Open weights won the argument in 48 hours

On Friday 24 July, a three-page policy letter titled Open Weights and American AI Leadership went out under 25 signatories, among them Nvidia, Microsoft, Meta, IBM, Dell, ServiceNow, CrowdStrike, Hugging Face and Palantir. Jensen Huang shared it as his first post on X. The absences drew as much attention as the names: OpenAI, Anthropic and Google, the three companies selling access to closed frontier models, had not signed.

That lasted about a day. OpenAI added its name on Friday afternoon US time, after the gap had been widely noted and Sam Altman had already posted that he wants the US to win in both open source and proprietary models. Google followed over the weekend, and by Monday morning the letter's own signature page carried 77 names, SpaceX and Y Combinator among them. Among the frontier labs, Anthropic is the remaining holdout.

The letter asks Washington to avoid “premature restrictions on open models”. It also asks for expanded compute access for startups and researchers, public investment in shared datasets and evaluation frameworks, and for distillation, training one model on another's outputs, to be handled through targeted legal action rather than broad limits on the technique. It arrived four days after reports that the US administration was reviving a push to ban Chinese AI models, and it never mentions China.

For the first day the list read like a map of commercial interests, since nobody on it sold closed frontier access. OpenAI and Google signing complicated that. Both earn revenue from closed models, and both also publish open-weight models of their own, Google's Gemma and OpenAI's gpt-oss. I won't pretend to know what mix of conviction and politics put their names on the page. Nobody wants to be on the wrong side of a Washington fight over model restrictions, and the letter's defence of distillation protects practices the whole industry leans on. The signatures settle a narrower question. Whichever longer game each company is playing, all 77 are positioning for a market with open weights in it, and Jensen Huang said at CES in January that one in four tokens generated today already comes from an open model.

Palantir signing on day one fits its product. It sells the platform around the model rather than the model itself, so AIP works with closed APIs or open weights, whichever a customer is permitted to run. More usable models in more places is good for that business.

The letter is addressed to policymakers, but most of the arguments in it are ones procurement teams already make: not being locked into one provider, keeping control of your own data, and the freedom to test models properly and run them where you need to. All of it sits outside the model. Whatever you buy, open or closed, it arrives knowing nothing about your data or what an agent is allowed to do in your systems. Rahul covered this in Part 3 of our What is Palantir series, the model is the brain you rent or own, the ontology is the body you build. The data integration, the permissions, the evaluation work, the ontology itself: that work carries over when you swap models, and it matters more as the market keeps producing models worth swapping to.

So before signing an AI contract this year, I would want answers on model governance more than model choice:

1. Which open and closed models are we permitted to run, and who decides?
2. Are there residency or on-premises requirements?
3. How does the system get evaluated and red-teamed, independently of the model vendor?
4. Can work be routed across models on cost and performance?
5. What does replacing a model cost once workflows depend on it?

All five are already on our own discovery checklist and the letter gives buyers a good reason to start asking them too.