Teaching Everyone to Fish for Tokens
| Source: Interconnects (Nathan Lambert)
Tags: Nvidia, Nemotron, open-source AI, AI strategy, OLMo, model training
Nvidia's $26B investment in near-open-source AI models is a deliberate chip-demand strategy: the more companies that can build and train their own models, the more Nvidia hardware they buy—a potentially self-sustaining flywheel that also prevents OpenAI and Anthropic from monopolizing AI intelligence.
Details
Nathan Lambert, a former Ai2 researcher who helped build the OLMo open models, argues that Nvidia's aggressive funding of open and near-open-source AI models (like Nemotron, at a reported $26B) is fundamentally a chip-demand play. The logic: a world where many companies can build their own token machines means massively distributed inference demand, all running on Nvidia GPUs. Lambert draws a parallel to how Linux became self-sustaining once it achieved critical mass—but notes the analogy breaks for open-weight models (weights only, no training recipe), which are closer to specific software versions. Truly open-source models (full training recipe + data + code, like OLMo or Pythia) are closer to the OS analogy and are what Nvidia is investing in. The piece lays out two futures: either the strategy works—generating far more chip demand than the $26B investment costs—or capital intensiveness drives more companies out of model training entirely, concentrating AI in a few large labs. Lambert points to Databricks and 01.ai backing away from training as early signals, though he notes they may be anomalies. This is insightful analysis from a practitioner, but represents commentary and strategic interpretation rather than breaking news.