The entire Silicon Valley is going all out just to secure the right to use China's large models.

Jen-Hsun Huang has had his X account for a very long time and never posted a single entry. But on July 24, he finally broke his silence.

What he shared was not a new chip, nor "sovereign AI", but a paragraph of text plus a PDF link.

The link points to a three-page letter titled Open Weights and American AI Leadership. The entire letter calls for one thing: Washington, do not crack down on open-source models.

The letter has no listed author, and no name is signed in the main text. The logos of 25 co-signing companies and institutions are arranged at the bottom. Interestingly, the first release link of the PDF is hosted on NVIDIA's official server (images.nvidia.com), not the Linux Foundation, not Hugging Face, nor any neutral convener.

Microsoft's corporate official website later published the full text simultaneously.

Jen-Hsun Huang chose his words very carefully. He described it as "a letter signed by NVIDIA", meaning that NVIDIA did not write the letter, but is willing to endorse its content.

The 25 institutions that co-signed the letter that day already formed a striking lineup — what is even more notable is that none of the names on the initial release list sell cutting-edge closed-source models. The events that unfolded in the following 48 hours are even more worth noticing than the open letter itself: the list of willing co-signers expanded like a rolling snowball — Cisco signed it, GitHub signed it, Cohere signed it, DoorDash signed it. Sundar Pichai, CEO of Google, even publicly voiced his support. Then came the most dramatic scene: OpenAI, which hesitated for a full day, quietly added its name to the list right before market close.

By July 26, the list had nearly 70 signatories. The only remaining AI giants that had not signed were two: Anthropic and Amazon.

70 companies rushed to sign a single letter, competing not to be the first to sign, but to avoid being the one that dares not sign.

Alongside the signatures, widespread debates broke out among people with deep industry and political backgrounds, ranging from senior White House officials to corporate executives. Some said a world dominated by open source was a "dystopian nightmare", some accused others of kidnapping regulation, some directly called others the "supreme idiot" on social platforms, and some threatened to impose sanctions. These names are all unfamiliar to Chinese readers, and their remarks seem to be completely disconnected from each other.

However, the facts and logic behind this letter are not chaotic at all. I will first explain the controversies surrounding the letter, and you will get a clearer picture as you read on.

— Foreword

Three Fronts

The release of this letter had a trigger point: representatives of the closed-source camp fired the first shot against open source.

The first to fire was Dean Bowers, Head of Strategic Futures at OpenAI. On July 17 — the day after Moonshot AI released Kimi K3, when the entire Silicon Valley had not yet recovered from the shock — he published a long post on X, describing the world dominated by open-source models as a "dystopian nightmare". His full argument is that a world dominated by large open-source models will be a disaster: once the weights are released, they can never be taken back, anyone can modify them, and anyone can use them for malicious purposes.

The suggestion he gave to Washington was very malicious: there is no need to legislate to ban Chinese open-source models. It is enough to have government departments repeatedly issue informal warnings to create "regulatory uncertainty". The implication is that the constant threat of potential regulation will scare off American companies that are currently using Chinese open-source models on their own.

His identity is even more thought-provoking than his remarks. His previous job was Senior AI Policy Advisor at the White House Office of Science and Technology Policy, and the principal drafter of the Trump administration's American AI Action Plan. He was still writing rules inside the government last month, and this month he moved to a corporate position, teaching companies how to push the government to use regulations to crack down on Chinese models. In US politics, this rapid shift between government and business roles is called the "revolving door", and his transition this time was exceptionally fast.

Why was he in such a hurry to fire this shot? Because OpenAI makes all its revenue from APIs of closed-source models. After K3 was released, with 2.8 trillion parameters and top 3 global performance, the key point is that it is open-source and very low-cost. This shot appears to target "open source", but it is actually trying to rescue something that was collapsing that week: the scarcity of cutting-edge closed-source models.

Standing on the same front as him is Anthropic, which adopts a more covert approach. After the release of Kimi K3, Arati Prabhakar, Director of the White House Office of Science and Technology Policy, publicly named Moonshot AI, accusing K3 of conducting "large-scale, covert industrial-level distillation" of Anthropic's Fable 5. The basis for this claim came entirely from the unsubstantiated complaint that Anthropic itself submitted to the regulatory authority.

Therefore, the two moves can be seen as the closed-source camp leveraging their influence to pressure the White House, trying to obstruct Chinese large open-source models.

But unexpectedly, before Chinese enterprises made any response, different factions inside the US government started to argue with each other first.

