Nvidia Is Quietly Betting Against OpenAI’s Push to Ban Kimi K3

Moonshot AI’s Kimi K3 has 2.8 trillion parameters and undercuts Anthropic and OpenAI on price by a wide margin. Axios reported this week that the Trump administration is reviving plans to restrict Chinese open-weight models like it inside the United States, though Politico says the Commerce Department will not act anytime soon.

The fight already cost one government official a job, and it has surfaced a conflict of interest nobody in Washington has had to answer for yet. Nvidia, the company with the most money riding on the outcome, is quietly betting against the labs pushing for a ban.

A Beijing Benchmark Reopens an Old Washington Fight

Moonshot AI, a Beijing-based startup backed by Alibaba and Tencent, released Kimi K3 on July 16. The model activates just 16 of its 896 experts for any given token, yet still reasons over a 1-million-token context window and costs $3 per million input tokens and $15 per million output tokens to run. Moonshot says the full weights, which would let anyone download and self-host the model rather than rent it through an app, will post publicly on July 27.

That pricing runs roughly a third of what Anthropic charges for its flagship Claude Fable 5, which costs $50 per million output tokens on comparable work, according to a pricing comparison reported by Tom’s Hardware. The gap is the whole argument for OpenAI’s head of strategic futures, Dean W. Ball, who argued the government should manufacture regulatory fear around open-weight models, reasoning that cheap competitors would deter the capital spending frontier labs still need. A source close to the administration told Axios that a leading AI lab or its allies pitch some version of that idea to the government every three to five months.

Braden Hancock, the co-founder of Snorkel AI and a research partner at the Laude Institute, does not dispute the economics. “Strong, frontier-caliber open source models will place a squeeze on the margins and will bring down the prices of the frontier companies,” Hancock told TechCrunch. He does not think total AI usage falls as a result, calling that outcome “quite the opposite” of what actually happens.

Ball’s Argument Didn’t Survive the Weekend

The pushback was immediate. Yann LeCun, the Turing Award-winning AI researcher, argued that open and proprietary AI can grow side by side rather than trade off against each other. Martin Casado, a general partner at the venture firm Andreessen Horowitz, made the case that open source has historically expanded whole industries instead of shrinking the companies built on top of them.

Ball retracted his position within days. He walked back the claim that a regulatory crackdown was the White House’s “best strategy” and that open-weight models necessarily slow down the technology’s advance.

The underlying question stuck around. Axios reported this week that the administration is weighing whether to add Chinese AI labs to the Commerce Department’s Entity List, a trade blacklist that would force American companies to get a license before using their technology. Politico’s reporting says Commerce is not ready to pull that trigger.

The Fight Has Already Cost Washington a Job

Chris Fall, the director of the Commerce Department’s Center for AI Standards and Innovation (CAISI), resigned on July 20, just three months after taking the job, the department confirmed. CAISI is the agency responsible for testing and evaluating advanced AI systems, including the Chinese models now at the center of the ban debate.

Fall’s exit follows a pattern. His predecessor, Collin Burns, a former researcher at Anthropic and OpenAI, left within a week of being named to the post in April, after sources told The Washington Post he had been pushed out over his past ties to Anthropic.

Before either of them, the job belonged to David Sacks, the venture capitalist and Trump adviser, who held the title of White House AI and Crypto Czar before stepping down from that role in March. Nobody has filled it since.

1. July 16: Moonshot AI releases Kimi K3, calling it the largest open-weight model built to date.
2. July 17: Sacks tells Axios that regulatory fear is “how you lose the AI race,” as the first reports of a renewed restriction push surface.
3. July 20: Axios reports the administration is reviving plans to restrict Chinese AI models, Politico reports Commerce will not act soon, and Chris Fall resigns as CAISI director.

That leaves the agency without permanent leadership while it is still supposed to be evaluating whether Kimi K3 and its peers pose a real security risk.

Nvidia Wants More Kimis, Not Fewer

Nobody in this fight has more money riding on the outcome than Nvidia, and Nvidia is not asking Washington for protection. It is doing the opposite, building its own open-weight models because a market with only two or three well-capitalized AI labs is worse for chip sales than a crowded one.

Nvidia debuted its Nemotron 3 family of open models for agentic AI workloads and has kept expanding the lineup with speech, safety and search models since. “Open innovation is the foundation of AI progress,” said Jensen Huang, Nvidia’s founder and CEO, when the family launched. Developers who fine-tuned Nemotron on Nvidia’s own Blackwell chips have reported running it for roughly 90 cents per million output tokens.

