For three years Washington's AI plan could be written on a napkin: keep the best chips at home, keep them out of China, and let American labs turn that hardware lead into a model lead. This week a coalition of the country's largest technology companies effectively told the government that half of that plan is missing. Led by Nvidia, and signed by Microsoft, Meta and others, an open letter argued that American AI leadership will now depend on building a strong ecosystem of open models, the kind anyone can download and run. OpenAI and Google, whose businesses lean the other way, later added their names too. The message, as CNBC put it, exposes a blind spot: America has a chip strategy, and now it needs an open-model strategy.
The distinction matters more than it sounds. Models like ChatGPT and Claude are reached through services their makers control. Open-weight models can be pulled down, customised, and run on a company's own machines. That makes them cheaper, keeps sensitive data in house, and lets outside researchers inspect a powerful system for flaws. "The data that these companies have is really their strategic asset," said Vipul Ved Prakash of Together AI; ship it off to a closed model and you risk handing over your "business' recipe."
The awkward evidence sits in the traffic logs. On OpenRouter, which routes developer requests across many models, Chinese systems accounted for 48 percent of tracked usage in the last week of June, up from 20 percent a year earlier. American models fell to 32 percent from 74 percent. Labs like Zhipu and Moonshot have been releasing capable open models at a fraction of the cost of the leading US systems, and the world has quietly started building on them. China has every reason to keep pushing: each cheaper open model chips away at the premium American firms charge for closed access.
That is exactly why the incentives split. OpenAI and Anthropic depend on keeping their best work proprietary, and Anthropic's Dario Amodei has warned that a model anyone can download is also a model anyone can modify and is harder to monitor. Those concerns are not imaginary. But a parallel fight shows how the ground is shifting. Model distillation, the practice of training a small "student" model on a powerful "teacher's" outputs, has become a US-China flashpoint, with Anthropic accusing DeepSeek, Moonshot and MiniMax of trying to extract capabilities from Claude. The very act of accusing rivals of copying reasoning traces marks a change: the prize being guarded is no longer just the chips, it is the knowledge baked into the models.
So Washington faces a choice with two doors. The Trump administration is reportedly weighing restrictions on Chinese models. Yet a ban would not un-release weights already circulating worldwide; it would mostly leave American developers watching from the sidelines while everyone else keeps building on Chinese foundations. The alternative is to treat open models the way the country treated chips: give universities and startups real compute, steer government contracts to American open-model teams, and fund the security tooling needed to run downloadable models safely. The open question is whether a government that spent three years learning to say no can learn to out-build instead. As one line from the week put it, the future of AI may not be decided by who builds the smartest single model, but by who builds the models everyone else runs on. Right now, increasingly, that is China.