Washington is heading into another round of AI rulemaking, and the industry is not waiting quietly. This week, two dozen companies and organisations, spanning direct commercial rivals and groups with little obvious shared interest, signed a joint open letter urging US policymakers to protect open-weight AI. The signatories include Meta, Microsoft, Nvidia, IBM, Dell Technologies, CrowdStrike, Palantir, ServiceNow, Hugging Face, Perplexity, Mistral, Andreessen Horowitz, Y Combinator, the Linux Foundation and Mozilla. For anyone deploying AI inside a business, this matters, because the rules written in the next legislative cycle will decide whether self-hosted models remain a serious option or become a compliance headache. Below is a straight read of what the letter says, why the coalition looks the way it does, and what procurement teams should take from it.
This article covers:
What the open letter actually asks for
What open-weight AI means, and how it differs from closed models
The security argument the signatories are making
Why distillation gets its own carve-out
Who signed, and why the coalition looks the way it does
What this means for procurement and policy teams
What the open letter actually asks for
The letter, published as a PDF and reported by AI News, does not propose specific legislation. It is a positioning document that lands ahead of anticipated action on AI policy in Washington. The concrete requests are that lawmakers expand compute access for startups and researchers, fund shared training datasets and evaluation frameworks, and avoid what the signatories call "premature restrictions" on open-weight AI.
The framing is deliberately historical. The signatories compare the current fight over model weights to the open-source software debate of the 1980s, arguing that capability circulating freely reaches more users, gets inspected by more researchers, and does not concentrate power in a handful of well-resourced labs. The letter puts open weights forward as the mechanism by which AI reaches "factories, hospitals, farms, classrooms, and main street businesses."
What open-weight AI means, and how it differs from closed models
Open-weight AI refers to models whose trained parameters, the numerical values learned during training, are published for anyone to download, inspect, modify, and run on their own hardware. That is different from a closed model such as those offered by OpenAI or Anthropic, where the weights stay inside the vendor and users only reach the model through a commercial API.
The distinction matters because it changes control after release. With an API-only product, the vendor can update behaviour, throttle access, or withdraw the service. With open weights, once the file is downloaded, the user runs it locally on their own terms. That is the appeal for regulated industries and sovereign deployments. For a look at a current frontier open-weight release, see.
The security argument the signatories are making
This is the most pointed section of the letter, and it inverts the usual framing. The signatories concede the standard concern: once weights are released, they are beyond the original developer's control, modified versions become hard to trace, and a fine-tuned variant with safety guardrails stripped out can circulate with no recall mechanism.
Their answer is that prohibition is the wrong response. The argument rests on a cybersecurity comparison. Defenders facing AI-equipped attackers need access to models with comparable capability to detect and simulate threats, and permission-gated closed systems do not easily provide that. Closed models, the letter argues, are not inherently safer, because they can be breached, misused, or fail in ways external researchers cannot observe. Concentrating capability behind a small number of closed providers creates single points of failure rather than removing them.
Open models, by contrast, let outside teams examine behaviour, run red-team exercises, and identify vulnerabilities across many organisations. The parallel drawn is to the "open-source is more secure than obscurity" argument from decades of software security debate. The letter does not cite specific vulnerability-discovery data for AI systems, so the analogy is asserted rather than measured.
Why distillation gets its own carve-out
The letter carves out space for distillation, the technique where one model's outputs are used to train or improve a second model. Distillation is standard practice in machine learning, used for evaluation, validation, and transferring capability from a larger model to a smaller one.
The signatories draw a line between distillation as a legitimate technique and "unlawful efforts to extract value from closed models," arguing the former should not get swept up in restrictions aimed at the latter. This reads as a direct response to disputes that flared after the rise of Chinese models such as DeepSeek and Kimi, when several US labs alleged that rival models had been trained by distilling outputs from their closed systems without authorisation. The recommended fix is targeted legal action, not blanket restrictions on a technique the entire field depends on.
Who signed, and why the coalition looks the way it does
The signatory list is the most interesting part of the story, because it groups companies that would not usually appear on the same document. The commercial motives sort into a few clear groups.
Signatory group | Interest in open-weight AI |
|---|---|
Hardware and chips (Nvidia, Dell, IBM) | Wider deployment sells more compute |
Hyperscalers and platforms (Microsoft, Meta) | Open models expand ecosystems around their tools |
Open-model labs (Mistral, Hugging Face, Perplexity) | Their products depend on open releases |
Enterprise software (Palantir, ServiceNow, CrowdStrike) | Customers want self-hosted options |
Venture and community (a16z, YC, Linux Foundation, Mozilla) | Portfolio and mission alignment |
The infrastructure providers have the clearest commercial motive. Nvidia, IBM and Dell benefit whenever a wider range of deployable models exists, because more open ecosystems sell more chips, servers and services regardless of which lab produced the weights. Meta has a separate reason: its Llama family is one of the most-downloaded open-weight lines in production, part of the same broader AI push covered in.
What this means for procurement and policy teams
Treat the letter as an indicator of where major infrastructure and chip providers want the regulatory conversation to land, not as a settled outcome. No legislative proposal is attached, and the politics remain unresolved.
Three practical takeaways for teams weighing open-weight versus closed-model deployments:
Factor policy risk into total cost of ownership. Restrictions on distillation or open releases could shift the economics of self-hosted AI within a single legislative cycle.
Do not assume a closed vendor is automatically more compliant. The security argument in the letter, whether or not you buy it, is now part of the record regulators will read.
Keep an evaluation path for both approaches. Vendor comparisons made six months ago may need refreshing. covers the internal capability side of that decision.
FAQs
What is open-weight AI?
Open-weight AI describes models whose trained parameters are published for anyone to download, inspect, modify, and run on their own hardware. That contrasts with closed models like those from OpenAI or Anthropic, which are only reachable through a commercial API and whose weights stay inside the vendor.
Who signed the open letter?
Signatories include Meta, Microsoft, Nvidia, IBM, Dell Technologies, CrowdStrike, Palantir, ServiceNow, Hugging Face, Perplexity, Mistral, Andreessen Horowitz, Y Combinator, the Linux Foundation, Mozilla and others, totalling about two dozen companies and organisations.
Does the letter propose specific legislation?
No. It is a positioning document ahead of anticipated AI policy action in Washington. The concrete asks are to expand compute access, fund shared datasets and evaluation frameworks, and avoid what the signatories call "premature restrictions" on open-weight models.
Why does the letter defend distillation?
Distillation is a standard machine learning technique used across the field. The signatories argue that legitimate distillation should not be lumped in with unlawful efforts to extract value from closed models, and that misappropriation is better handled through targeted legal action.
What should procurement teams take from this?
Policy risk is now a live variable. Restrictions on open releases or distillation could change self-hosted economics within one legislative cycle, so keep both deployment paths on your evaluation shortlist rather than committing on today's rules.