great suppliers and integrates all of that. The systems integration piece is probably ASML’s real secret sauce.
The individual technologies are fairly well known in terms of what they are and how to approach them. But the real secret sauce is experience. Some of this cannot be learned by reading a paper and recreating it. Even if you know all the parts that go into it, you still have to know how to make everything work together, and that only comes from experience. That is why the idea that they have had a lab up and running in Dongguan for a year, where they have been putting whatever machine they have through its paces, matters. They will have to gain experience in a real environment with real tools and real pieces if they are going to make progress.
You cannot steal this. You cannot steal the knowledge through cyber means. I worked in the cyber domain for many years, and there is only so much you can steal through cyber. You cannot steal engineering know-how or decades of experience tweaking a light source or an optics system. Those things cannot be shortcut. In the Reuters article, they said there were a number of engineers who had worked for ASML, so China is not starting from scratch on everything. And in this era, the amount of information available on these technologies is much greater than it would have been 20 years ago.
So the flip side is that China is not starting from zero. It has a head start. Huawei and its engineers have a lot of experience working with TSMC on advanced systems. The hope for China is that it can bring all these pieces together and get to something more advanced. No other country could do this - not even close. No other country has the depth of engineers, the resources, and the incentive. That is the key, as I noted in the paper. Export controls give China the incentive to do this. Absent that incentive, market forces would be at work and it would be a different ballgame.
Kyle Chan (27:04)
It feels like they have no choice.
Paul Triolo (27:25)
They have no choice. Exactly. Without that incentive, market forces would dominate and it would be a different ballgame.
Kyle Chan (27:32)
Could you talk more about some of the other players you mentioned, like AMEC, NAURA, and maybe some of the players working on wafer production, packaging, and testing? Which ones are you following closely, and how have they evolved over time? SiCarrier also seems interesting.
Paul Triolo (27:56)
Each one has a complicated story. NAURA is a good example. It was going along for a long time as a fast follower, copying some advances from players like Lam Research, Applied Materials, and KLA. But then, as a result of the October 2022 end-use controls - 14 and 16 nanometers for logic, 128 layers for memory, and 18-nanometer pitch for DRAM - U.S. companies in China had to pull their people out of facilities operating at those thresholds.
Right away, YMTC hired a lot of engineers who were let go when U.S. companies pulled people out and had to lay off local staff. The same thing happened at SMIC. The advanced equipment inside those fabs was then made available to Chinese competitors, including companies like NAURA in etch and deposition. Those companies learned a lot from that equipment. One industry source I talked to said it was the biggest unforced technology transfer in history. When the foreign companies were there, their people maintained the equipment and nobody else could touch it. Once those people were out, all bets were off.
If you look at charts in one of my papers, NAURA is moving along and slowly up the innovation curve in areas like plasma etching. Then, as a result of export controls, it starts moving more quickly into real innovation. It learned a lot from access to U.S. tools. If you look at NAURA’s revenue chart, it is off the charts. But more important than revenue is the learning.
What people miss is that toolmakers have to innovate to meet the requirements of their customers. It is a very close relationship. Toolmakers do not just sell a tool and walk away. They work closely with fabs and front-end manufacturers, who are working with the customers whose designs they are building. That forms a virtuous cycle: the fab has to up its game to meet the customer’s requirements, and the toolmaker has to up its game to meet the fab’s requirements. That is how innovation happens.
By making it possible for Chinese toolmakers like NAURA, Piotech, and AMEC to be more engaged in that loop with fabs like SMIC and Hua Hong, export controls drag those companies up the innovation chain in a way that would not have happened otherwise. If the fabs had not been forced to work with them, they would not have done it.
That is the secret sauce of TSMC. TSMC is not competing with the design companies, so companies come to TSMC and say, “We have this new, difficult design.” Cerebras and its Wafer-Scale Engine is a good example. That was really hard to do. Cerebras worked very closely with TSMC, and TSMC worked very closely with toolmakers to fabricate the Wafer-Scale Engine. That is where the magic and innovation happen.
The flip side is also why U.S. developers want to be in China. Chinese manufacturers are innovating, and U.S. companies want to take advantage of that. Innovation happens both ways. U.S. companies want to cross-license some of the innovations that happen in Chinese fabs. YMTC’s XTacking technology is a good example. It is already important for DRAM and could be important for 3D DRAM. Just before YMTC was put on the Entity List, one of our clients wanted to license that technology, but it could not after the listing.
