Jensen Huang's first X post ever was about policy, not chips

Jensen Huang has run Nvidia since 1993 and had never posted on X. On July 24 he finally did, and he did not use it to sell a chip.

@jensenhuang shared a coalition letter titled "Open Weights and American AI Leadership" and summed up his position in a line: "The world needs both frontier closed models and frontier open models." Forbes put the post at around 11 million views. The letter launched with 25 signatures. Inside of a day it had 50, and by July 27 coverage had it at 77, with OpenAI and Google both joining after launch.

What the letter is asking for

Open weights means a model you can download and run on hardware you control, rather than one you rent through somebody's API. Meta's Llama models, Mistral, and Moonshot's Kimi are the names most people have heard.

Washington is weighing restrictions on Chinese open-weight models, and there is a live fight over whether distillation, training a model on another model's outputs, counts as normal research or as theft. The letter asks policymakers not to restrict downloadable models as a category. It separates distillation as a legitimate technique from outright misappropriation of a closed model's value, and asks for rules aimed at the second instead of bans aimed at the first. Its security argument is the one worth reading twice: if defenders cannot run models comparable to what attackers are running, a closed-only ecosystem becomes one very large single point of failure.

The company that didn't sign

Anthropic. That absence is what the AI corner of X actually spent the week on. Amazon, Anthropic's largest investor, was also missing at the 50-signature mark.

White House AI adviser @DavidSacks went at the omission publicly, writing that nobody is demanding all software be open source, only that "open-weight AI should be allowed," and finishing with "Read the room."

Anthropic answered on July 27. Dario Amodei published a statement saying the company "has never advocated for a ban on open-weights models," and laid out a counter-proposal: keep advanced chips out of China with real enforcement against smuggling, act against industrial-scale distillation, and require pre-release safety testing for any sufficiently capable model regardless of how it is licensed. Critics read that third item as a restriction wearing a lab coat, since a testing mandate lands hardest on the people shipping weights they can never recall.

The week got charged enough that @karpathy spent part of July 27 denying he had quit Anthropic, after people noticed his X bio had changed and decided a famous open-source advocate must have walked over the letter. "weird misinformation to find circling on twitter, no," he replied. A bio edit got read as a resignation letter. That tells you how hot the room was.

My read

Both sides are arguing their book, and both are still right about something.

Huang sells the hardware that open models run on everywhere, so a world of many models is a world of many buyers. Anthropic sells access to closed frontier models, and a safety-testing mandate is a cost its competitors would have to absorb and it already absorbs. Motives noted. The arguments survive them anyway: a concentrated model supply really is fragile, and weights you cannot take back really are hard to un-ship.

What this means if you run a business

Skip the politics and look at the dependency. When you build a process on top of an AI model, you are picking a landlord. In June, Anthropic switched Claude Fable 5 off for every customer on earth under a US export-control order. That was not an outage or a bug. It was policy, executed in about a day.

You do not need to self-host anything to be ready for that. You need to know, for each automated process, which model it calls, what breaks if that model goes dark tomorrow, and how long a swap would take. Most owners we talk to cannot answer the first question, never mind the third.

The prep work is unglamorous. Write down every place AI touches your operations. Keep your prompts and business logic in your own files instead of buried in a vendor's interface. Test one alternate model per process, so switching is a checkbox rather than a project.

If you want that mapped for one workflow, New Face Design runs a free process audit: what the process costs you now, where AI actually earns its place, and what you would do if your provider disappeared on a Tuesday. The labs are arguing about who controls the off switch. It is worth knowing where yours is.