Silicon Valley's smallest companies and its largest AI labs are now openly at war over whether the U.S. government should block access to Chinese open-weight models, and the stakes are higher than most AI policy fights get.
On July 22, the Little Tech Association sent letters to President Trump, Commerce Secretary Howard Lutnick, and White House OSTP Director Michael Kratsios signed by nearly 200 venture-backed startups. The ask was simple: don't restrict access to Chinese open-weight AI models, including those from Moonshot AI and Alibaba. The argument was sharper: a blanket ban wouldn't stop China. It would stop them.
Harry Godfrey, the association's executive director, called for "a scalpel rather than a sledgehammer," arguing that any security measure should be targeted at specific bad actors rather than the category of open-weight models as a whole. The letter cited the June 2026 export controls on Anthropic's Claude Fable 5 as a model: restrictions applied to named models, not to an entire class of technology that thousands of American startups depend on to build products and stay competitive with larger rivals.
The trouble is that the firms on the other side of this debate aren't exactly neutral parties. Anthropic sent its own letter to Capitol Hill in June accusing Alibaba's Qwen team of running what it called the largest known distillation attack on its Claude models, describing 25,000 fraudulent accounts used to illicitly access its API and 28.8 million exchanges scraped in the process. OpenAI aligned alongside Anthropic, and together the two most powerful closed-model labs in the world are now pushing Washington toward the exact restrictions that would tighten their grip on the U.S. market. The startups aren't wrong to name that conflict of interest plainly.
The policy fight sharpened further when Moonshot AI released Kimi K3 in July. The model is a 2.8-trillion-parameter open-weight system that the company claims outperforms every rival except Anthropic's Claude Fable 5 and OpenAI's GPT-5.6. Kratsios told the press that Moonshot had distilled Fable to build it, and that the company had developed a sophisticated internal platform to run large-scale distillation attacks while switching between access methods to evade detection. He also alleged Moonshot had acquired servers equipped with Nvidia's GB300 chips and accessed the same hardware in Thailand. Treasury Secretary Scott Bessent followed with a threat of sanctions and export-control blacklisting.
Moonshot hasn't confirmed the allegations. Some experts have noted the timeline is awkward: Fable has only been publicly available since July 1, which doesn't leave much runway to distill a 2.8-trillion-parameter model from scratch. That skepticism doesn't mean nothing happened. It means the evidentiary bar matters, and the White House hasn't fully cleared it in public yet.
Frankly, the distillation question and the open-weight question are being run together in Washington in a way that serves the incumbents more than it serves clarity. Distillation of a proprietary model without authorization is IP theft, full stop. Accessing an open-weight model released to the world is something else entirely. Treating them as the same problem to justify the same policy response is exactly the kind of imprecision that tends to benefit whoever already has market share.
Who actually pays if the ban goes through
The startups signing the Little Tech letter aren't just making a theoretical argument about competition. They're describing their cost structure. Open-weight models from Chinese labs, including Alibaba's Qwen series and Moonshot's Kimi lineup, are among the most capable and cheapest options available to a small team building an AI-native product. Cutting off access doesn't redirect founders toward some neutral alternative. It redirects them toward OpenAI and Anthropic, the exact companies lobbying for the restrictions.
Anthropic has already updated its terms of service to block access for any company with majority Chinese ownership, defined as at least 50 percent. U.S. Trade Representative Jamieson Greer has called distillation a form of IP theft. The government's posture is clearly moving toward restriction. Whether that restriction stays surgical or becomes categorical is the question the Little Tech letter is trying to force into the open before the decision gets made quietly.
What's unusual about this fight is how clearly the commercial incentives line up. The large labs want protection from Chinese competition and can dress it as national security. The startups want access to cheap, capable models and can dress it as innovation policy. Both framings are partially true and partially self-serving. The difference is that Anthropic and OpenAI get regular meetings in Washington and the 200 startups had to organize a letter-writing campaign to get heard at all.
The White House, for its part, reportedly did not seriously consider a blanket prohibition during internal discussions. That's a narrow comfort for founders. A targeted restriction on specific models or specific uses could still close off a meaningful part of the open-weight market without ever being called a ban. The startups know this. Their letter isn't really about the word "ban." It's about who controls what the next generation of AI products gets built on, and who in Washington gets to decide.
Also read: Uncle Bob Martin says he no longer reads AI-generated code and the developer world is split • Samsung and Broadcom's $200 billion chip pact is a direct challenge to TSMC's grip on AI manufacturing • Etched raises $300M at a $10.3B valuation on a bet that transformer-only chips will displace Nvidia at the inference layer
