Nvidia's CUDA Moat Faces Threat from AI Coding Agents

Nvidia's CUDA technology has been its key moat in the AI industry for two decades. This software layer turns Nvidia's silicon into a platform that developers can build on, and it took years to develop.

However, a new threat is emerging: AI coding agents that can write low-level software themselves. Jeremy Nixon, a former Google Brain researcher and founder of Infinity, claims that his team used these agents to rebuild CUDA-like software for chip firm D-Matrix in just 10 hours.

This development puts pressure on Nvidia's second advantage - the ecosystem built around CUDA, which includes millions of lines of company code and workflows. Agents can now do in hours what used to take specialist teams years, making it easier to switch to a rival chip.

Nvidia does not dispute the trend, but claims that developers rely more heavily on CUDA's libraries every year. The company also uses coding agents to build CUDA faster itself.

However, the sharper threat is a shift in what AI chips are for. As the industry moves from training models to running them, buyers care less about peak performance and more about running AI cheaply. This favours software that works across different chips, not software welded to one vendor.