A group of technology heavyweights has asked United States policymakers to go easy on open-weight artificial intelligence, the freely downloadable models whose internal settings anyone can inspect and reuse.
Nvidia, Microsoft and Meta signed the public letter, alongside startups including Replit and Reflection, arguing that the country should lead the world in developing and deploying such models.
The framing is patriotic, but the motivation looks commercial.
The tell is who stayed away.
OpenAI and Anthropic, the two labs with the strongest closed-source models, did not sign.
That split maps neatly onto economics rather than principle.
If closed models capture most customer revenue and margin, the firms selling the surrounding layers, the chips, the middleware and the hosting, are left fighting over scraps.
An open ecosystem, by contrast, commoditises the model itself and pushes value towards infrastructure, which is precisely where the signatories sit.
History supports the bet.
When proprietary software dominates a market, open alternatives tend to emerge and capture the value in other ways, a pattern running from Netscape's 1998 browser release through Red Hat's support business for Linux to companies such as Databricks and ClickHouse that grew out of open-source projects.
The signatories are wagering that large language models follow the same script, ending in a mix of proprietary and open-weight products rather than a closed-model monopoly.
The immediate trigger was foreign.
Early reactions to Moonshot AI's Kimi K3 model sharpened a debate that began as a narrow worry about Chinese systems and widened into a fight over open-weight regulation in general.
That widening is the danger the letter is really aimed at.
A central concern is that regulators cannot easily tell a foreign open-weight model from a domestically built one, and may reach for rules broad enough to catch both.
The other worry is distillation, the practice of training a smaller model on the outputs of a larger one.
Critics say open models lean on distillation from proprietary systems to close the performance gap, and proponents fear a ban on the technique.
The awkward fact for critics is that closed labs distil too, using their own models to train the next generation.
Sweeping AI rules could therefore hobble American open-weight efforts while doing little about the overseas models they were meant to address.
Notable absences complicate the picture.
Google did not sign, and neither did Amazon, despite hosting open-weight models on its cloud, though Amazon has already dissolved its in-house proprietary model team and may yet come round.
For now the letter reads less as a statement of conviction than as a coalition of firms that would profit if the model layer stayed cheap and open.
