Gary Marcus blames Meta's open-source strategy for the shift
Zuckerberg and LeCun’s unilateral decision to open source Llama likely (partly) catalyzed China’s AI industry — and may have done truly massive harm to American business interests.
We are now starting to see the consequences.
This is a pretty striking shift toward Chinese models by American AI startups since the start of the year. https://substack.com/@profgmarkets/p-200029541
Positive users praise Chinese LLMs for real utility and open-source competition while negative users insult the surge in US adoption as driven by greedy expensive American AI or dismiss the topic outright.
see eg https://www.scmp.com/tech/tech-trends/article/3271858/ai-race-alibaba-tencent-quickly-adopt-metas-new-llama-31-model-amid-excitement and https://medium.com/the-endless-forge/zuckerbergs-llama-fiasco-did-meta-ignite-china-s-ai-takeover-a388ffc37aaa though the full story is perhaps not known
Zuckerberg and LeCun’s unilateral decision to open source Llama likely (partly) catalyzed China’s AI industry — and may have done truly massive harm to American business interests.
We are now starting to see the consequences.
American AI startups are routing far more app traffic to Chinese LLMs.
Through OpenRouter, weekly token consumption was mostly driven by U.S. models through much of 2025. But from early 2026, Chinese models suddenly became the main growth engine.
AI model market is becoming less about brand loyalty and more about raw utility.
This is a pretty striking shift toward Chinese models by American AI startups since the start of the year. https://substack.com/@profgmarkets/p-200029541
@nxthompson OpenRouter doesn't really make sense for US models, you can use Anthropic/OpenAI/Azure/AWS directly for that. So I wouldn't read much into the US/Chinese ratio on OpenRouter.
@nxthompson It is quite disappointing to see that the US 'Big 3' currently hold only a 35% market share. In my view, Gemma 4 31B has come the closest to rivaling the Chinese models. To capitalize on this, Google needs to make a much stronger effort to deploy and promote it through gCloud.
@nxthompson OpenRouter is not a representative sample. It's mostly prosumers/solo devs, so biases heavily towards not just Chinese models but, e.g., DeepSeek Flash (not Pro). At companies, people are not just using Claude/GPT, but, e.g., Opus (not Sonnet). So literally the opposite.
@nxthompson Who actually supplies tokens for Chinese models
@nxthompson What could go wrong with speaking to the commies about proprietary stuff? Retards.
@nxthompson OpenRouter is where price-elastic builds live. The chart mostly shows where developers go when cost is the variable. Frontier usage doesn't clear the same channel. The barbell's showing up in the data. Thread:
@nxthompson CCP subsidy program working as intended.
@nxthompson This data is only for people accessing models through OpenRouter. ie: it's self-selected for queries that aren't using Claude, Gemini, or ChatGPT directly. I'm more likely to use Chinese models through something like OpenRouter, too, because I wouldn't want to use them directly.
@nxthompson Why are you assuming this is representative of the token consumption of American AI startups? Seems like a crazy jump to make...
@GaryMarcus DeepSeek could be the black swan. None of the frontier US companies is using Mixture of Experts (MoE) model which is good enough for most jobs at a fraction of the cost Claude Haiku 4.5: $1.00 per 1M input tokens $5.00 per 1M output tokens DeepSeek .14 per M input .28 per output
@nxthompson @grok which Chinese model is used most currently ?
@GaryMarcus It’s like saying the transformers paper should not have been released, or going back, NSF grant funded (and therefore mandated to be released to the public) research should not have happened in ANNs in the 80s/90s. Just a terrible take all around
@nxthompson Mirrors Chinese car production graph.
While I don’t disagree that Chinese models are gaining traction I think it’s worth mentioning that users tend to use individual platform APIs (Anthropic / OpenAI/ Google) when using closed models because there’s no cost advantage to using open router for those. So the conclusion here is that Chinese OPEN SOURCE models are beating us ones - which isn’t surprising given that Meta changed its strategy with Llama
@nxthompson Only like 20% of users buy iphones but aapl is worlds largest phone co
@nxthompson It doesnt say "among American AI startups" anywhere on here.
@drorpoleg will see how things go now that tokenmaxxing has declined.
@GaryMarcus I only use U.S. models when I need SOTA, mostly programming tasks. Otherwise I use locally hosted qwen or similar.
Gary Marcus blames Meta's open-source strategy for the shift
Zuckerberg and LeCun’s unilateral decision to open source Llama likely (partly) catalyzed China’s AI industry — and may have done truly massive harm to American business interests.
We are now starting to see the consequences.
This is a pretty striking shift toward Chinese models by American AI startups since the start of the year. https://substack.com/@profgmarkets/p-200029541