"Letting open models run wild only benefits China." Just as such concerns were spreading across the AI industry, China's Moonshot AI released "Kimi K3," a 2.8-trillion-parameter open-weight model (a method in which the trained model's internals are released for free, allowing anyone to modify and run it) on July 16, 2026, further intensifying this mood. Yet on July 24, NVIDIA CEO Jensen Huang took a different stance in his first-ever X post, expressing support for an industry letter titled "Open Weights and American AI Leadership," signed by 25 organizations including NVIDIA, opposing tighter regulation of open-weight models. The letter's signatory list does not include OpenAI, Anthropic, or Google. Behind this pattern of non-participation, NVIDIA's own business calculations become visible.
Jensen Huang Breaks His Silence: What's Inside the Letter Backed by 25 Organizations
Multiple media outlets reported the fact that Huang had never posted on X before as "an unprecedented break in silence." The occasion he chose for his first post was neither a product announcement nor an earnings report, but a statement of opinion on regulatory debate. The letter "Open Weights and American AI Leadership" was signed by 25 organizations including NVIDIA, Microsoft, Meta, Palantir, Hugging Face, IBM, Andreessen Horowitz, Perplexity, Linux Foundation, Mozilla, Mistral, CrowdStrike, Dell, ServiceNow, and Y Combinator.
In his post, Huang stated, "Open models strengthen safety and cybersecurity, accelerate innovation and adoption, and enable sovereignty. The world needs both frontier closed models and frontier open models." Against calls for tighter regulation on security grounds, he presented the opposite argument: that regulation itself would undermine security. This argument itself is not new. It simply applies the logic long repeated in the open-source software world—that transparency enhances safety—to AI models. Things get complicated the moment you look at who is voicing this logic most loudly.
The points raised in the letter can be organized into roughly three arguments. From a security standpoint, the claim is that models whose internals are public make it easier for third parties to discover vulnerabilities. From the standpoint of the pace of technological innovation, the logic is that companies and researchers no longer need to rebuild foundational models from scratch and can instead focus on application development. And from a sovereignty standpoint, the argument is that this leaves room for countries to operate AI without entrusting their domestic data to foreign clouds. Each of these carries a certain persuasive force, but the fact that the companies championing these three benefits most loudly are GPU sellers cannot be ignored when gauging the neutrality of the claims.
The 2.8-Trillion-Parameter Kimi K3 Evokes Memories of January 2025
Just before the letter was published, what had been heightening industry wariness was Moonshot AI's Kimi K3. The Beijing-based company announced on July 16 a 2.8-trillion-parameter open-weight model, described at the time of release as one of the largest open-source models ever. It is said to be roughly 75% larger in scale than the comparison point, DeepSeek V4 Pro (approximately 1.6 trillion parameters). The full weights were scheduled to be released by July 27, and the reason industry wariness intensified was the string of assessments that its performance was closing in on top-tier closed models.
A scale of 2.8 trillion parameters demands massive memory and computational resources just to run. For individuals or small and mid-sized businesses to operate Kimi K3 as-is is not realistic without a dedicated GPU cluster. Even if open weights embody the ideal that "anyone can use it," in practice, only organizations able to procure sufficient computing resources can fully reap the benefits.
On January 27, 2025, NVIDIA's stock plunged following China's DeepSeek announcing a low-cost AI model, and the company lost roughly $590 billion in market capitalization in a single day. Converted at the exchange rate at the time (assuming 1 dollar = 150 yen), this amounts to roughly 88.5 trillion yen, recorded as one of the largest single-day declines in U.S. stock market history. The scale of Kimi K3 evokes memories of this "DeepSeek shock." Each time a Chinese-made open-weight model closes the performance gap, it has left the market with anxiety over whether U.S. technological dominance is being shaken.
What NVIDIA lost in the January 2025 stock plunge was market capitalization, not GPU sales opportunities themselves. What DeepSeek's emergence spread was the concern that "massive amounts of GPUs may not be necessary for high-performance AI"—it was not that the existence of open-weight models itself threatened NVIDIA's business. Damage to market capitalization and damage to GPU demand are two different things.
