Silicon Valley's AI Incumbents and Startups Just Picked Sides Over China

On Wednesday, the Little Tech Association, a coalition of about 200 startups that includes Y Combinator, sent a letter to White House science adviser Michael Kratsios and Commerce Secretary Howard Lutnick asking the Trump administration not to ban Chinese open-weight AI models (WIRED). The letter landed as Anthropic, in a June post, publicly accused Alibaba of running a large-scale distillation campaign to extract its AI capabilities — a charge corroborated by CNBC, the BBC, and Business Insider. Two days before the letter, Kratsios posted on X that Moonshot AI, the Chinese lab behind the Kimi model family, had built its Kimi K3 by distilling Anthropic's Fable 5 through what he called a sophisticated internal platform for large-scale distillation against U.S. models.

Anthropic and the Little Tech Association now sit on opposite sides of the same fight, and neither is the underdog. Anthropic, the lead voice for the AI incumbents, argues that open-weight models spread without the safety guardrails US labs have spent years building. Open-weight models are AI systems whose core parameters, the numerical settings that determine how a model behaves, are published so anyone can download, run, and fine-tune them. Once a set of weights is on the internet, it can land on Hugging Face, a public hub where developers share models, on GitHub, in a third-party cloud, or on a laptop in a lab in Shenzhen, and the original lab cannot revoke it. That is the distribution surface the Little Tech Association is defending.

According to WIRED's reporting, Benchmark's Bill Gurley has framed the divide publicly as pro-diffusion startups against pro-regulation incumbents, and CSIS analyst Yasir Atalan has pointed out that open-weight models spread through Hugging Face, GitHub, cloud providers, and third-party inference platforms, the same surfaces that would make a US ban unenforceable in practice. The startup coalition's argument is that a ban would not stop the weights from circulating; it would only stop Americans from running them. The incumbents' argument is that without a ban, the safety work US labs have paid for gets stripped off in a fine-tune and shipped to users the labs never vetted.

Distillation is the operation both accusations describe. To distill a model is to train a smaller or cheaper model on the outputs of a larger one, so the student picks up the teacher's behavior without copying its weights. The accusation against Alibaba was that it ran a large-scale operation to harvest those outputs at scale through Anthropic's API. The claim against Moonshot, made in a single X post and not yet independently corroborated, is that the same playbook was used to compress Fable 5 into Kimi K3.

Both sides have commercial interests in their framing. Anthropic's safety case protects a proprietary moat, the kind of advantage that depends on keeping weights out of competitors' hands. The Little Tech Association's free-market case protects a distribution moat that depends on weights staying downloadable. A US ban on Chinese open-weight models would gate the distribution surface without closing the safety gap, because the underlying weights are already public and the same models can be fine-tuned in jurisdictions that do not participate in the ban. The letter's practical ask is to keep that surface open and let the market, rather than the Commerce Department, decide which models Americans run.

The next checkpoint is Kratsios's office. The White House has not said when it will respond to the letter, and the Commerce Department's Bureau of Industry and Security has not opened a public rulemaking on Chinese open-weight models. The first concrete answer is likely to come from whether a ban shows up in an executive order, an export-control rule, or an Entity List addition. Until then, the weights keep moving and both lobbies keep writing.