Hancock says the logic is simple. Nvidia does better, he told TechCrunch, “if there are dozens or hundreds of companies building AI than rather than two or three that are well capitalized enough to make their own chips,” which is one reason behind its investment in Nemotron. A handful of dominant labs buying chips is a worse customer base for Nvidia than a crowded field of open-model builders who are all still buying chips to compete on top of them.

The four companies at the center of the fight are not aligned the way the ban debate assumes.

That is a very different bet than the one OpenAI and Anthropic are making inside the same industry.

Could Chip Controls Succeed Where a Model Ban Would Fail?

Sam Bresnick, a China-focused research fellow at Georgetown’s Center for Security and Emerging Technology (CSET), says restricting Chinese chatbots does less than restricting the Nvidia chips that train them. Washington already limits which processors can reach China. Tightening that chokepoint, he argues, would matter more than trying to ban software millions of people can already download.

Bresnick asks why the government should intervene on the software side at all, questioning the fairness of protecting American labs “from competitors that are being locked out from the U.S. market based on their origins.” A better fix, he argues, is to stop selling Nvidia’s H200 processors to China outright, rather than policing software already running on servers worldwide. “That,” he says, “could potentially keep us out of this thorny debate about banning open source technologies that huge numbers of US companies want to use.”

The main point is the U.S. would be very well served to have its own very capable, much less expensive open models. It just clashes with the approach the frontier labs have taken.

Bresnick made that point to TechCrunch. No cheap, capable American open model exists yet at Kimi K3’s scale, which is part of why restricting the Chinese versions does not solve the underlying problem by itself.

The Guardrails Cut Both Ways

Officials also point to safety guardrails built into American frontier models through an opaque government-directed process meant to stop US systems from being used to break into computer networks or help design weapons. Chinese models carry no equivalent constraints, which officials frame as a risk. Sacks has been posting examples of security teams turning to Chinese models to close gaps when US frontier systems refuse the same task.

Strip away the competitive interest and the security case against Chinese models comes down to three claims, none of them fully settled:

• Data exposure: the fear that Chinese-made models could funnel US data back to Beijing, similar to the reasoning behind the US ban on importing modern Chinese electric vehicles. Researchers who study the models say those running on US servers are unlikely to phone home, though nobody calls it impossible.
• Ideological bias: the concern that a model trained under Chinese government constraints carries a hidden slant, though it is unclear what that would even mean for something like a coding task.
• Missing guardrails: the argument that Chinese labs skip safety testing US frontier models go through, a complaint that cuts the other way when American companies need a model willing to do sensitive but legitimate work.

Coinbase is one company already living inside that tradeoff. Its chief executive, Brian Armstrong, has said the exchange runs models including GLM-5.2 and Kimi in production, cutting its overall AI spending nearly in half even as actual token usage climbed, according to Tom’s Hardware.

Graduate Students Are Already Voting with Their GPUs

Advocates for open models argue the frontier labs have invented a false choice between innovation and closed weights. Hancock says the real risk from Chinese open models is not hidden code. “The bigger impact of having these open source models come from China is less that they’re sneaking in back doors, and more that they are owning the innovation,” he told TechCrunch.

He points to PyTorch, Meta’s open-source machine learning library, which pulled the research community toward one shared standard. “PyTorch became the industry standard because it was open source,” Hancock said, and rival frameworks from closed labs faded by comparison.

He and other advocates worry Chinese LLMs are on the same path. US graduate programs already build primarily on open-weight Chinese models, and Hancock estimates half the papers students study now come out of Chinese institutions, even as American frontier labs grow more guarded about publishing their own research.

That undercutting pattern is not new. It is already pulling US startups toward cheaper Chinese alternatives on cost grounds alone, well before Washington votes on anything.

Clem Delangue, the CEO of Hugging Face, a platform for open AI collaboration, says restricting access would not fix what worries people about these models in the first place. “Restricting open models wouldn’t make AI safer,” he said. “It would simply hide the risks, concentrate power in the hands of a few and make it harder for the next generation of builders, researchers, academia, non-profits, governments to participate in making AI safer and more beneficial for all.”

Moonshot says the rest of Kimi K3’s weights post publicly on July 27. Whatever Washington decides before then, the model will already be sitting on servers anyone can reach, in a country that requires no license to download it.