Somebody is paying for the open weights, and it is not you

Nathan Lambert wrote this week about who actually funds open models, and his framing is sharper than the usual Linux comparison. He splits open source models, meaning a full recipe with data and training code, from open weight models, which behave more like a specific version of software you install and build on top of. Weights are fairly transient. A recipe is something anybody can pick up, modify, and press run on.

Then the money. He passes along a report that Nvidia is spending twenty six billion dollars on nearly open source models, and reads the motive plainly. Nvidia wants a world where countless people can build token machines, because that world buys an enormous number of chips. Meta, he argues, is using the same tool for a different end, releasing strong open weights to slow the revenue growth of the companies that sell tokens. Both are commoditizing their complements, in his words. One is teaching everyone to fish. The other is flooding the zone.

The honest version of the downside is the part I would keep. If neither financing path returns, open models fork toward efficiency, modifiability, and specialization, filling a long tail while closed models hold the most valuable ground. He calls that his most likely outcome, which is not a thing people say on stage.

The vault’s economics notes are a lens here rather than a settled fact base, and the lens still helps. Open weights work as bargaining pressure on frontier providers, and every lane owes you a total cost answer covering serving, latency, reliability, review time, and integration.

So the practical question for anyone building on an open model is not whether the weights are good. They are. It is who is paying for the next set, and what that party is actually buying.