On July 24, twenty-five technology companies published a three-page letter asking Washington not to place premature restrictions on downloadable AI models. Nvidia’s Jensen Huang, who had never posted on X before, used his first post to promote it. By the next day the list had doubled, and it has kept growing since — Microsoft, Meta, IBM, Dell, Palantir, Hugging Face, Y Combinator, Mistral, Andreessen Horowitz.
Three notable companies weren’t on it: Anthropic, Amazon, and Google. Google and OpenAI added their names later, after the omission drew attention. Anthropic still hasn’t.
I should say where I sit before going further. AICV runs on Anthropic’s models. That makes this uncomfortable to write, and worth writing anyway.
the detail nobody led with
Tom’s Hardware caught the thing that explains the entire roster: none of the original twenty-five signatories sells access to a closed frontier model.
That’s not a coalition organized around a principle. It’s a coalition organized around a position. Chipmakers, cloud operators, enterprise software, venture funds — every one of them makes money when AI is cheap and everywhere. The companies selling premium access to models you can’t download are the ones who didn’t sign, or signed late.
Which doesn’t make the argument wrong. It makes it interested, and those are different things. The letter’s actual case holds up: openness removes a single point of failure, defenders need capabilities comparable to attackers, and distillation is a normal technique rather than theft. Anthropic’s position also holds up on its own terms — Dario Amodei wrote on Monday that the company has never advocated banning open-weight models and called the framing a red herring.
Two reasonable arguments, both made by parties with money on the outcome. That’s most of what a policy fight is.
why this reaches the desert
Open weights are the reason cheap AI exists.
Not the discount tier at a big lab — the actual floor. Downloadable models are what let a small shop run something for pennies, and they’re what forces the large labs to price against an alternative. The New Stack put it plainly: running an open Chinese model isn’t a fringe act, it’s how a lot of small builders keep their bills survivable while the major labs get more expensive.
A couple of weeks ago I wrote that the AI price collapse is a gift to small operators — that a restaurant or a clinic can now run agents that used to need an enterprise contract. I still think that. This is the argument about whether it stays true.
If downloadable models get restricted, the cheap end of this market narrows to whatever a handful of closed labs decide to charge. It wouldn’t happen loudly. The tools you use today would keep working. The floor would just quietly stop falling.
what a local operator does about it
Directly, nothing. This gets settled in Washington by people who have never heard of the Coachella Valley.
But there’s a smaller move worth making. If you adopted AI in the last year, some part of your operation now assumes it’s cheap. Find out what you’d do if that changed — which task you’d stop automating, what it would cost you to go back to doing it by hand. Not because a price hike is coming, but because you should know your own exposure before someone else decides it for you.
Every argument in AI right now arrives dressed as safety or sovereignty. Most of them are also arguments about price. The question isn’t which side you’re on. It’s whether you noticed your operating costs are being negotiated in a room you’re not in.
The full letter is on microsoft’s site.
source: open weights and american ai leadership (july 24, 2026); tom’s hardware; cnbc; the new stack.