← Back to feed

Microsoft published “Open Weights and American AI Leadership” on July 25 — a position paper and signatories list arguing that open-weight AI models strengthen competition, improve security, and are necessary for sustained US technological leadership. The document went live with more than two dozen signatories and grew to 35 in under a day.

The coalition is notable for who signed and when. NVIDIA CEO Jensen Huang used the statement as the occasion for his first-ever post on X. OpenAI, which was not listed at launch, added its signature on the same day — a reversal of its typical posture on open-weights policy.

Who Signed

Named signatories at launch include Meta, NVIDIA, Palantir, and a range of mid-tier AI and infrastructure companies. Intel and AMD both signed. The full list grew to 35 by July 25, spanning cloud providers, chipmakers, enterprise AI vendors, and foundation model labs.

Notably absent: Anthropic and Google DeepMind. Both companies have open-source commitments (Gemma from Google, various research releases from Anthropic) but neither signed the declaration. The statement targets Washington and explicitly frames open weights as an American competitiveness issue, which may complicate participation from labs whose safety positioning relies on arguments for controlled model access.

What the Declaration Argues

The document makes four claims:

1. Open weights expand access — developers and researchers outside hyperscaler budgets can build on open models in ways not possible with API-only access
2. Open weights strengthen competition — closed-model monocultures consolidate AI capability with a small number of firms; open distribution spreads it
3. Open weights improve security — adversarial testing and vulnerability discovery happen faster when researchers can inspect weights directly rather than probing through an API
4. Open weights sustain American AI leadership — China’s open-source download share (1.15B downloads) already exceeds the US (723M) according to the ATOM Report; restricting American open-weight models would cede ground that has already been gained

The competitiveness framing responds directly to ongoing White House discussions about whether open-weight models above a certain capability threshold should be treated as export-controlled technology.

The Political Context

The open-weights question has been live in Washington since late 2025, when Kimi K3 and GLM-5.2 demonstrated that Chinese open-weight models could match or exceed US closed-frontier performance on specific benchmarks. The White House subsequently explored whether export controls should apply to open-weight releases from US labs.

The Kill Switch Act introduced the same week (requiring shutdown mechanisms on frontier AI) specifically carves out open weights — the bill acknowledges that downloaded weights cannot be switched off and focuses exclusively on hosted commercial deployments. The open-weights declaration and the Kill Switch Act thus represent parallel legislative and industry responses to the same capability concern, pointing in opposite directions.

Jensen Huang’s First Post

Jensen Huang has been an active voice in AI hardware and policy discussions without maintaining a personal social media presence. His first post on X referenced the coalition directly. The significance is less about the content and more about the signal: the CEO of NVIDIA — which supplies the compute infrastructure for every major AI lab — has chosen to make open-weights access his first public political statement. That positioning is unlikely to be accidental.

OpenAI’s Late Addition

OpenAI’s signature arriving after the initial launch is the most structurally interesting element of the coalition. OpenAI has historically occupied a middle position on open weights: it has released some models (GPT-2, Whisper, CLIP) while keeping its frontier models closed. The company’s decision to join the declaration on the same day it was published — but not at launch — suggests the statement evolved faster than expected and OpenAI chose to be included rather than visibly absent.

Whether this changes OpenAI’s actual release practices is a separate question. The declaration is a statement of principle, not a commitment to open-source specific models.