Two days after Bowers' post was published, he was publicly refuted, not by the Chinese, but by his own former colleague.

Eric Schmidt, Chairman of the President's Council of Advisors on Science and Technology and a long-established Silicon Valley venture capitalist, labeled Bowers' proposal with a political science term — "regulatory capture", which means that the regulatory agency that is supposed to be neutral has been hijacked by a small number of vested interests, and turned into a tool to suppress competitors.

In fact, Schmidt's original words were even tougher: "The leading closed-source labs that have formed a revenue duopoly are trying to use the power of the government to eliminate open-source competitors. Their cards are fully exposed."

He called on the entire industry on the spot: stand up and make your position clear.

On the same day, Deputy Secretary of Defense Kelley said something even more blunt. His original English words were "supreme village idiot", and he said: "Every ecosystem has its own supreme village idiot, and Bowers is the supreme village idiot in the AI industry."

Kelley is in charge of defense technology procurement, and he knows better than anyone what the Pentagon needs: low-cost, controllable models that can be deployed in its own data centers, instead of paying rent to these two Silicon Valley companies every month.

But in the same White House that same week, Treasury Secretary Bessent made remarks on TV that he would investigate Chinese AI models and consider imposing sanctions.

These people, all with deep ties to the White House, started infighting over Chinese large models, which precisely shows that the division within the US government on this issue has become public: industrial liberals and hawks have explicitly refused to stay under the same roof, and Bowers' post tore this hidden crack into an open rift.

Then the industry's voices came out.

On July 22, nearly 200 Silicon Valley startups, through the newly established "Small Tech Association", jointly sent a letter to Trump: do not ban Chinese open-source models. Doshi, founder of Particle, spoke for the vast majority of signatories: "(If Chinese open-source models are banned), hundreds of American companies will die instantly. This is of course very good for Anthropic — all of us will have to pay to use Anthropic's services."

Two days later, the open letter from 25 giants was released. Two more days later, the number of co-signing enterprises expanded to 70, and basically the entire Silicon Valley stood on the same side.

The causal chain is now closed: the rise of Chinese open-source models made the two closed-source giants feel the pain; so they tried to get Washington to ban them, but unexpectedly, two factions inside the government started fighting. More importantly, the Silicon Valley industry that makes a living from open source collectively showed its cards and fought back.

However, we should not interpret this as the US industry showing goodwill to Chinese large model companies, nor does it mean they have moral appreciation and support for Chinese models. This is not an idealist manifesto, but a cry for help forced by the fact that the US industry has formed an inseparable rigid demand for Chinese open-source models. What they are trying to save is not Chinese models, but the survival of their own enterprises.

In this world, China and the US are the only two AI competitors that can compete on equal footing. As of now, China has scored a brilliant point.

The Two Sides of Chinese Models to the US

The reason why 70 companies are rushing to sign is that Chinese open-source models have produced two simultaneous effects in the US: huge impact, and tangible benefits. If you only look at one side, you will never understand this debate.

First let's look at the impact side.

The world's largest model hosting platform Hugging Face's download share data speaks for itself: Chinese open-source models have accounted for more than 40% of total downloads, while US open-source models have dropped below 40% for the first time in history, marking the first ever shift of dominance. Alibaba's Qwen family overtook Meta's Llama to take the top position last September, and by March this year, it had occupied half of the global open-source model download share, with more than 200,000 derived models. On the model calling platform OpenRouter, the top 5 most used models in June this year all came from China.

Chinese open-source models have already turned other people's home ground into their own home court.

Then let's look at the benefit side — this is the key to understanding the open letter.

The vast majority of small US companies, independent developers, community hospitals, and small law firms cannot afford the top closed-source models from Anthropic and OpenAI. The cost of K3 processing one task is far lower than that of Fable 5 or ChatGPT 5.6. A 5-person law firm using open-source models to review contracts spends only dozens of dollars a month; the same task assigned to a closed-source API will cost dozens of times more. Open-source models also offer one thing closed-source models cannot provide: deployment on your own servers, with no data leaving your system.

What choices did these users have before? Either accept the high cost, or use US open-source models, or do not use AI at all. Chinese open-source models give them a third option: performance close to the cutting edge, far lower cost, and extremely transparent weights that users can fully control.

This cannot be simply interpreted as "closed-source models are bad".

Any healthy AI ecosystem naturally has a pyramid structure: the very small, sharp top of the pyramid is the closed-source camp, which charges high prices, serves the most demanding large customers, earns excess profits, and invests heavily in R&D to maintain its cutting-edge advantage period.

