NVIDIA CEO Jensen Huang on Open-Weight AI Models and Ecosystems

NVIDIA just wrote an AI Love Letter, about Open Weights: 💌 Except it's not really about open source. It's about who owns the toll booth underneath it. Here's the plot twist buried in "Open Weights and American AI Leadership": ▸ The letter launched with 25 signatures on July 24 ▸ It doubled to 50 within a day ▸ OpenAI wasn't on the original list — it signed AFTER getting called out publicly 👀 ▸ Anthropic and Amazon still haven't signed The company selling closed models... didn't want to sign a letter defending open ones. Shocking, I know. So why is a chip company leading the charge for open weight AI? One word: CUDA. 🔌 Twenty years of proprietary software glue that every model, open, closed, American, Chinese , runs on top of. NVIDIA doesn't care who wins the model war. Whoever wins is still buying the chips. And the long game is even better: NVIDIA's Chinese market share went from 95% to basically zero. A world where "open weights" (including Chinese ones) becomes untouchable policy is a world where the door to China reopens someday, without NVIDIA ever having to say the word "China" once. Which it doesn't, anywhere in the letter. 🎯 Oh, and here's the part that made me laugh out loud: The letter also asks Washington to protect frontier labs' IP from "unauthorized distillation." Reminder: this letter got written days after Anthropic accused a Chinese lab of distilling its own Fable 5 model. 🍿 So the same industry that trained on the entire internet without asking anyone's permission is now very concerned about IP theft, just, you know, in the direction that costs them money. The real question isn't "who's right." It's: does your AI strategy survive no matter who wins? CUDA is the real moat, the China long game, and what it means for how you should actually be building your model strategy, in this week's Substack. #AIStrategy #OpenWeights #NVIDIA #BusinessFirstAI #AIGovernance

Demis Hassabis’ most recent post (his first Substack!!) is one of the more important AI governance signals I’ve seen in a while, not because it says AGI is near, but because it reframes the issue as infrastructure. For years, AI governance has been discussed as principles, policies, councils, and controls bodies. All of it is necessary, but not sufficient. As models become more agentic, more autonomous, and more embedded into real workflows, governance has to become part of the architecture: testing protocols, independent evaluation, model release standards, cyber controls, auditability, incident response, and recovery. (My preference for speed of innovation is cringing, but I know this is right.) He suggests there should be a framework for a Frontier AI standards body with FINRA as an analogy. Financial markets learned long ago that trust at scale does not come from innovation alone. It comes from standards, oversight, shared operating discipline, and infrastructure resilient enough to absorb shocks. The same may now be true for frontier AI. This is certainly not a call to slow innovation, but it’s a call to make innovation as durable as possible. https://lnkd.in/gXX2mzzU #AI #AIGovernance #FrontierAI #ResponsibleAI #FinancialMarkets

At a first glance, yesterday's announcement of NVIDIA's $6B bet to develop open weight American foundational models is a pure "commoditize layer above me" play. It allows NVIDIA to protect its chip ecosystem by hedging against both Western frontier labs and their significantly cheaper Chinese counterparts and make compute the only scarce input. This move follows an earlier NVIDIA $5B bet on Safe Superintelligence and $800M bet in Reflection AI, another US startup building open weight models for predominantly Western customers. On a deeper level, a $12B bet is significantly more than any combination of worldwide government investments to build an open weight model ecosystem. It represents a state-like policy setting investment that if successful diminishes risk of frontier model distillation (i.e. intellectual theft), establishes future sustainable cost effective US alternative for small/medium size global businesses and forces Chinese models to either maintain velocity and efficiency of current deployment or attempt to follow American models by monetizing current models and moving away from pure open weight licensing. Of course, they can also try (and are trying) to beat NVIDIA at their own game by commoditizing the compute itself. For more opinions on open weight models: - Anthropic - Our position on open-weights models (https://lnkd.in/eH9R5pfm) - Stanford Institute for Human-Centered Artificial Intelligence (HAI) - Open weight models aren't enough (https://lnkd.in/efq-DsYG) - Layer 3 Labs - American Open Weight Models and who ships them (https://lnkd.in/egkQtHVX) - Jin et al - U.S. Technological Containment and the Rise of China’s Open AI Ecosystem (https://lnkd.in/eVnEwiG6) https://lnkd.in/eW8MuR2X

On Wednesday, Anthropic said it would use all of the compute capacity at Colossus 1, the SpaceX-owned supercomputer in Memphis, and within hours it raised usage limits for paid Claude users. Doubled rate limits and the removal of peak-hour throttling went live for Claude Pro, Max, and Claude Code subscribers. That same day, the NAACP and Earthjustice asked a federal court for emergency relief against xAI's Southaven power plant. The filing says the plant is powered by 33 unpermitted gas turbines in violation of the Clean Air Act, and likely makes the site the largest industrial source of nitrogen oxides in the greater Memphis area. Southaven is a majority-Black community in northern Mississippi already failing federal smog standards. The Southaven plant powers Colossus 2, a separate facility from the one Anthropic just leased. xAI used as many as 35 unpermitted turbines at Colossus 1 in 2024 before some were permitted and others removed. Three months before this week's compute deal, Anthropic published a public commitment to cover 100% of consumer electricity price increases caused by its data centers. The pledge applies directly where Anthropic develops data centers for its own workloads. Where it leases capacity, the company says it is "exploring further ways" to address those effects. So the picture this week is a company that is also suing the current administration after a Pentagon dispute over Claude's military use restrictions, while raising paid-tier capacity by leasing all available compute from an operator whose other Memphis-area facility is now the subject of a federal Clean Air Act enforcement action. My read is that ethics commitments hold inside the perimeter a company controls, and get blurrier the further out you go. We saw the same pattern in apparel, in conflict minerals, in cloud carbon accounting. Frontier AI is now running into the same issues. Every enterprise buying frontier AI inherits a version of this gap. Your AI vendor's carbon disclosures, community commitments, and ethics posture usually describe the facilities they own. The inference serving your traffic is probably running somewhere else, on terms your vendor did not write. A SOC 2 report and a model card are necessary diligence. They are not sufficient diligence on where workloads actually run, who absorbs the externalities, or which operators your vendor's principles do or do not extend to. When you ask your AI vendor to demonstrate their values, do you ask about the data centers they own, or the data centers actually serving your traffic? #AIGovernance #EnterpriseAI #AIStrategy #ResponsibleAI #DataCenters

Nvidia and two dozen companies signing an open-weights letter while Washington debates Chinese AI is a signal that “openness” is becoming geopolitics. Open weights aren’t just a research preference - they’re a supply chain. Whoever controls the distribution of capable models controls where innovation happens, which companies can build, and which nations can compete. The notable part: OpenAI, Anthropic, and Google not signing. That split matters. We’re watching the industry fork into two camps: “trust us with closed models” vs “trust the ecosystem with inspectable weights.” For enterprises, this is not philosophy - it’s risk management. Open weights can be audited, fine-tuned locally, and deployed with data residency. Closed models can be safer by default at the margin - until you’re locked into one vendor’s policy changes, pricing, and outages. Bold view: closed model lobbying dressed up as “safety” is just regulatory capture with a nicer label. If we let a few gatekeepers decide what’s “safe,” we’ll end up with nationalized innovation by proxy. If China’s progress is the concern, do we really win by concentrating AI power into fewer Western companies - or by distributing capability with hard, enforceable standards? #AI #OpenSource #Geopolitics #OperatorOS Playing it forward (get my CEO Playbooks): https://lnkd.in/gr_x7hgi https://lnkd.in/gt3SN5xu