The Calculation Behind OpenAI, Anthropic, and Google Not Signing
Looking over the letter's signatory list, one notices that OpenAI, Anthropic, and Google—the three companies whose main offerings are frontier closed models—are absent. Around July 21, U.S. Treasury Secretary Scott Bessent warned that if "industrial-scale distillation" by Chinese AI companies constitutes intellectual property theft, it could become grounds for sanctions or Entity List designation. Distillation refers to a technique in which a large volume of queries to a large-scale model and its responses are used as training material to teach a smaller model equivalent capabilities. Since it eliminates the need to gather training data from scratch, it can dramatically compress development costs, but it also tends to invite criticism from the original model's provider that their achievements are being freely exploited.
Bessent stated in an X post, "Being open source doesn't grant a free pass to harvest American intellectual property however you like." White House AI advisor Michael Kratsios also pointed out that Kimi K3 mimicked U.S. models through distillation techniques. Prior to this, in June 2026, Anthropic accused Alibaba's Qwen AI division of distilling from Claude at massive scale. Specific figures for the number of fraudulent accounts or communication instances vary depending on the report and will not be dealt with in this article, but the fact that Anthropic flagged this as an infringement on its own model is consistent across multiple reports.
Entity List designation is a sanction measure in which designated companies or organizations are, in principle, cut off from access to advanced U.S. technology. If the U.S. government applies this to Chinese AI companies, restrictions could extend not only to semiconductors but also to the use of software tools. The distillation controversy is both an ethical dispute and, in practical terms, a trigger for sanction risk.
For the closed-model camp, wariness toward distillation and support for stronger regulation are reasonable positions. If the situation is left unaddressed in which Chinese companies cheaply copy the output of models that cost enormous sums to train, the first-mover's competitive advantage will erode. It makes more sense to interpret OpenAI, Anthropic, and Google's non-signing not as opposition to the open-weight philosophy itself, but as a reflection of the fact that stronger regulation could serve as a means of protecting their own competitive advantage.
Looking at OpenAI's and Anthropic's revenue structures makes this interest even clearer. Both companies offer their models via API access or paid subscriptions, and charging fees while keeping the model's internals private is core to their business. The closer open-weight offerings come to matching this level of performance for free, the more the premise of this revenue model wobbles.
In Google's case, of parent company Alphabet's total revenue of $402.836 billion for the fiscal year ending December 2025, the advertising business accounted for $294.691 billion (approximately 73%), and revenue from the Gemini API alone is not disclosed. Even so, in wanting to maintain an edge while keeping its own Gemini model's internals private, Google shares interests pointing in the same direction as OpenAI and Anthropic. It is more natural to view non-participation in the letter as a practical decision reflecting each company's revenue-structure interests, rather than as a political statement in itself.
Behind the Open-Weight Threat Narrative, a Growing Pool of GPU Buyers
The letter argues that open models enhance safety and sovereignty, but it says nothing about what NVIDIA itself gains from this arrangement. The spread of open-weight models has the effect of distributing NVIDIA's GPU demand across a wider range than closed models do.
Here is how the mechanism works. Closed-model providers like OpenAI and Anthropic concentrate and secure computing resources with partner cloud providers such as Oracle, Microsoft Azure, or AWS, and sell access to users via API. In this case, the entities procuring GPUs in bulk are concentrated among a handful of major players: the model providers and their partner cloud operators.
Open-weight models, on the other hand, are downloaded and run on their own servers by companies, government agencies, and cloud providers individually. Because each user needs to fine-tune the trained model or run inference on its own data center, the entities purchasing and procuring GPUs expand beyond model providers to include companies, government agencies, and cloud operators. The stronger the demand for "sovereign AI"—countries wanting to run AI without sending their data abroad—the more this distributed purchasing behavior grows.
This alignment of interests is not limited to NVIDIA alone. Meta (developer of the open-weight model Llama) and Hugging Face (a company operating a distribution platform for open models) also appear among the letter's signatories. Both companies are also in a position where the more open weights spread, the more the value of their own products and platforms rises. Looking at the signatories as a whole, a picture emerges in which the majority are companies that stand to profit from the spread of open weights, rather than companies championing the open-weight philosophy itself.
