Lambert: Nvidia is spending $26B to seed open-model builders

Nvidia is reportedly spending $26 billion on open-source AI models. That is the number Nathan Lambert plants at the center of "Teaching Everyone to Fish for Tokens," his latest Interconnects post, and the framing he draws from it is flat: Nvidia wants you building your own model, not buying tokens from Anthropic or OpenAI.

"Nvidia wants a world where countless people can build token machines, so intelligence is not monopolized," Lambert writes. The chipmaker's incentive is direct. More independent model builders means more inference silicon sold. He wraps the strategy in a familiar image: "Nvidia wants to teach everyone to fish for tokens, so the ecosystem is self-sustaining."

Lambert sketches two ways the bet can land. In the first, the open recipe works, and Nvidia is "creating far more demand for their chips (and profits) than it costs to build the models." In the second, open models fork off from the leading closed models entirely, drifting toward "on-prem agents with private data on repetitive business tasks" rather than chasing the frontier.

He contrasts the approach with Meta's, which he says is "strategically flooding the zone with tokens" through open-weight releases (Muse Spark 1.2 is his example), undermining rivals whose business is selling tokens by the million. Along the way the essay flags a naming shift underway in the field: what used to be "pretraining, midtraining, post-training" may collapse toward "pretraining, reasoning training, and post-training" as reasoning becomes its own labeled phase.

Two of the researchers we track were passing the piece around when it went up.