Why Nvidia is spending billions to help everyone build their own models

Nathan Lambert makes a clean case for why Nvidia keeps pouring money into open models: every company that trains and serves its own model buys more chips, so Nvidia's interest is to teach everyone to fish for tokens rather than sell them fish. He points to a reported $26 billion Nvidia is putting toward open-source model work. The company does not need those models to earn API revenue the way OpenAI or Anthropic do. It needs the demand for compute that a wide field of independent builders creates.

Lambert separates this from Meta, which he describes as flooding the zone with tokens through releases like Muse Spark 1.2, a move aimed more at the competitive landscape than at growing a healthy base of model developers. Nvidia's version is closer to seeding an ecosystem it profits from at the hardware layer. The catch is that building a base model from scratch keeps getting harder and less transparent. Lambert notes the vocabulary is starting to shift from "pretraining, midtraining, post-training" toward "pretraining, reasoning training, post-training," and the reasoning stage is where the hard-won, undisclosed technique now lives. As that complexity rises, fewer groups can produce a competitive base model, and open weights risk concentrating among a handful of well-funded labs. He notes that many companies still run workflows on Llama 3, a reminder that older open weights keep paying off long after release.

Why it matters

If you depend on open models, watch who funds them and why. A world where open weights come mainly from Nvidia's incentives or a few hyperscalers is more fragile than one with many independent trainers, and it shapes which models you will actually be able to build on in a year or two.