One thing you don't mention is the evolving pricing models. Deepseek are the category definers here (offering "nearly as good", but at 1/10 the cost of proprietory models). There is some erosion in this gap, from both ends, but for day to day work, DS-flash is excellent.
What I find interesting is why, and how? Is the Chinese govt subsidising them as it does batteries and EVs, to corner the global market? Or is it a result of the US sanctions making the Chinese companies focus on efficiency as well as functionality - a constraint not faced by the venture funded US model labs?
I expect the frontier, closed labs to compete much more closely here too. I don’t think it’s subsidies, just a much more competitive price point as open models run at much lower margins (especially if Kimi/glm start charging around 30% revenue share too).
I feel like the pressure of serving a billion users is much higher to target efficiency, rather than having less training compute (which does indeed shift you to a different approach to training your models).
My view is the sanctions massively backfired. Creating not only a homegrown model industry, but a homegrown GPU/NPU industry too. The US child have had a hungry, compliant customer, instead it has a vertically integrated competitor with ideological support. Whereas both the US and China can't afford to fall, only the US might be in the space where they can't afford to succeed either
RL is increasingly the critical piece of post-training and distillation is much messier to apply there (unclear if I would count it at all). The 1-2 months is based on that, and how the top chinese labs have great models they can use to build their envs and data pipelines.
I think the chinese labs are more focused, and fast following is much easier than pushing for dramatic breakthroughs that US companies are best at.
Like most things, I would happily update with more information, but I don't see the evidence that distillation is an existential prop up for the chinese labs performance.
Super helpful post. Hopefully WH and and Congress take a wide field view on this topic, data like this helps them with clarity in their internal discussions
The balance of power question usually gets counted in releases and benchmark scores. There is a commercial version of it that nobody measures, which is who actually sells the tokens, and the two answers are not the same.
We index published prices across 49 vendors. This week the open-weight basket holds 215 models: 104 built in China, 83 in the United States, 20 in France. Those models carry 1,611 listings, and 1,074 of them sit with inference platforms and neoclouds rather than with the labs that released the weights.
So China is ahead on releases and America is ahead on distribution. A rule written around where weights come from is aimed at something different from where the serving revenue and the customer relationship actually sit.
Testimony prepared for Congress is a market document as much as a technical one. Releasing weights is usually described as openness. Operationally it is a transfer of liability: whoever deploys the model owns the failure, and the lab that published it owns none of it. That is a durable advantage, and it does not appear in any benchmark comparing the two countries.
Thank you for this deep dive into the open-weights landscape. Your data on the US-China gap is fascinating, but as a practitioner using AI for physical construction and BIM, two of your points hit incredibly close to home.
First, your observation that Chinese models excel at narrow, benchmarked tasks (like coding) but lag in open scientific domains like physics is the exact bottleneck I face daily. Benchmarks measure syntax and logic, but my work requires physical grounding—material science, local building codes, and structural integrity. If US labs want to win the real economy, they need to optimize for physical and spatial reasoning, not just code generation. A model that can write a Python script but doesn't understand the thermal conductivity of clinker bricks in a damp climate is useless on a construction site.
Second, your example of HuggingFace relying on a Chinese open model for cyber-defense because closed US models refused the prompt highlights a massive blind spot: the "alignment tax." Closed US models are becoming so over-aligned and safety-tuned that they frequently false-refuse benign, highly technical queries. This paranoia pushes practitioners directly into the arms of open-weight models (often Chinese) just to get our actual jobs done.
If the US wants to mitigate the risks of open weights, it must fix the false-refusal problem in closed models and fund highly capable, less-neurotic US open models. Otherwise, the global micro-economy of industry-specific fine-tunes will inevitably be built on a Chinese foundation, simply because they are the most frictionless tools available.
Thanks again for keeping the discourse grounded in reality.
This is a great summation, especially for your Congressional audience. Can I ask whether this was in front of a committee for testimony or just interested offices?
It has been getting delayed and I don't have a full attendee list yet, but I expect this to end up being staffers (as more hearings have popped up given the AI discourse in the last week or two).
There will only be more hearings/engagements soon.
Nathan, I pay for your content and enjoy it. I also, purchased your RL text book. I think you can make your long form posts (and text material) more accessible by taking a less academic tone, and by unpacking your run-on sentences into multiple and shorter sentences. Truth be told, I frequently have to use ChatGPT to get the full meaning of your posts.
Excellent round up.
One thing you don't mention is the evolving pricing models. Deepseek are the category definers here (offering "nearly as good", but at 1/10 the cost of proprietory models). There is some erosion in this gap, from both ends, but for day to day work, DS-flash is excellent.
