50 Firms Signed the Open-Weights Letter. Anthropic Didn't

What you need to know

• The document. An open letter titled "Open Weights and American AI Leadership", posted on X by Nvidia's Jensen Huang on 24 July 2026. Per Forbes, the post reached roughly 11 million views, and signatures kept arriving while the coverage was still being written.
• The growth. It launched with 25 technology company signatories and had doubled to 50 by 25 July 2026 — around two dozen names added inside a day.
• The ask. That Washington not restrict downloadable AI models, and that policymakers avoid measures which could "stifle competition" or "drive innovation overseas". The letter argues openness "prevents lock-in" and avoids "single points of failure".
• The trigger. A brewing policy fight over whether to restrict Chinese open-weight models.
• The absences. As of late July 2026, Amazon and Anthropic have not signed. Forbes called that pairing "the most interesting thing missing from the list", noting Amazon is Anthropic's largest investor and a major infrastructure partner. Neither has publicly explained the decision.
• The legal status. None. This is advocacy. Nothing has changed about what you can download this morning.

Why this is a cost question before it is a policy question

Strip away the framing about leadership and competitiveness and the letter is about one mechanical thing: whether a model's parameters can be downloaded and run on hardware you control. That single property is the hinge on which most cost-constrained AI architecture turns.

If you can download weights, you have an exit. When API pricing moves against you, when a provider deprecates the model your evaluations were tuned against, when your data-residency requirements harden, you can move the workload onto your own GPUs and eat a known infrastructure bill instead of an unknown per-token one. If you cannot download weights, that exit closes and renting becomes the only option. Everything after that is a negotiation you have no leverage in.

Which teams feel that first is not evenly distributed. A well-funded San Francisco startup burning venture money on API credits treats inference cost as a line item to optimise later. A twelve-person team in Pune or Leeds serving a domestic market at domestic price points does not have "later" — the arithmetic either works at launch or the product does not ship. We have run those numbers in detail before, in the break-even maths for self-hosting DeepSeek V4-Pro on eight H100s, and the conclusion is consistent: above a certain sustained token volume, owning the inference wins, and below it renting wins. Restricting downloadable weights would not change that curve. It would delete half of it.

The signatory list, read as a map of interests

The most useful way to read the list is not as a headcount but as a cross-section of who benefits when model weights move freely. What follows is an interpretation of business incentives, not a claim about anyone's stated motives — every company here signed a document about competitiveness and none of them published a rationale keyed to their revenue model.

Read that way, the pattern is hard to miss. The infrastructure and distribution layer — the companies that sell the picks, shovels, racks, registries and runtimes — is almost fully represented. These are businesses whose revenue scales with the number of places a model can run. The layer with the least obvious incentive is the one selling metered access to a model that only exists behind their endpoint: for that business, a freely downloadable competitor of similar quality is a substitute good, not a complement.

That is an interpretation, and it is worth stating its limits. OpenAI and Google both joined after publication despite substantial API businesses, which cuts directly against a tidy incentives story. Palantir and Block do not slot neatly into any of these boxes. And Meta and Mistral have released open weights for years, so their position is a continuation rather than a calculation. The map is useful; it is not a proof.

The two names that are not on it

Forbes flagged Amazon and Anthropic as the most interesting omission, and the reason it is interesting is structural: Amazon is Anthropic's largest investor and a major infrastructure partner, so the two absences are not independent observations.

Beyond that, there is very little to say responsibly. Neither company has publicly explained why it did not sign. A letter of this kind circulates fast — Huang posted it, it reached roughly 11 million views per Forbes, and signatures arrived within hours — and there are a great many ordinary reasons a large company is not on a one-day-old list, from legal review timelines to a simple disagreement about wording. Absence from a voluntary advocacy document is weak evidence of anything.

The awkward part: the strongest open model right now is Chinese

The letter's timing creates a genuine tension rather than a rhetorical one. It was prompted by a policy fight over whether to restrict Chinese open-weight models, and it lands in a week when the most capable open-weight model available is Chinese.

Moonshot AI's Kimi K3 is a 2.8-trillion-parameter open-weight mixture-of-experts model with native vision and a one-million-token context window, posting 93.5% on GPQA Diamond — described in reporting as the best open score published to date. Its full weights were released by 27 July 2026. We covered what K3 actually does and where it tops the leaderboards when it landed.