The industry benefits from cross-licensing, which pushes the whole thing forward. When you disrupt that cycle, or encourage it in the case of Chinese toolmakers, you get different outcomes. I have seen detailed charts showing Western equipment and Chinese replacements across almost every category, especially etch, deposition, parts of metrology, cleaning, and many pieces of the processing chain. Much of that is internal and not publicly available. But for almost every Western piece of equipment, there is now a Chinese equivalent. By equivalent, I mean roughly up to the capability of the Western equipment used in China.
Even before U.S. export controls, Western toolmakers were not selling their best equipment to China because of controls through Wassenaar and other mechanisms. There were always controls around the most advanced tools. Western toolmakers still maintain a lead, but in China they had been able to sell tools a few generations behind their most advanced products. That was a huge source of revenue for them.
Now the worry for the toolmakers is that companies like NAURA are starting to compete with U.S. toolmakers outside China. Because of export controls, those Chinese toolmakers have moved far up the curve. There is nothing to prevent U.S. manufacturing companies like Intel from buying Chinese tools. That is one of the ironies. Chinese memory makers are making memory that U.S. PC makers can buy, but U.S. toolmakers cannot supply tools to those memory makers. The controls have created a strange asymmetry.
I believe TSMC has been trialing some NAURA tools. If the Chinese tools are good and a company needs them, it will buy them. One of the big outcomes of export controls is enabling Chinese toolmakers to compete outside China. U.S. toolmakers are not only denied access to the biggest wafer manufacturing market in the world; they also now have to compete with Chinese companies outside China. NAURA is definitely competitive globally. AMEC is competitive to some degree, and Piotech is rapidly coming up the curve. All of those companies are worth watching.
Kyle Chan (36:21)
That is another fascinating thing about the semiconductor industry. You have so many feedback loops. If you can close the loop between producer and customer, they can iterate back and forth through a co-design and co-evolution process. You see this with chip equipment makers and fabs, with fabs and chip designers, and with AI chip designers and their end customers, the AI companies.
At each stage, this loop can close and the feedback loop can really kick into gear. From the Chinese AI company perspective, they may have been happier working with NVIDIA chips. They were more comfortable with the CUDA ecosystem and happier with NVIDIA hardware. Chinese AI chipmakers themselves might not have had that opportunity before. Now they have the opportunity to work together.
Paul Triolo (37:25)
Right. DeepSeek is now working with Huawei to optimize inference on Ascend chips.
Kyle Chan (37:31)
Could you say more about that layer as well? We saw the recent DeepSeek announcement, and now it seems a number of Chinese AI companies are keen, at least at the inference level, to switch over to domestic Chinese chips and kick off that broader feedback loop.
Paul Triolo (37:52)
That is a great observation, Kyle. I saw this when I was in China recently. We talked to a lot of AI companies. It is a similar dynamic. If the U.S. is going to cut off NVIDIA GPUs, everyone in China would still like to use NVIDIA GPUs, but everyone realizes they cannot count on that. They cannot count on access, or on any logical path forward.
Take H200s, for example. The Trump administration had a strange Truth Social post in the summer saying it would no longer force NVIDIA to downgrade its GPUs to sell to China. But very few H200s have flowed to China. Chinese companies are now in the same situation as toolmakers and fabs: they would love to keep developing on NVIDIA CUDA, because CUDA is a very capable library. Once you understand how CUDA works, and your development cycle is based on it, you do not want to switch. Developers want to learn one system, and CUDA is a complicated system to learn. Switching to another system is not trivial.
That is what Jensen Huang means when he talks about 50 percent of the world’s AI developers. Many of those developers were trained on NVIDIA hardware. Now you also have a complicated situation where many robotics companies in China are using NVIDIA’s Jetson Orin system for robotics models, running those on robots. Those are not export-controlled yet, and they also use CUDA. Across many parts of the Chinese tech ecosystem, there are still these dependencies.
For advanced GPUs, however, China is in a transition period. Many companies still have legacy NVIDIA hardware: A100s, A800s, H800s, L40Ss, and other systems that were allowed to go to China. Those are still useful. The useful life of GPUs is four, five, or six years, and some are still very capable systems.
Every AI developer you talk to in China shows you its hardware stack. Usually there is a middleware layer, and under it is a heterogeneous hardware stack: NVIDIA, AMD, Huawei, Biren, MetaX, Moore Threads, and other players. Many companies have developed middleware to make the underlying hardware transparent to the developer, but that is easier said than done.
