A cross-industry statement titled Open Weights and American AI Leadership now lists 35 signatories on its live Microsoft-hosted page, and OpenAI is among them - alongside NVIDIA, Meta, IBM, Hugging Face, Palantir, GitHub, Cisco, Cohere, Nous Research and Prime Intellect. That contradicts the version of the story that travelled furthest, in which OpenAI was the closed-lab holdout. The document names no bill, executive order, or rule it opposes.
Key facts
- The anchor number: the roster grew from 25 organisations on the original NVIDIA-hosted PDF to 35 on the live page, with no published revision history.
- When: circulating July 24, 2026.
- Who: hosted by Microsoft; signatories span chipmakers, clouds, model labs, VCs, security vendors and open-source foundations.
- Primary source: the live statement and signatory list, with the original NVIDIA PDF as the earlier document state.
The letter defines open weights plainly: models people can download, inspect, modify, and run themselves. Its argument runs on four rails. Local, adaptable models cut vendor lock-in and make AI economical for specialised work. Openness widens competition across models, chips, clouds, applications and services. Open weights are a cyber-defence capability, because defenders can inspect and test models locally - a point that acquired uncomfortable support this month when Hugging Face had to use a self-hosted open model for incident forensics after hosted frontier models refused to analyse the malicious artefacts. And distillation is a normal engineering practice for improvement, evaluation and validation, so alleged theft should be met with targeted legal remedies rather than a broad ban on the technique.
That last plank is the one with live political stakes. Distillation - training a smaller model on a larger one's outputs - has become the pressure point in US-China AI policy, with Washington having already alleged that Moonshot distilled Anthropic's Fable and the Treasury Secretary floating sanctions over it. The coalition is asking for the technique to be protected as legitimate while carving out actual misappropriation, which is a narrower and more defensible ask than the headlines suggest.
But the letter is not fighting a named enemy. It cites no pending bill, no executive order, no agency rule, no release threshold. And the closest current federal text is not a ban: the White House AI Action Plan already says the federal government should create a supportive environment for open models and facilitate compute access, and the June executive order calls for a voluntary frontier-model framework while explicitly disclaiming authority for mandatory licensing or pre-clearance of model releases. This coalition is pressing an already sympathetic administration to hold its line - not rebutting an identified threat.
The signature confusion is worth walking through, because it shows how a document can change under a story. The original NVIDIA-hosted PDF carried 25 organisations, including Andreessen Horowitz, Black Forest Labs, CrowdStrike, Dell, Hugging Face, IBM, the Linux Foundation, Meta, Microsoft, Mistral, Mozilla, NVIDIA, Palantir, Perplexity, Replit, ServiceNow and Y Combinator. The live Microsoft page adds ten more - Cisco, Cohere, DoorDash, Fireworks AI, GitHub, Nous Research, OpenAI, OpenClaw, Palo Alto Networks and Prime Intellect - without any revision note. Anyone reading the earlier PDF and noting OpenAI's absence had a defensible observation about a dated document state, and an indefensible conclusion about a refusal.
OpenAI's own stated position never fit the holdout framing anyway. Its open-weights policy post calls open versus closed a "false choice" and says both are needed. Its gpt-oss safety documentation supplies the qualification the coalition glosses: release is genuinely different because determined users can fine-tune around refusals and the developer cannot revoke access. OpenAI says it shipped gpt-oss only after adversarial fine-tuning tests failed to cross its high-capability thresholds.
The strongest counter-argument is not that closed models are safer. It is that public weight release is an irreversible capability-distribution decision, so it should be risk-tiered rather than treated as a default good. The International AI Safety Report 2026 finds both sides: released weights widen research, local deployment and scrutiny, but they also make safety training easier to strip out, make monitoring harder, and cannot be recalled once copied. An Oxford Martin AI Governance Initiative submission sharpens it into policy: preserve research and audit access through structured access or APIs, but restrict full-weight release for systems that materially enable catastrophic misuse. The letter itself concedes the core risk in one line - released weights cannot reliably be recalled or traced after modification.
The honest caveat is about whose interests are aligned here. A coalition containing GPU vendors, cloud providers, enterprise integrators, model hosts, venture funds and open-model labs benefits from open weights for several quite different commercial reasons. That is a broad commercial alignment, not a neutral technical consensus, and it is the fairest lens for reading a document that asks for compute access, shared datasets, and freedom from restrictions nobody has yet proposed.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)