The American answer is not absent, and it is not weak. Thinking Machines Lab released Inkling on 15 July 2026 under Apache 2.0 — 975 billion total parameters with 41 billion active, with native text, image and audio reasoning — which we wrote up as America's largest open-weight model shipping under a genuinely permissive licence. So the picture is not one-sided. But two weeks apart, one of these is a 975B model and the other is a 2.8T model with the best published open score on a hard reasoning benchmark, and the argument for openness is being made by an American coalition in exactly that window.

The honest reading is that "keep weights downloadable" and "restrict Chinese weights specifically" are not opposites but they are difficult to hold together. A rule that permits downloads in general while blocking a particular country's models has to be enforced at the download, which means someone has to police model distribution — and the infrastructure for policing distribution is the infrastructure the letter is arguing against building. Meanwhile the practical effect of a country-specific restriction on a Bengaluru or Bristol team is not neutral: whatever else you think of it, K3 currently sets the ceiling on what free-to-run capability looks like, and a rule that removes it from your options removes it from your budget too. That is the tension, and no version of this fight resolves it cleanly.

What actually changes for a builder: nothing, yet

To be blunt about it: as of 29 July 2026, this letter has changed precisely nothing about your stack. No model has become unavailable. No licence has changed. No agency has acted. If you were planning to pull weights this afternoon, pull them.

What the episode does do is price a risk that was previously invisible. The right response to a risk you cannot control is not to predict it — it is to make yourself cheap to move. Four positions are worth taking now, and all four are worth taking even if the policy fight fizzles entirely, because each of them also protects you against ordinary commercial events like a price rise or a model deprecation.

• Keep an open-weight fallback path warm. Not necessarily in production — but you should know which open model your core prompts run acceptably on, and have the evaluation numbers to prove it. A fallback you have never measured is a hope, not a plan.
• Do not build your core loop on provider-specific features. Proprietary tool-calling formats, bespoke caching semantics and provider-only structured-output modes are all fine at the edges and expensive in the middle. Keep the part of your system that decides things portable.
• Know your self-host break-even number. Not a vibe — an actual figure in tokens per month, with your traffic shape and your GPU rental rate. Our vLLM production playbook walks through the throughput maths that number depends on.
• Archive the weights you actually depend on. If an open model is load-bearing for you, having a copy in your own object storage costs very little and removes a dependency on someone else's registry policy. Check the licence permits redistribution or retention for your use before you do it.

India and the UK are exposed to the same rule, differently

Neither of our markets gets a vote on what Washington decides, which is precisely why the exposure is worth understanding.

India. The economic logic of the IndiaAI Mission's compute programme is subsidised GPU-hours — and subsidised GPU-hours are only useful if there is something capable to run on them. We covered the scale-up when IndiaAI compute passed 34,000 GPUs, and the entire proposition for a startup taking that capacity is: rent cheap hardware, run open weights, avoid paying frontier API rates in dollars. Remove downloadable weights from that equation and the subsidy buys you idle silicon. This is the clearest case anywhere of a national compute policy resting on a distribution policy set in another jurisdiction.

The UK. British teams face a separate domestic AI regulatory track, which will govern how they deploy systems regardless of what any US letter says. But domestic regulation governs deployment; it does not manufacture supply. The pool of open models a London or Manchester team can build on is determined overwhelmingly by release decisions made in the US, China and France. A UK team can be fully compliant at home and still find its cost base rewritten by a rule it had no part in.

The common thread is dependence without representation. That is not an argument for panic — it is an argument for the portability discipline above, which is cheap insurance in both markets and useful even in the world where nothing happens.

The honest read

Fifty signatures in a day is a real signal about industry alignment, and a weak signal about outcomes. Open letters are the cheapest artefact in policy: they cost a legal review and a logo, they commit nobody to anything, and they are written to be quoted rather than to be implemented. This one is unusually well-supported across the infrastructure layer, which tells you the picks-and-shovels half of the industry has converged on a position. It does not tell you what any government will do, and this article is not going to pretend otherwise.

What is worth carrying forward is smaller and more durable than the news cycle. Downloadable weights are the mechanism that keeps AI costs negotiable for teams outside the funding centres. That mechanism is currently a norm rather than a guarantee, it is being argued about in public for the first time, and the argument is happening in a week when the strongest example of it is Chinese and the strongest American answer is two weeks old. Builders who treat portability as a standing engineering requirement — rather than a reaction to whatever the coalition-of-fifty becomes next month — will be fine either way, and we will keep tracking the fight itself in our AI policy coverage.