Companies like DeepSeek may still be training models on NVIDIA hardware, whether legally held hardware, smuggled hardware, or access to Blackwell systems overseas in places like Japan. But DeepSeek is looking long-term on the inference side. We are in a transition from an 80/20 training-inference balance to something closer to 80/20 inference-training. DeepSeek’s decision is: if we are probably not going to see hundreds of thousands of NVIDIA chips useful for inference, we need to bite the bullet and move to the Ascend series. They are hoping Huawei, working with SMIC and CXMT on high-bandwidth memory, can ramp production later this year and early next year to the point where clusters will be sufficient.
All the other Chinese companies are in the same boat. Optimizing on Huawei hardware is now the preferred option, at least for inference. Training is more complicated. But for inference, the feeling is that Huawei Ascend chips are good enough. They may not be optimized for training, but they are sufficient for inference. Clusters like CloudMatrix and the N+1 systems are sufficient. They may not be as good as Blackwell or Cerebras for inference, but they are good enough.
DeepSeek is very good at optimizing software runtimes to reduce compute requirements and optimize GPU-to-GPU transfer rates. That was a key part of DeepSeek-V3. They figured out how to get around restrictions on NVIDIA GPU-to-GPU transfer rates. The H800s and A800s were made to fall under the export-control threshold by restricting transfer rates. DeepSeek figured out how to send only the most important information over the right link. There are multiple GPU-to-GPU transfer channels, and DeepSeek optimized around that because it had to. It is very good at programming low-level hardware.
That is why DeepSeek is unique. If you read its papers, it is programming at the hardware level. If you use CUDA as an engineer, you do not need to program the hardware directly. But if you can program the hardware directly, you may not need CUDA, and it can be better. Not many companies can do that. DeepSeek can. That is why DeepSeek optimizing on Huawei hardware is meaningful. It can program the hardware directly.
Huawei has the CANN architecture for neural networks, its answer to CUDA. But it is different because Ascend processors are not true general-purpose GPUs; they are neural processing units. Moving from GPU to NPU is another challenge because they do not operate the same way. Huawei is reportedly feverishly redesigning the Ascend series to be more like a GPGPU in future generations.
The other players - MetaX, Biren, and Moore Threads - have lead engineers who came from NVIDIA and AMD. What they are designing are real GPGPUs. That is an important distinction. But Huawei is the biggest player and has most of the capacity at SMIC, so it will produce the most hardware. If Huawei starts doing its own manufacturing and SMIC can dedicate more capacity to domestic GPU players, then some of those other players will be big competitors in the inference space. It will be easier to transition from NVIDIA CUDA to their architectures because those engineers worked with CUDA and their architectures are more CUDA-friendly.
Right now, AI developers have to think about what hardware they will run on, how much hardware they need, and how to make sure their engineers know how to develop on whatever hardware is available. We are in a transition period away from NVIDIA and toward an emerging system that includes Huawei and smaller players, some of which will become big. I suspect MetaX and Biren in particular are very good companies, founded by engineers with lots of industry experience. AI developers now have to work closely with hardware developers, and there is a virtuous feedback loop there.
Kyle Chan (47:04)
Cambricon also seems quite interesting. In the near term, Chinese AI companies may have to allocate resources, time, and effort to making this transition. That could affect how fast they move on model development if you look only at that dimension. But over the medium to long term, it is a bet that they are not suddenly going to be flooded with NVIDIA GPUs. They will have to make this painful transition at some point, so they might as well do it now and set themselves up for a future where they have a range of alternatives.
Paul Triolo (48:11)
Right. One other thing to mention is that Jensen Huang did a very good job on the Dwarkesh podcast explaining this as a five-layer cake. The layers are energy, chips and compute infrastructure, models, and applications. In all the areas besides compute, Chinese companies and the Chinese ecosystem are very strong. On networking and infrastructure, for example, Huawei is very good at putting together optical networking systems. At a systems level, they can do a lot.
In my paper, I tried to bring out the need for a new way of thinking about this. It is not just about GPU-to-GPU performance comparisons. You have to go up to the systems level and ask how these clusters will work together. China is even doing distributed training in a way that is not really being pursued in the West. I have written a couple of Substack pieces on that.
