Open letters: 235 signatures for open weights in the US

On 24 July, Open Weights and American AI Leadership was published, an open letter shepherded by Microsoft and signed by 235 companies in the sector: NVIDIA, Amazon, Y Combinator, The Linux Foundation and, added later, OpenAI. It is not a technical document or a product roadmap. It is a piece of political pressure with a specific goal: to keep the US government from limiting or banning open weight models on safety grounds.

Simon Willison has gathered that letter and the rest of the recent run of open letters in a post on his blog, originally written as a section of his sponsors-only newsletter. His reading is blunt: the text is built to counter any regulatory instinct in that direction, and the concern is not hypothetical. As a reminder, Willison links to what happened with Claude Fable 5 in mid June, when a government directive affected access to the model.

What the letter argues

The line Willison highlights is the one that best sums up the strategy: "Relying solely on closed models is not inherently safe". The argument flips the usual framing of the debate. For the past few years, the discussion around open weights has almost always been posed as a risk problem: if anyone can download the weights, anyone can strip out the filters. The letter answers from the other side: concentrating all capability in a handful of closed providers is also a risk, just a different kind (dependency, opacity and a single operational point of failure).

The fact that 235 companies signed it, and not only the open source community, is what matters. NVIDIA sells hardware that pays off better the more people train and serve their own models. Amazon hosts a good share of that infrastructure. Y Combinator funds startups that lean on open models so they do not depend on someone else's API. OpenAI's late signature, from a company that until recently was the textbook closed model case, is the most commented detail.

Why now

The context comes from Willison himself, who links to an Axios piece from 20 July about the competition between the United States and China in open weight models, with Kimi as the reference on the Chinese side. That is the nerve of the matter: if Washington restricts open weights inside its borders, the global ecosystem does not disappear, it simply leans on the models published elsewhere.

The Claude Fable 5 episode in June worked as a practical warning. A model served exclusively by its provider can become inaccessible through a decision that neither you nor the provider made. Anyone running production on top of that API felt it within hours.

Who this is useful for

If you build products on LLMs, this is not abstract policy. It touches three very concrete decisions:

1. Contingency plan: having an escape route to an open weight model, even one that performs worse, stops being a theoretical exercise and becomes part of the design.
2. Integration architecture: layers like MCP help because they decouple the tools from the model calling them. Switching providers costs less if the logic lives in MCP servers rather than embedded in prompts tied to one specific API.
3. Compliance and data: for clients with residency or audit requirements, a model you can deploy yourself is still the simplest answer to defend.

For teams working mostly with Claude, the useful takeaway is not "drop the closed model". It is: do not design as if access were guaranteed forever.

Our take

A letter with 235 signatures moves little on its own, but it maps out clearly where the industry's interests sit when safety comes up. At ElephantPink we keep building on Claude because it is what works best in our projects today, while keeping our integrations decoupled enough that we do not depend on that decision still holding next year.