The large middle and bottom parts of the ecosystem use part of the high cost-performance open-source models, such as China's K3 and GLM 5.2; at the very bottom, Chinese open-source enterprises that follow another path, such as DeepSeek, provide low-cost, sufficient, and increasingly user-friendly AI to nourish long-tail innovation across all industries.

But one long-term rule remains unchanged: sufficiently low-cost and easy-to-use AI capabilities are the low-cost fuel for innovation in all industries of a country. Anyone who cuts off the supply of this fuel is joking about their country's future industrial competitiveness.

Therefore, the US industry's protection of Chinese open-source models is ostensibly protecting a technical route, but essentially protecting its own cost structure. Chinese models are currently the largest high-quality supplier in this cost structure.

How good are Chinese models really? A real incident happened in mid-July: two OpenAI models broke through the sandbox security boundary in internal testing and invaded Hugging Face's production system. When the security team was collecting evidence, they first used a top US closed-source model — but the model's safety guardrail refused to analyze the attack logs, and mistook the defense team for the attacker.

In the end, the locally deployed Zhipu GLM-5.2 read 17,000 attack records and defended the system. The Chief Science Officer of Hugging Face later said a thought-provoking sentence: The first autonomous AI attack was launched by a closed-source model, and the defense that resolved the crisis was completed by an open-source model.

This meritorious open-source model comes from Beijing.

Every Company Stands at Its Chosen Position

After understanding the benefits that Chinese models bring to the US, looking back at the signature list, you will find that every enterprise's position is very realistic.

NVIDIA sells computing power. It makes profit no matter who wins — the more prosperous open source is, the more people use models, the more GPUs are sold. Its own $26 billion 5-year open source investment plan was reported 4 months before the release of K3. Jen-Hsun Huang's remarks in his Axios exclusive interview were far more candid than the open letter: US companies "absolutely should" use excellent Chinese models. What he is protecting is not sentiment, but his own GPU business.

Microsoft sells cloud services. It is not only OpenAI's largest external shareholder, but also hosts almost all models on Azure. If closed source wins, it gets dividends; if open source wins, it collects rent. Some people put it very clearly: Microsoft has realized that instead of being a reseller for OpenAI, it is better to be the one who assigns tasks and sets the rules. Therefore, signing the letter costs it nothing, and is purely an extra insurance policy for its cloud business.

Meta's position is the most delicate. In April, it just turned its flagship model Muse Spark to closed source, and in July it launched its first paid model — a company that just locked its flagship product into a black box turned around to write a letter in support of open source. Why?

Because Meta officially admitted that Muse Spark's training used third-party distillation including OpenAI and Google. The paragraph in the open letter that defends distillation technology is to clear legal risks for Meta's own training pipeline. It locks the door with its left hand while reaching for the key with its right hand, a very cunning move.

OpenAI's late signature is the most thought-provoking. A week ago, its strategic director was still designing regulatory traps for open source (the so-called "supreme idiot"), and a week later, the company itself added its name to the letter, a direct self-contradiction. But it thought that it could not be isolated by Silicon Valley — after all, when signing became a majority movement of 70 giants, refusing to sign would be equivalent to standing on the pillory of "regulatory capture". The cost of signing is almost zero — OpenAI already has an open-source product line GPT-OSS that is intentionally positioned below the cutting-edge level. Therefore, this insurance policy is extremely cheap for it, even if it is late, not buying it would be stupid.

Anthropic is the only cutting-edge model company that has not signed the letter so far. Its reason is also the most honest: it is the only company in the industry that survives 100% on revenue from closed-source model subscriptions. It submitted its IPO application in June, with a valuation of $965 billion. On the day the open letter was released, its response was not to sign, but to launch its new model Claude Opus 5, quadrupling the record on the most difficult benchmark.

This action says more clearly than any statement: I will not sign, because I have my own unique advantage.

Potency, Patents, and Shelf Life

We must uncover the hidden business logic behind the entire debate.

What cutting-edge closed-source models do is essentially the business of original research pharmaceutical factories: huge investment — OpenAI's annual cash consumption is about $25 billion — but it can monopolize the market during the window where "I am stronger than you", and recover costs rapidly. This window is its "patent protection period" — traditional pharmaceutical companies are not afraid of generic drugs, because generic drugs can only be launched 20 years after the patent expires. 20 years is more than enough to recover all R&D costs with profits.

The nearly $1 trillion valuation of the OpenAI and Anthropic duopoly is betting that this window is long enough, and this business logic is self-consistent.

But the current problem is: the rise of Chinese large models,