A moat is only worth defending if you own it. That plain fact runs under most of the American AI fight from the last week of July 2026, and it explains why Jensen Huang picked the document he did for his first post on X.
The post shared a three-page letter, "Open Weights and American AI Leadership," dated July 24, 2026, co-signed by about 25 companies and groups. Huang's line was that open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty, and that the world needs both frontier closed models and frontier open models. It drew more than 11 million views. For a man who had never posted before, that is a well-aimed debut.
The argument is clean. It opens on the 1980s open-source software movement, claims that shared code now carries most of the internet and much of U.S. government and military infrastructure, and calls this a shared foundation of institutional sovereignty. Open-weight models are defined as models anyone can download, inspect, modify, and run on their own hardware. Four benefits follow: cheaper access matched to the job, competition across the stack, customer control against vendor lock-in, and safety through transparency rather than obscurity. The letter concedes the real risk, that once weights are released they are beyond the developer's control and modified versions are hard to trace, then argues the answer is not prohibition.
The passage that actually fights a policy battle is near the end, on distillation. The letter defines distillation as using one model's outputs to help train or improve another, calls it a long-standing and legitimate technique, and asks that concerns about unlawful extraction be handled with targeted legal frameworks instead of sweeping bans. That is not an abstract worry. The letter landed days after the Trump administration was reported to be reviving a push to restrict Chinese AI models, after Moonshot AI's Kimi K3 reset expectations for what a downloadable model can do. Treasury Secretary Scott Bessent said the administration would examine Chinese models for IP theft and could sanction the companies, claiming U.S. watermarks turned up in Chinese systems. That is an official's assertion, not a proven fact, and worth treating as one. The letter never names China, Moonshot, or DeepSeek. It did not have to.
The signatory list is a map of who wants the gate open
Look at who signed and the story tells itself. Nvidia, Microsoft, Meta, Mistral, IBM, Dell, Palantir, CrowdStrike, Hugging Face, Mozilla, the Linux Foundation, Andreessen Horowitz, Perplexity, Replit, ServiceNow, Y Combinator. Chipmakers, server vendors, cloud operators, security firms, venture funds. Everyone who profits when AI gets cheaper and more portable, and nobody who sells access to a closed frontier model.
The conspicuous absences are OpenAI, Anthropic, and Google. Those are the three with the most to lose if open weights close the capability gap, because their premium comes from per-token access to something you cannot download. Sam Altman said he wanted the U.S. to win in both open source and proprietary models and was glad to see the letter, which is an endorsement of a document his own company did not sign. Some reports later said OpenAI added its name, leaving Anthropic as the lone holdout, but that timeline was moving and I would not state it as settled.
The bookmark that started me on this is Sundar Pichai quote-posting Huang. Pichai backs the letter on behalf of Google, stresses Google's open-source history, and points to Gemma, DeepMind's family of open-weight models, as proof. Fair enough, Gemma is real and Google does ship it. But Google did not sign either. Cheering from the sidelines while keeping your name off the page is its own position. Nobody in this is neutral, and nobody is pretending otherwise once you read the incentive.
The CUDA joke and where the piracy claim actually stands
Julian Schrittwieser, the researcher behind AlphaZero and MuZero work, caught the tension with one line. He said he was excited that Huang is a believer in open source now, and looked forward to the CUDA and GPU driver open source release. That lands because CUDA is proprietary and is Nvidia's real moat. Open for Nvidia means open where it sells more chips and closed where it owns the platform. Every open model that ships anywhere drives demand for the hardware to run it. Huang is talking his book, and it is a very good book.
David Sacks, the administration's AI point man, quote-tweeted that jab to reframe it: nobody is demanding all software be open source, only that open-weight AI be allowed, and he cast Anthropic as working to kneecap the open ecosystem. This sits inside a running feud. After Anthropic policy chief Jack Clark published an essay on technological optimism and appropriate fear, Sacks accused the company of running a regulatory-capture strategy built on fear-mongering, and later said the real issue was Anthropic backdooring AI regulation through blue states. Dario Amodei pushed back, saying the company and the administration want the same thing.
