U.S. export controls have made China substantially more innovative all across their technology spectrum. That really is the point. They are now the undisputed leader in open source AI. The capabilities of their models in agentic AI are world-class now. They've made significant strides and have a meaningful edge in countless technologies.
The United States needs better long-term strategy, not simply to hamper its Global competitors if it wants to stay ahead. It needs to create an inviting place for AI Talent to come, but it's done the opposite. Now China's AI Talent density in engineering and research is fundamentally ahead.
Your testimony landed April 22. DeepSeek V4 dropped April 23, Tencent Hy3 the same day. Both open-sourced, both matching or rivaling frontier closed models.
Your point 3 on export controls is playing out in real time. The V4 tech report explicitly links pricing to Huawei Ascend 950 production timelines in H2 2026. They wrote it in the footnote: pricing will drop significantly when Ascend 950 super nodes ship at scale. Export controls created the constraint. The constraint accelerated domestic chip adaptation.
The efficiency framing matters too. Tencent Hy3 (295B total, 21B active) achieves a 7% activation ratio, more efficient than anything at this scale. Your distinction between raw model performance and full-stack adoption is proving out quickly. The winning race is not just at the model layer.
Very well stated, Kyle. I think it would be beneficial to ordinary Americans if those congressmen take you points seriously.
One opportunity you may have missed, in my view, is the caution on AI Safety. Currently, there is little effort spent to prevent AI from going rogue on us Homo Sapiens. And the geopolitical race has made the cooperation near impossible. If the US policymakers shift the stand to be more like yours, that may create some rooms for possible cooperation with China in this area because they will not be too fixated on this “Winning AGI”.
As someone not super-tapped into the AI and industrial policy conversation, this was super helpful framing! And was so heartened to hear you raise energy and energy costs as an important issue - curious if there was any specific response to or discussion on this specific aspect.
Thank you! Yes, the data center energy cost issue is becoming a major political issue in the US. Some members of Congress asked about this and may propose related legislation going forward.
Meanwhile, Anthropic is flirting with the idea of raising the price of Claude Code by 500%, presumably because of how compute constrained they are. We can see flickers of supply chain + infrastructure mastery taking over model primacy as a driver of success already.
Love the piece, Kyle! I’m curious about your take on that third point. Like in the next 5 to 10 years, do you think the US can pivot its policies to make AI truly benefitting the people? Right now, it feels like a lot of the big AI tech giants are doing more fear-mongering than anything else, and most people seem to have a pretty negative association about the whole thing.
I'm in China so I can see the government pushing on AI productivity alnong implementations in manufacturing, medical fields, etc. I think a lot of young folks here are hoping embodied AI could help with elderly care when they get old. But the anxiety over losing jobs is also here. I suppose people in China see the updates in policies but are no less lost than Americans.
Kyle, one of the most interesting parts of this discussion is the distinction between AI as a stand-alone technology and AI as part of a much larger technological stack.
AI is unfolding rapidly, but its integration into economic and social life will likely take decades rather than years. The real transformation may come less from AI alone and more from how it pairs with other technologies, infrastructure systems, and industrial processes over time.
The grid constraints and broader infrastructure bottlenecks you mention seem especially important because they reveal that diffusion may ultimately matter as much as frontier model performance.
I think in a lot of these areas, America has a unique disadvantage. In America, you are rewarded for scale nowadays, it seems, and profit. And in a lot of ways, I see that as being counterproductive to what the goal needs to be. Not to mention, I’ve messed around with AI a lot, and I believe it will be transformative in some ways, but I also think the industry overcompensates a lot as well. AI still has a lot of problems that will be very difficult to solve in order for it to be as impactful as the industry thinks it’s going to be, but that’s just my opinion.
Yes this is a great question. Going forward, we may likely see a multi-track strategy from Chinese AI labs with the most powerful frontier models being closed and less powerful but still very useful models being open. Where to draw the line will be a bigger question over time.
U.S. export controls have made China substantially more innovative all across their technology spectrum. That really is the point. They are now the undisputed leader in open source AI. The capabilities of their models in agentic AI are world-class now. They've made significant strides and have a meaningful edge in countless technologies.