What further reinforces this structure is the CUDA software environment NVIDIA has built up over many years. Many companies that run open-weight models on their own use tool sets optimized for CUDA in both training and inference. Regardless of which open model wins out in competition, much of the training and inference infrastructure ends up running on NVIDIA-made GPUs. Which open-weight model wins or loses in the competition is only a secondary matter for NVIDIA's revenue.
In other words, the letter's logic that "open models enable sovereignty" overlaps directly with the logic of expanding NVIDIA's customer base. The stated rationale of countering Chinese models and the practical benefit of expanding NVIDIA's own GPU sales channels arrive at the same conclusion. The letter says nothing about this alignment. Nor does it specify which regulatory provisions or bills the signatory companies have concretely in mind, and it makes no mention of the impact on NVIDIA's own sales to China. This is the one point the letter does not touch.
The Letter Arrived 10 Days After Testimony That H200 Shipments Were "Minimal"
The timing of the letter's publication also deserves attention. Since 2023, the United States has progressively tightened export regulations on AI semiconductors to China. Since August 2022, A100 and H100 chips have required licenses for export to China, effectively treated as near-total embargo.
The H20, introduced afterward, was also temporarily added to export restrictions in April 2025 after it was revealed that DeepSeek had used H20 chips to develop a high-performance model, and exports resumed later that year in July following Huang's visit to China. Since entering 2026, the focus has shifted to export of the higher-performance H200 to China. On May 31, the Department of Commerce announced new guidance to close loopholes around the higher-tier Blackwell model, tightening the regulatory net further.
Then, on July 14, at a hearing before the U.S. House Foreign Affairs Committee, Jeffrey Kessler, Under Secretary heading the Commerce Department's Bureau of Industry and Security (BIS), testified that although H200 export licenses for China had reached approximately $10 billion, actual shipments remained "minimal." Ten days after it became public that licenses had expanded while actual shipments stayed limited, Huang made his first-ever X post backing the letter defending open weights. With the export of semiconductors themselves still constrained, Huang chose instead to voice opposition to regulation in a different domain: AI models.
The two are separate regulatory targets, but for NVIDIA, the conclusion points in the same direction. If constraints on exports to China ease, the sales channel for its products expands; if regulation of open-weight models is held back, the base of purchasing entities expands. There is no direct evidence linking the two, but both align in a direction that pushes up GPU demand.
A Choice That Extends to Japanese Cloud Providers Too—What the Letter Doesn't Say
This confrontation is not a story that plays out solely on the other side of the Pacific. Japanese AI development companies, including those affiliated with NTT and SoftBank, already face a choice between using overseas closed models or running open-weight models on their own as the foundation for their services. If the push for open-weight regulation strengthens in the U.S., the very range of models available for selection could narrow, and domestic cloud providers' GPU procurement is structurally susceptible to regulatory trends via their reliance on NVIDIA products. For domestic startups building services on top of open-weight models, if U.S.-made models become harder to obtain, the alternatives could shift toward Chinese-made models. Stronger U.S. regulation could invite the very outcome the U.S. itself would most want to avoid.
The letter's claim that "the world needs both models" is not, in itself, off the mark from the standpoint of diversity in technology choice. But the fact that the party most enthusiastically championing this argument is a company that can sell computing resources regardless of which camp wins is a crucial piece of context that readers need in order to evaluate the letter.
Whether the letter actually moves policy will be tested by whether Congress actually takes up regulatory legislation that specifically names open-weight models, and whether the sanctions Bessent warned of are actually invoked. There is no clear timeline for when bill deliberations in Congress might begin, but sanctions do not need to wait for new legislation. Entity List designation is an administrative procedure decided by majority vote of the End-User Review Committee, an interagency body chaired by the Department of Commerce with voting members from the State Department, Department of Defense, and Department of Energy, and it can be enacted without any change in law. The Treasury Department is a non-voting consultative member, but Bessent's warning could serve as material bolstering the committee's judgment.
If Chinese AI companies' "industrial-scale distillation" is recognized as intellectual property theft, blocking access for the companies involved could be carried into effect through this committee's decision without going through congressional deliberation. The blueprint for who keeps buying GPUs could be set in motion by an administrative decision that bypasses Congress entirely.