Microsoft spearheaded the initiative and today published the open letter “Open Weights and American AI Leadership.” The document is signed by a substantial group of companies – Nvidia, Meta, Palantir, Hugging Face, and about twenty others – and urges policymakers to resist the temptation of imposing broad or premature restrictions on open weight models.

The request is sharp: draw a clear line between legitimate model distillation and misappropriation. This is no technical footnote; it is the heart of a conflict that is reshaping the economics of AI and, with it, the incentives for those who develop, distribute, and run Large Language Models.

Looking at who signed and who did not already tells a story. Among the signatories we find companies that sell infrastructure (Nvidia with GPUs, Microsoft with Azure and its hybrid presence, Palantir with analytics and government deployment platforms) and organizations that thrive on open ecosystems like Hugging Face. Meta, for its part, has released models like Llama under permissive licenses, building influence and community-driven improvement. For all these players, open weight models are an enabling layer: the more freely weights circulate, the more inference and training workloads multiply, and the greater the demand for compute – on-premises, in the cloud, at the edge. This is not philanthropy; it is the alignment of an entire bloc of industrial interests.

On the other side, the glaring absentees – OpenAI, Anthropic, Google – have built their advantage on closed models accessible only through APIs. For them, regulation offers a double shield: it protects against reputational risks while simultaneously raising barriers for anyone trying to replicate their capabilities with open, self-hosted alternatives. The letter, by rejecting blanket restrictions, strikes directly at that mechanism.

For those evaluating on-premise deployments or environments with data sovereignty constraints, the stakes are enormous. Open weight models are the raw material of self-hosting: they allow organizations to download weights, apply quantization, fine-tune on proprietary data, and build inference pipelines that never leave the corporate perimeter. Any restriction, even if designed for extreme cases, risks slowing the development of efficient variants – the very variants that today let an organization run an LLM on its own hardware, lowering TCO and maintaining control over data residency.

The emphasis on distillation is deliberate. Technically, distilling a large model into a smaller, specialized one is the primary method for bringing advanced capabilities to machines with limited VRAM, in air-gapped or edge scenarios. Defining a legal boundary that separates this practice from intellectual property theft means protecting the technical conditions that make local deployment possible. Without that distinction, every attempt to compress a model for on-premise use could be exposed to litigation, freezing innovation.

This is not a game confined to Silicon Valley. The entire letter is framed around “American leadership,” but the implications are global. For European companies, for instance, the availability of open weight models is often the prerequisite for adopting generative AI while complying with the GDPR without handing over data to third parties. Any unilateral regulation risks cascading effects on market dynamics and on the architectural choices of those who deploy in-house.

In this clash, the real battleground is not the model itself but the control layer over infrastructure. The letter's signatories are telling lawmakers: do not hand the future of AI to those who run proprietary APIs, because the cost in terms of competition, sovereignty, and adaptability would be enormous. Beyond the rhetoric, that stance draws an alliance between those who produce the building blocks of AI and those who put them to work, leaving out those who prefer to sell the house already built, keys in hand, on a monthly subscription.