The United States needs better long-term strategy, not simply to hamper its Global competitors if it wants to stay ahead. It needs to create an inviting place for AI Talent to come, but it's done the opposite. Now China's AI Talent density in engineering and research is fundamentally ahead.
Thank you, Michael. Many excellent points here!
Your testimony landed April 22. DeepSeek V4 dropped April 23, Tencent Hy3 the same day. Both open-sourced, both matching or rivaling frontier closed models.
Your point 3 on export controls is playing out in real time. The V4 tech report explicitly links pricing to Huawei Ascend 950 production timelines in H2 2026. They wrote it in the footnote: pricing will drop significantly when Ascend 950 super nodes ship at scale. Export controls created the constraint. The constraint accelerated domestic chip adaptation.
The efficiency framing matters too. Tencent Hy3 (295B total, 21B active) achieves a 7% activation ratio, more efficient than anything at this scale. Your distinction between raw model performance and full-stack adoption is proving out quickly. The winning race is not just at the model layer.
Very well stated, Kyle. I think it would be beneficial to ordinary Americans if those congressmen take you points seriously.
One opportunity you may have missed, in my view, is the caution on AI Safety. Currently, there is little effort spent to prevent AI from going rogue on us Homo Sapiens. And the geopolitical race has made the cooperation near impossible. If the US policymakers shift the stand to be more like yours, that may create some rooms for possible cooperation with China in this area because they will not be too fixated on this “Winning AGI”.
As someone not super-tapped into the AI and industrial policy conversation, this was super helpful framing! And was so heartened to hear you raise energy and energy costs as an important issue - curious if there was any specific response to or discussion on this specific aspect.
Thank you! Yes, the data center energy cost issue is becoming a major political issue in the US. Some members of Congress asked about this and may propose related legislation going forward.
Meanwhile, Anthropic is flirting with the idea of raising the price of Claude Code by 500%, presumably because of how compute constrained they are. We can see flickers of supply chain + infrastructure mastery taking over model primacy as a driver of success already.
In a 'war' .... Everybody loses.
Love the piece, Kyle! I’m curious about your take on that third point. Like in the next 5 to 10 years, do you think the US can pivot its policies to make AI truly benefitting the people? Right now, it feels like a lot of the big AI tech giants are doing more fear-mongering than anything else, and most people seem to have a pretty negative association about the whole thing.
I'm in China so I can see the government pushing on AI productivity alnong implementations in manufacturing, medical fields, etc. I think a lot of young folks here are hoping embodied AI could help with elderly care when they get old. But the anxiety over losing jobs is also here. I suppose people in China see the updates in policies but are no less lost than Americans.
Kyle, one of the most interesting parts of this discussion is the distinction between AI as a stand-alone technology and AI as part of a much larger technological stack.
AI is unfolding rapidly, but its integration into economic and social life will likely take decades rather than years. The real transformation may come less from AI alone and more from how it pairs with other technologies, infrastructure systems, and industrial processes over time.
The grid constraints and broader infrastructure bottlenecks you mention seem especially important because they reveal that diffusion may ultimately matter as much as frontier model performance.
I think in a lot of these areas, America has a unique disadvantage. In America, you are rewarded for scale nowadays, it seems, and profit. And in a lot of ways, I see that as being counterproductive to what the goal needs to be. Not to mention, I’ve messed around with AI a lot, and I believe it will be transformative in some ways, but I also think the industry overcompensates a lot as well. AI still has a lot of problems that will be very difficult to solve in order for it to be as impactful as the industry thinks it’s going to be, but that’s just my opinion.
how to square the various advantages of open weight models with risks for misuse?
from your written testimony
> Due to differing structural and commercial incentives, American AI labs are less motivated to release strong open-weight models with some exceptions
i dont think, for instance, that it would be a good idea for anthropic to release mythos open weight
Yes this is a great question. Going forward, we may likely see a multi-track strategy from Chinese AI labs with the most powerful frontier models being closed and less powerful but still very useful models being open. Where to draw the line will be a bigger question over time.