What I find interesting is why, and how? Is the Chinese govt subsidising them as it does batteries and EVs, to corner the global market? Or is it a result of the US sanctions making the Chinese companies focus on efficiency as well as functionality - a constraint not faced by the venture funded US model labs?
I expect the frontier, closed labs to compete much more closely here too. I don’t think it’s subsidies, just a much more competitive price point as open models run at much lower margins (especially if Kimi/glm start charging around 30% revenue share too).
I feel like the pressure of serving a billion users is much higher to target efficiency, rather than having less training compute (which does indeed shift you to a different approach to training your models).
My view is the sanctions massively backfired. Creating not only a homegrown model industry, but a homegrown GPU/NPU industry too. The US child have had a hungry, compliant customer, instead it has a vertically integrated competitor with ideological support. Whereas both the US and China can't afford to fall, only the US might be in the space where they can't afford to succeed either
Too soon to tell imo, but yes the leaky enforcement seems like the worst of both worlds.
Hi Nathan - great piece! Curious how you came to the 1-2 month estimate for distillation gap
RL is increasingly the critical piece of post-training and distillation is much messier to apply there (unclear if I would count it at all). The 1-2 months is based on that, and how the top chinese labs have great models they can use to build their envs and data pipelines.
I think the chinese labs are more focused, and fast following is much easier than pushing for dramatic breakthroughs that US companies are best at.
https://www.interconnects.ai/p/glm-53-how-chinese-labs-keep-stride
Like most things, I would happily update with more information, but I don't see the evidence that distillation is an existential prop up for the chinese labs performance.
Thanks for the response!
Super helpful post. Hopefully WH and and Congress take a wide field view on this topic, data like this helps them with clarity in their internal discussions
The balance of power question usually gets counted in releases and benchmark scores. There is a commercial version of it that nobody measures, which is who actually sells the tokens, and the two answers are not the same.
We index published prices across 49 vendors. This week the open-weight basket holds 215 models: 104 built in China, 83 in the United States, 20 in France. Those models carry 1,611 listings, and 1,074 of them sit with inference platforms and neoclouds rather than with the labs that released the weights.
So China is ahead on releases and America is ahead on distribution. A rule written around where weights come from is aimed at something different from where the serving revenue and the customer relationship actually sit.
Testimony prepared for Congress is a market document as much as a technical one. Releasing weights is usually described as openness. Operationally it is a transfer of liability: whoever deploys the model owns the failure, and the lab that published it owns none of it. That is a durable advantage, and it does not appear in any benchmark comparing the two countries.
Dear Nathan,
Thank you for this deep dive into the open-weights landscape. Your data on the US-China gap is fascinating, but as a practitioner using AI for physical construction and BIM, two of your points hit incredibly close to home.
First, your observation that Chinese models excel at narrow, benchmarked tasks (like coding) but lag in open scientific domains like physics is the exact bottleneck I face daily. Benchmarks measure syntax and logic, but my work requires physical grounding—material science, local building codes, and structural integrity. If US labs want to win the real economy, they need to optimize for physical and spatial reasoning, not just code generation. A model that can write a Python script but doesn't understand the thermal conductivity of clinker bricks in a damp climate is useless on a construction site.
Second, your example of HuggingFace relying on a Chinese open model for cyber-defense because closed US models refused the prompt highlights a massive blind spot: the "alignment tax." Closed US models are becoming so over-aligned and safety-tuned that they frequently false-refuse benign, highly technical queries. This paranoia pushes practitioners directly into the arms of open-weight models (often Chinese) just to get our actual jobs done.
If the US wants to mitigate the risks of open weights, it must fix the false-refusal problem in closed models and fund highly capable, less-neurotic US open models. Otherwise, the global micro-economy of industry-specific fine-tunes will inevitably be built on a Chinese foundation, simply because they are the most frictionless tools available.
Thanks again for keeping the discourse grounded in reality.
Best regards,
Vladimir Bernadsky
Open models are definitely reshaping the AI landscape exciting times ahead.
Hi Nathan I just published this to complement yours: https://theperceptionshift.substack.com/p/the-fifth-race-why-the-future-of?r=5q3x5r&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true
This is a great summation, especially for your Congressional audience. Can I ask whether this was in front of a committee for testimony or just interested offices?
It has been getting delayed and I don't have a full attendee list yet, but I expect this to end up being staffers (as more hearings have popped up given the AI discourse in the last week or two).
There will only be more hearings/engagements soon.
Nathan, I pay for your content and enjoy it. I also, purchased your RL text book. I think you can make your long form posts (and text material) more accessible by taking a less academic tone, and by unpacking your run-on sentences into multiple and shorter sentences. Truth be told, I frequently have to use ChatGPT to get the full meaning of your posts.