Nvidia Stops Making Chips and Starts Making Models

Image on Unsplash

Nvidia launched Nemotron 3, a family of open-source AI models. This marks an unusual shift for the world's most dominant semiconductor manufacturer. 

The models are designed to be faster, cheaper, and more efficient than previous offerings, and the company is positioning them as an answer to the rise of open-source models from Chinese AI laboratories that have begun challenging Nvidia's market dominance.

Nvidia has spent the past two years dominating the hardware layer of AI development. Companies buy Nvidia chips to train and run their models. But as the market matures, the economics are shifting. Customers want cheaper inference. They want models that run on less expensive hardware. They want open-source alternatives to the proprietary models from OpenAI and Google. Nvidia's response is to stop being purely a hardware company and become a full-stack AI provider.

Chinese AI labs have released dozens of open-source models this year that are competitive with frontier models but available for anyone to use without licensing fees. Meta released Llama 3.1. Mistral released models that outperform some closed alternatives. As the competition moves from hardware to software, companies that control the entire stack gain leverage. Nvidia is moving to control the stack rather than just the chips.

However, Nvidia's decision might actually accelerate the fragmentation of the AI market. If Nvidia produces models, then customers have less reason to buy chips from Nvidia to run other companies' models. The company is betting that controlling both the hardware and software will give it more competitive advantage than controlling just the hardware. Whether that strategy works depends on whether Nemotron 3 can compete on quality and efficiency with models from companies whose entire focus is model development rather than chip manufacturing.