The Chinese system is adapting in different ways and taking advantage of its strengths. Energy is a real issue, but China has a lot of energy. At the application layer, companies like ByteDance, Alibaba, and others are very good at developing applications people want to use, and now using AI. They have advantages there that are not captured when you talk only about who has the most compute.
I will be talking about this next week at the American Chamber of Commerce in China conference, where I am going to debate Chris Miller about whether export controls are working. One thing I will point out is that a lot of people keep coming back to, “We have more GPUs than they do, so we are winning.” Dwarkesh kept coming back to that on the Jensen Huang podcast, and it was a little embarrassing because Jensen was pointing to the five-layer cake and saying China has millions of GPUs.
We are in a different ballgame. It depends on what you mean by winning and what you mean by competition. What I argued in this paper and earlier papers is that this is a bigger ecosystem competition, not just a GPU comparison. GPUs are important, and they matter for whether the U.S. can develop the most advanced models. That is true to some degree. But it is not that simple, because model development, post-training, and other parts of the process do not always rely on infinite scaling of compute capacity. Chinese companies have innovated around the lack of compute.
For inference, Chinese models are going to be much cheaper to run, so more people will run them. I like to talk about this at two levels. First, there is the economic level: diffusion and which countries will see more uptake of AI. China will probably have some advantages there. Then there is the absolute level: who can get to superintelligence first. There, the U.S. lead in compute may be more relevant. But even there, there are issues: how quickly Chinese companies can innovate and produce more capacity for model developers.
The gap in leading models is debated. I have seen estimates all over the map in the last couple of days - either even, or the U.S. being eight months ahead. In any case, everybody thinks Chinese models will catch up at some point. The size of that lead matters. Mythos has also become an important narrative in the debate over who is ahead and whose models are more capable. We are in a very interesting period in U.S.-China AI tech competition.
Kyle Chan (53:17)
I wanted to ask one last question. Beyond AI and AI chips - or maybe this is still related to AI - we see a whole bunch of Chinese tech companies jumping into semiconductors, designing their own chips, and creating their own ASICs. In the U.S., Google has TPUs and Amazon has Trainium. But what is really incredible to me is how many Chinese tech companies have their own chip design units and see this as a strategic core capability. Why is this happening? How is it able to happen? And what does that mean longer term for China’s tech industry in the AI+ realm?
Paul Triolo (54:15)
That is a great question. In the U.S., hyperscalers like Google, Meta, and Amazon at some point did not like paying the premium for NVIDIA hardware. Those are general-purpose GPUs. The idea of optimizing ASICs - application-specific integrated circuits - for a particular company and its model development became attractive. Companies like Broadcom and Marvell can do ASIC design, and the hyperscalers still use TSMC to manufacture those chips on the U.S. side.
You have Amazon Trainium, and Google has been doing TPUs for a long time. Traditionally, cloud services companies had different clusters of hardware. Depending on the workload, TPUs might be more expensive, and if a workload was optimized for TPU features, you would run it on TPUs. I have a friend at NIH who would have to decide: can we pay for TPUs, or should we use cheaper GPUs? That has been going on for a long time in the West.
But funding and recruiting a semiconductor design team is not easy. Lenovo in China once wanted to get into semiconductor design, but a decision was made that it would be too expensive and difficult. Alibaba, Tencent, and Baidu have the resources to do it. Baidu’s Kunlunxin is going public in Hong Kong, although its valuation is not that high. Alibaba’s DAMO Academy has assembled a very good team, and it is making PPUs. Alibaba is like Google in that it has already sold some of those chips; I think China Telecom bought some for a cluster in China.
In the U.S., those chips were originally all in-house because they were optimized for the hyperscaler’s own workloads. Now companies like Google and others are going to sell them. The problem is that you need people who know how to write software for them. If you are trained on CUDA and somebody gives you Trainium, you have to figure out how to use it. The beauty of NVIDIA is that these are general-purpose GPUs. They may not be optimized for a particular workload, but through CUDA and other tools you can optimize them pretty well.
China has the same dynamic, but Chinese companies are more constrained. Just as everyone in the West uses TSMC, in China they have to use SMIC to produce these PPUs. Tencent has its own chips and has been doing this quietly for a long time. It has a very capable design team. Baidu has also been designing Kunlun chips for a long time. I think they were originally using TSMC, and that is a complicated issue because I think none of these companies are on the Entity List. But there are restrictions on what TSMC can manufacture as a result of the Foundry Due Diligence rule pushed out at the end of the Biden administration. As designs reach a certain point, TSMC may have to pull the plug on cooperation.
