Interconnects AI · 2026-08-17 · notable
Nathan Lambert: 'Teaching Everyone to Fish for Tokens' — Nvidia's $26B bet
Nathan Lambert argues Nvidia is spending $26 billion on open-source models so companies train their own instead of buying tokens from OpenAI or Anthropic. His August 17 Interconnects post lays out two futures for that bet.
Interconnects reads Nvidia's $26 billion open-model spending as demand creation for GPUs, not a bid to win the model race.
What is it?
Nvidia is spending $26 billion on open-source model work, and Nathan Lambert's August 17 Interconnects post asks what the chip maker wants back for it. The answer sits in the post's own subtitle: Nvidia wants you building your own model, not buying from Anthropic or OpenAI. Lambert reads the spending as a way to grow chip demand rather than an attempt to beat the closed labs at their own game.
How does it work?
The essay traces that money through the open-model supply chain — Nvidia's own Nemotron models, plus earlier public efforts like AI2's Olmo and EleutherAI's Pythia. Lambert points out that most teams today fine-tune existing open weights such as DeepSeek V4 Flash, Inkling Small and GLM 5.X instead of pre-training from scratch, and he reads Meta's Muse Spark 1.2 release as the same move: make tokens cheap and common.
Why does it matter?
Lambert sets out two futures for the bet. In the first, open recipes work well enough that many companies train, and Nvidia sells the chips. In the second — which he calls the more likely one — open models drift toward efficiency, modifiability and specialization, filling a long tail of enterprise-specific tasks while closed models keep the high-value work in knowledge work, drug discovery and software engineering. He notes Databricks and 01.ai left model training, but calls them anomalies rather than a trend.
Who is it for?
open-source AI followers, ML infra leads, teams deciding whether to train