America Can't Ban What Its Own Leaders Praise

Washington is racing to sanction Kimi K3 — and the Americans it claims to defend won't close ranks. On July 21, OpenAI president Greg Brockman called Moonshot's model "pretty good," said distillation claims are "too early" to judge, and declined to join the ban debate. The next day, Nvidia CEO Jensen Huang told Axios U.S. companies should "absolutely" use Chinese open models. Treasury threatened sanctions anyway.

What Brockman said:

• "It's a pretty good model," Brockman told Bloomberg on July 21, referring to Kimi K3. "There's no question about it."
• Asked whether Moonshot distilled from OpenAI's systems, he said it is "too early" to tell — adding, "This is something we do always monitor for."
• Brockman estimated China could be as little as four months behind the U.S. in model development.
• He called the idea that open-weight models are free a "misconception": "These are still big models that run on lots of compute, and compute is expensive."
• Brockman, who with his wife donated $25 million to Trump's MAGA Inc. super PAC in September 2025, declined to wade into whether Chinese models should be banned. He said only that "America continues to lead" in AI.

What OpenAI's corporate line says:

• In a statement to Bloomberg, OpenAI said open models are important for "democratising access to AI."
• The company added: "Advances in Chinese open-weight models are not an argument against openness."
• That puts Brockman closer to Jensen Huang than to his own policy shop.

What Huang said the same week:

• "These Chinese models are excellent," Huang told Axios on July 22. "Open-source models that are excellent should be used."
• Asked whether American companies should be allowed to use Chinese AI models, he answered: "Absolutely."
• Huang rejected backdoor fears as a "misconception," arguing companies can fine-tune and sandbox downloaded weights.
• "There's no scenario where China runs U.S. companies off the road," he said. "Zero possibility."
• His economic logic is blunt: cheaper open models expand AI adoption, which means more Nvidia chips, more data centers, more services.

What Washington is pushing anyway:

• On July 22, White House science chief Michael Kratsios accused Moonshot of distilling Anthropic's Fable to build K3 and accessing banned Nvidia GB300 chips in Thailand.
• Treasury Secretary Scott Bessent posted: "Open source is not open season on American IP." He warned sanctions and Entity List designations are "on the table."
• Bessent had said July 21 that officials were "finding watermarks of our U.S. large language models on many of the Chinese models."
• Moonshot has not responded. The company plans to release Kimi K3's full open weights on July 27.

The model nobody in Silicon Valley will disown:

• Moonshot released Kimi K3 on July 16 — a 2.8-trillion-parameter model that Artificial Analysis scores at 57 on its Intelligence Index, ranking third overall behind Claude Fable 5 and GPT-5.6 Sol.
• K3 topped Arena's frontend coding leaderboard within 24 hours of launch, ahead of Anthropic and OpenAI flagships, according to The Next Web.
• Moonshot paused new Kimi subscriptions last week after demand maxed out its GPU cluster — the same weekend Bloomberg reported it is racing toward a $30 billion Hong Kong IPO.
• Brockman and Huang are not dismissing a toy. They are acknowledging a frontier-class product four days before it ships as a downloadable ZIP file.

The split inside OpenAI:

• Dean Ball, OpenAI's head of strategic futures and a former White House AI adviser, has argued the U.S. should restrict or effectively ban Chinese open-weight models, TechCrunch reported July 22.
• Ball's position aligns with closed-lab lobbying in Washington. Brockman's public praise does not.
• The same week Axios reported OpenAI and Anthropic are aligning to warn policymakers about Chinese open-weight models, Brockman refused to join that chorus on camera.
• Anthropic thanked Kratsios on X for naming Moonshot. OpenAI's president would not confirm the theft.

Why the fracture matters for markets:

• OpenAI and Anthropic are both racing to prove revenue models ahead of expected IPOs. Kimi K3 prices inference at a fraction of U.S. frontier APIs.
• Washington's sanctions threat re-prices procurement risk. Silicon Valley's refusal to vilify the model re-prices the probability any ban sticks.
• If the lab Treasury is defending and the chipmaker every hyperscaler depends on both say use it, compliance officers — not policymakers — become the enforcement layer.
• Alphabet and Tesla both reported negative free cash flow July 22 while raising AI capex. Investors are already asking whether trillion-dollar rack bills can survive commoditized models. A sanctions war the industry won't fight makes that math worse.

What enterprises should watch:

• July 27 is the enforcement deadline that isn't one — once weights are public, download bans are performative.
• Named accusations without public evidence still freeze pilots at regulated firms. Treasury designations attach to vendors, not GitHub repos.
• The fight is no longer U.S. vs. China. It is Washington vs. its own AI-industrial base over whether openness is a national asset or a national threat.

Convina's view: Washington wants a sanctions coalition. Silicon Valley just told you it doesn't exist. Brockman — who gave Trump's super PAC $25 million — praised the model Bessent wants to blacklist and wouldn't confirm the crime. Huang, whose chips power every side of this fight, told American firms to run it. OpenAI's corporate statement explicitly rejects using Chinese progress as an excuse to close weights. You cannot enforce a ban the defendants won't endorse, especially four days before the product becomes a free download. Treat July 27 as a procurement event, not a policy finale — the geopolitical discount on Chinese open weights just collided with the people who build the racks, and they picked openness.