Microsoft、Nvidia、Metaら24社がオープンウェイト支持、AnthropicとOpenAIは署名せず

imageThe fight over open-weight AI entered a new round over the last week.

On July 22, White House officials accused China’s Moonshot AI of stealing American IP through distillation and floated sanctions and repercussions. Anthropic’s policy chief called the alleged operation “industrial espionage.” Then the startups showed up, with nearly 200 co-signing a letter to President Trump urging him not to cut off the open models they rely on. On Friday, a statement signed by 25 major organizations warned Washington against “premature restrictions” on open weights. And Nvidia’s Jensen Huang, asked by Axios whether open labs should be allowed to distill closed models, answered that distillation is “fundamental to intelligence.”

Let me be clear: Running an open-weight Chinese AI model is not a fringe act, and it’s not disloyal. It’s how many AI-native startups and developers keep their token bills under control because Anthropic and OpenAI are becoming too expensive to build on at scale.

Washington wants to draw the line between legitimate openness and Chinese theft. Inside the technology industry, the more revealing line is open versus closed. Microsoft, Nvidia, Meta, and 22 other organizations signed the statement, warning against premature restrictions and defending distillation as a legitimate model-development technique. Anthropic and OpenAI did not sign.

For developers, the stakes are not abstract. The affordable model they can download, adapt, and run themselves is exactly what this fight could make harder to access.

Developers already picked open weights because of cost

Startups are on the front line here. America’s frontier labs priced many of them onto Chinese open weights. Nearly 200 companies, Y Combinator and Proton among them, told the President through the new Little Tech Association not to cut off Chinese open weights, and to use “a scalpel rather than a sledgehammer.” Particle founder Shuail Doshi put it plainly: “There’ll be hundreds of companies that instantly die. It’s great for Anthropic. We’re all going to have to spend money on Anthropic.”

Why the panic? It’s the price. I’ve been talking about this for months. Moonshot’s Kimi K3 is a near-frontier model that’s extremely powerful. My colleague Jessica Wachetel ran Kimi K3 head-to-head against Anthropic’s best coding model across a bug fix, a refactor, and a feature build; Kimi K3 wrote the same code as Claude Fable 5 for about a third of the cost. It was roughly four times slower.

Claude Fable 5 vs. Kimi K3: same code, a third the cost, four times slower

Model

Price (in / out)

3-task cost

3-task time

Code produced

Claude Fable 5Anthropic

$10 / $50

$5.98

~7 min

Baseline

Kimi K3Moonshot · open weight

$3 / $15

$2.13

~28 min

Same result

Source: The New Stack, Jessica Wachtel’s Fable 5 vs. Kimi K3 head-to-head. Prices are published API rates per 1M tokens (input / output).

For a startup counting tokens, four times slower is a fine price for a third of the bill. And they’re paying it. On OpenRouter, Chinese models have accounted for more than 30% of token usage by U.S. customers every week since February, peaking at 46%. OpenRouter is only one slice of the market, but it captures the direction. A meaningful share of developers has already voted with their tokens, routing real workloads toward much-cheaper Chinese models.

The pitch for Opus 5, which debuted on Friday, is almost all about cost. It matches or beats Anthropic’s pricier Fable 5 across most of the company’s own benchmarks while costing $5 per million input tokens and $25 per million output tokens, half of Fable 5 and the same as the older Opus 4.8. Anthropic’s case is cost per finished task, not cost per token: A model that solves the job in one pass can be the cheaper one even at a higher sticker price. It’s a fair point, and it helps narrow the gap. Anthropic cut the price of its frontier performance in half, and yet, it still charges more per token than Kimi K3.

The tech industry is split

So who actually wants powerful open weights restricted? Anthropic has taken the clearest position, and Axios reports that OpenAI has aligned with it in warning Washington about powerful Chinese open-weight models. The White House, meanwhile, is trying to draw a narrower line by defending open weights generally while punishing Chinese companies it believes used distillation to steal American capabilities.

Across from them are many of the companies that supply the chips, clouds, tools, and applications the AI economy runs on. The statement they signed, “Open Weights and American AI Leadership,” calls distillation “a widely used technique” and warns policymakers not to conflate legitimate model development with unlawful extraction. Its signatories include Microsoft, Nvidia, Meta, Hugging Face, IBM, Mozilla, Palantir, Perplexity, Mistral, and Y Combinator. Anthropic and OpenAI did not sign it. Right now, no ban has been drafted; the White House calls the reporting “baseless speculation,” and our own Amanda Caswell notes that the distillation allegation itself has not been backed by publicly verifiable evidence. But the sides are now on the record.

Huang, speaking to Axios, talked about whether open labs should be allowed to distill closed models and rejected the blanket theft framing. American companies should “absolutely” use excellent Chinese models, he said, and distillation itself is “fundamental to intelligence.” He added that privacy and contract violations should still carry consequences. But learning from other systems, in his view, is not theft. It’s how intelligence works.

Everyone here is talking about their book. Nvidia sells its chips to Anthropic and OpenAI as eagerly as to anyone running open models, and Microsoft is one of OpenAI’s largest investors. These are not companies with a simple grudge against the closed labs. They have money on both sides, and they still signed this letter anyway.

So look at the record. Distilling a model and pirating a book are not the same legal question. But both controversies began with the same accusation: An AI company benefited from material it did not own. The conduct and the law are different, and the hypocrisy is still hard to miss: A court this week approved Anthropic’s 2025 agreement to pay $1.5 billion for downloading and retaining pirated books (the court separately held that using books to train Claude was fair use). Anthropic has also spent more on federal lobbying in the first half of the year than in all of 2025. Its open-source record is real and beside the point: Anthropic backs the open source that expands Claude’s ecosystem, like MCP, while warning Washington about open weights that can compete with it.

It’s okay to use open weights

This is where it stops being Anthropic’s story and starts being yours.

A few weeks ago I wrote about the day a sudden government ruling forced Anthropic to pull Fable, and the developers who had built on it scrambled to swap in other models. They scrambled because Fable is a hosted service, which is the kind of dependency that can change without notice. Open weights are different. If you’ve already downloaded them and can run them on infrastructure you control, they’re far harder to take away than an API someone else owns.

The catch is that much of today’s Chinese-model usage still isn’t like that. Many companies and developers access Kimi and other open-weight models through hosted APIs that could disappear overnight if policy changes. Those developers would end up scrambling just like the Fable users did.

So don’t just experiment with an open-weight model. Download one. Run it. Know that your stack works before you need it. Not because Kimi is about to disappear, but because the tools you build on can now be switched off by people you’ve never met.

The post Microsoft, Nvidia, Meta and 22 others defended open weights. Anthropic and OpenAI didn’t sign. appeared first on The New Stack.