This shows how capable the Chinese system is of producing multiple semiconductor design teams that can design very good ASICs optimized for AI workloads. The trend is not surprising. The difference is that these companies operate under much greater constraints than leading U.S. hyperscalers. There is also murkiness around whether some chips are still being manufactured by TSMC or Samsung. I will probably have to do a Substack on that.
Xiaomi, for example, is still using TSMC to manufacture 3-nanometer smartphone SoCs, which have some GPU capacity. But the controls were designed for data-center GPUs, not smartphone SoCs with GPU capacity. There is a clearer line there. Some of the Chinese ASICs have GPU-like capabilities, so the performance issue is complicated. When do they run into the Foundry Due Diligence rule?
Those kinds of rules put a big burden on TSMC to figure out what its customers’ chips are capable of doing. In many cases, TSMC is producing the wafer and may not know what the final package will look like once memory and other components are incorporated. I do not know how TSMC figures that out. That is why Huawei was able to manufacture something like three million 910B wafers at TSMC using a cutout company. That prompted the due-diligence rule because TSMC probably rightly claimed it did not know what the wafer alone was capable of doing. U.S. controls are also tied to things like memory bandwidth, which can only be measured once you have the final packaged product, and those were shipped back to China for packaging.
This shows how difficult it is to write export controls and draw lines when the industry is dynamic and changing every day. Those export controls were originally written for technologies like WMD, which never changed much - in some cases technologies from the 1940s and 1950s. Now you are trying to write export controls for technologies that change every day.
Kyle Chan (1:01:03)
With this global supply chain and many different players, applying a nuclear analogy to semiconductors just does not work. It is a fundamentally different industry, a different set of technologies, and it works in a very different way. The nuclear analogy is an easy fallback because it sounds familiar on the surface, but when you dig in, semiconductors seem structurally very different.
Paul Triolo (1:01:39)
I think Jensen Huang said it best: nobody needs nuclear weapons, but everybody needs AI. Everybody needs semiconductors. It is not a very good analogy. I have been critical of some of the major labs on this. Dario Amodei has likened AI to nuclear weapons, and I think that is a really bad analogy.
It is interesting because it raises the question: why should private companies have a capability like nuclear weapons? The U.S. nuclear weapons program was a government program, not an entire private sector. If you make the nuclear analogy, you are almost making the argument that the government should take over and nationalize the labs. People in the industry who make that analogy need to be careful, because they are essentially making an argument for the government to nationalize those labs.
I have a long paper coming out on the conditions under which China and the U.S. would consider nationalizing their AI labs. That is a huge question in my mind, and Mythos only makes it more so. We could do a whole podcast just on that.
Kyle Chan (1:02:45)
That would be amazing. Your work has been incredible. I will link to your Twitter account, @pstAsiatech, which is a fantastic source for the latest in Chinese tech developments. I will also link to your Substack, *AI Stack Decrypted*, if that works for you.
Paul Triolo (1:03:11)
I have a new one coming out today on the Compute List, version 2, which I hope you will find interesting.
Kyle Chan (1:03:17)
Awesome. I will include a link to that directly as well. I will also link to the *American Affairs* piece we have been referring to throughout this episode. Let me know later if there are other items you want to link to. I want it to be as easy as possible for people to find your work, your writing, and your latest thinking on this. This has been incredibly helpful. I really appreciate it.
Paul Triolo (1:03:43)
Kyle, I want to thank you too, because we did not talk about rare earths at all, and I appreciate it. I have been doing a lot of rare earths work over the last year, including this week, because we have a number of clients. As you can imagine, any high-tech company in the U.S. relies on rare earths. That issue has not gone away, I can assure you. It is a big deal.
It is also tied up in the semiconductor domain, because all semiconductors are touched at some point by rare earth materials or by equipment with motors driven by rare earth magnets. That is another interesting topic. Thank you. You are asking all the right questions. I was very excited to see your appointment at Brookings, a favorite D.C. think tank of mine. Ryan Hass, who is a longtime friend and a great commentator on China, is there. I have always been a big fan of Brookings, and I am glad you are there.
Kyle Chan (1:04:58)
Thank you, Paul. It is very exciting to be in this space together. I will wrap up by saying: 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 am your host, Kyle Chan. Thanks for joining, and see you next time.