Then came the sharper reply in the thread, that Anthropic's models were trained on illegally torrented material, grab everything from GitHub, pirate every book, then train on it. Here I want to be exact, because accusations aimed at a named company should be held to the record.
- The book piracy part is well supported. In Bartz v. Anthropic, Judge William Alsup ruled in June 2025 that training on books was fair use, but that downloading pirated copies from sites like LibGen, PiLiMi, and Books3 to build a permanent digital library was not. Anthropic is believed to have copied up to seven million books that way.
- The company settled for about 1.5 billion dollars, roughly 3,000 dollars per work across some 500,000 titles, the largest copyright class-action payout in U.S. history, and agreed to destroy the pirated files. Because it settled, Alsup's fair-use ruling never became binding precedent.
- The GitHub and PhD papers part is not established by that case. Bartz was about books from three pirate sources. Presenting code and academic-paper piracy as adjudicated fact is overstating the record.
- The nuance that matters: training on the books was ruled legal. The liability was for pirating and hoarding the copies. It was the acquisition, not the training, that crossed the line.
So the tweet is true where it stays on books and inflated where it jumps to everything else. That distinction is exactly the kind that gets flattened when a claim travels fast, and flattening it helps whichever side is shouting.
A gate that decides actual lives
While a room full of American executives argued about markets and moats in front of millions of viewers, a different gate failed on the other side of the world, one that decides real futures rather than market share. On July 25, 2026, India's Education Minister Dharmendra Pradhan resigned over the NEET-UG paper leak, the exam that gates entry to medical college for more than 20 lakh students.
The letter is careful. He addresses students directly, cites 40-plus years in education, thanks the Prime Minister, and walks the timeline: irregularities surfaced in the May 3 exam, the government handed the probe to the CBI, cancelled the exam, ran a re-exam on June 21, moved the test to computer-based mode from next year, and declared results on July 16. He says he took responsibility from day one and would not let the exam mafia harm any meritorious student.
Then the tell. "The events of the past 10 days have saddened me. This is not a matter of personal prestige." He writes that he will not let the country's youth get trapped in a web of confusion, points to the situation at Jantar Mantar, warns that anti-national forces should not benefit, and says he resigned so no student's future gets tangled in legal complications. Read the incentive in the words. The resignation is framed as protecting students from confusion and guarding national unity, not as an admission that the country's most important exam leaked again.
Again is the key word. NEET leaked in 2024 too, with a Patna paper leak, a grace-marks row, and a Supreme Court case, and Pradhan survived that one while resisting resignation calls. In 2026 it leaked despite the 2024 anti-cheating law and an NTA revamp, and this time he is out. What forced it was not the letter's noble self-account. It was pressure from the street: the satirical Cockroach Janta Party youth movement, student unions, and Sonam Wangchuk's 26-day hunger strike at Jantar Mantar, which he ended on July 24 after government assurances. The same week, the Cabinet cleared a tougher anti-paper-leak bill with up to 10 years in prison, fines up to 10 crore rupees, and fast-track courts meant to finish cases within three months.
The material stakes here dwarf the AI drama. More than 20 lakh students sat an exam, had it cancelled, sat it again, and waited on delayed counselling, with reports of real distress running through it all. This is a gate that decides whether a poor kid from a small town becomes a doctor. It got a fraction of the Anglophone volume that the moat fight commanded, which tells you something about whose problems get narrated as urgent.
Both letters are pleas dressed as principle. The AI one turns market self-interest into the language of sovereignty and safety, and it is not wrong about openness even while it talks its book. The resignation one turns a system's second failure into a story about shielding the young from confusion. Neither is a lie exactly. Each is the polished side of a claim whose rougher version, follow the money in one and count the leaks in the other, is closer to the truth. When two documents land in the same week asking you to trust the framing, the useful habit is to ask who guards the gate and what they gain by keeping the story simple.