Nvidia’s open-model push aims to commoditize token production and sustain chip demand

Nvidia is making a high-stakes economic wager that open-weight and open-recipe models can sustain a decentralized AI ecosystem, reducing reliance on closed providers like Anthropic and OpenAI. The company is reportedly investing $26 billion in this strategy, aiming to create a world where countless organizations build and run their own models rather than purchase tokens from a handful of dominant API providers.

The open-model ecosystem—exemplified by projects like AI2’s OLMo and EleutherAI’s Pythia—differs from open-weight releases by providing full training recipes, datasets, and code, enabling companies to reproduce and modify models. This approach contrasts with the transient nature of open-weight models, which often serve as components in larger software stacks. Nvidia’s Nemotron series follows this model by releasing training data and code where legally permissible, positioning itself as a facilitator of a self-sustaining token economy.

The success of Nvidia’s strategy depends on whether open-model development can remain financially viable. Analyst Nathan Lambert outlines two potential futures: one in which open models become competitive enough to drive massive demand for inference hardware, justifying the $26 billion investment; and another in which capital intensity forces most companies out of training, leaving open models to occupy a long-tail of specialized, efficiency-focused applications such as enterprise-specific agents running on-premises with private data.

Lambert suggests that the current trend—where fewer organizations release base models and more focus on post-training or fine-tuning—may accelerate if training becomes too opaque or capital-intensive. This shift could further reduce incentives to invest in open-source AI, particularly as post-training increasingly resembles large-scale pretraining in complexity. Revenue-sharing licenses and other monetization experiments for open models are emerging as attempts to sustain near-frontier development, but their success remains uncertain.

The strategic divergence between Nvidia’s push for open-model proliferation and hyperscalers like Meta— which release strong open-weight models to commoditize their competitors’ token sales—underscores a broader industry tension. While Nvidia seeks to teach the ecosystem to “fish for tokens,” Meta’s approach of flooding the market with open models may slow the revenue growth of closed API providers like Anthropic and OpenAI by making intelligence a more abundant and less monetizable resource.