NVIDIA, Microsoft, and Meta Tell Washington Not to Lock Down Open AI

There's a coalition letter making the rounds this week, and the guest list alone tells you something: NVIDIA, Microsoft, Meta, Google, OpenAI, IBM, Cisco, Palantir, Andreessen Horowitz, Hugging Face, and dozens more, spanning chipmakers, cloud giants, startups, and open-source nonprofits like the Linux Foundation and Mozilla. Rivals who compete fiercely for enterprise AI contracts are, on this one issue, standing shoulder to shoulder.

The pitch draws a direct line back to the 1980s open-source software movement, arguing that the same transparency that let engineers everywhere study and improve code is what America needs now for AI weights, the actual parameters that make a model run. The argument isn't just philosophical. Open weight models let a hospital, a farm, or a small manufacturer grab an existing model and adapt it instead of training something from scratch or paying frontier-model prices for every routine task. That matters because, as the letter puts it, AI's economic payoff depends on matching the right-sized model to the right job, not throwing the biggest model at everything.

There's also a competition argument buried in here that's worth sitting with. Open weights don't just create rivalry between the handful of companies building frontier models, they spur competition across chips, cloud infrastructure, and applications too. Fewer companies controlling the technology means fewer single points of failure, and the letter is blunt that concentrating advanced AI in a small number of closed systems is itself a risk, not a safety feature.

On security, the coalition pushes back hard against the instinct to lock things down after something goes wrong. Yes, once weights are released, nobody can fully control what happens to modified versions downstream. But the letter's counter is that defenders need models as capable as the ones attackers already have access to, and that a broad community poking at open models tends to find and fix vulnerabilities faster than a closed shop ever could. Closed models get breached too, the letter notes, they just fail in ways nobody outside the company can see.

The letter also tries to draw a bright line around distillation, the practice of training one model using another's outputs, insisting it's a legitimate and long-standing technique for improving models rather than some kind of theft. Unlawful extraction of value from closed systems is a separate problem, they argue, one that calls for targeted legal remedies rather than broad restrictions that would sweep up ordinary model development along with it. The ask to policymakers boils down to three things: more compute access for startups and researchers, investment in shared datasets and evaluation tools, and resisting the urge to slap premature restrictions on open models before the harms are even demonstrated.