AI Agents Recreate CUDA-Like Software in 10 Hours, Challenging Nvidia's Moat
Updated
Updated · Business Insider · Aug 3
AI Agents Recreate CUDA-Like Software in 10 Hours, Challenging Nvidia's Moat
2 articles · Updated · Business Insider · Aug 3
Summary
Infinity said AI coding agents rebuilt CUDA-like software for chip startup D-Matrix in 10 hours, underscoring how Nvidia’s long-defended software edge may be easier to copy.
CUDA has long locked customers into Nvidia through ready-made AI code, debugging tools and workflows built over millions of lines, but rivals argue agents can sharply cut the time needed to build alternatives.
Inference is adding pressure because companies increasingly want cheaper, cross-chip software rather than maximum training performance, a shift some startups say weakens CUDA’s hold on deployment workloads.
Google, Amazon and Microsoft have already built software around their own AI chips, while Amazon documents previously identified CUDA as a major barrier to wider adoption of Trainium and Inferentia.
Nvidia says the same trend could reinforce its lead: it already uses AI coding agents to speed CUDA development, and supporters argue verification and optimization—not code generation—remain the harder moat.