Teaching Everyone to Fish for Tokens

According to Nathan Lambert at Interconnects, Nvidia is gambling twenty-six billion dollars on open-source AI to reshape the entire market. The strategy is elegant: give away the recipe and the models so that nearly everyone builds their own AI systems—and buys Nvidia chips to do it. But there's a catch. Training frontier models is brutally expensive, and it's not clear Nvidia's investment will ever repay itself. Lambert sketches two possible futures. In the first, the open-source path works, and Nvidia becomes indispensable to thousands of companies building on top of open models. In the second, and the one Lambert thinks more likely, open models splinter into a different ecology—smaller, specialized systems for on-prem enterprise tasks and long-tail use cases, unable to compete with frontier closed models from Anthropic and OpenAI. The pressure is mounting from an unexpected angle: training is getting more abstract and opaque. Fewer companies are building base models from scratch, and more are fine-tuning existing ones. Meanwhile, Meta and other hyperscalers are dumping powerful open-weight models into the market not because they're philanthropists, but to hamstring their competitors' token sales. So Nvidia's teaching everyone to fish for tokens—but Meta is strategically flooding the zone with them.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton