AI agent breaches Nvidia's 20-year CUDA moat in 10 hours
NVIDIA
A startup called Infinity used its AI agent Ignition to develop a CUDA-like software stack for AI chip startup d-Matrix in just 10 hours. This highlights the growing threat to Nvidia's CUDA ecosystem, especially in the inference market, but experts say true disruption is far off.
Jeremy Nixon, founder of AI software startup Infinity and former Google Brain researcher, revealed that his team's AI agent Ignition spent only 10 hours building a CUDA-like software for AI chip startup d-Matrix. Ignition generates GPU kernels, runs tests, finds errors, measures performance, and rewrites code automatically, with humans providing high-level direction. The news highlights that while training AI models still relies heavily on CUDA, the inference market is becoming a battleground where CUDA's dominance is weaker. Companies like d-Matrix, Rebellions, and Cerebras are targeting this gap. However, experts caution that 10 hours of generated code is not equivalent to 20 years of CUDA's ecosystem, libraries, and verification tools. Bing Xu, founder of INT21, notes that verification is the biggest bottleneck, not code generation. Chris Lattner of Modular says coding agents bring incremental improvements and hype is exaggerated. Nvidia is also using AI agents to develop CUDA faster, so the race depends on whether competitors can catch up faster than Nvidia iterates.
Source: QbitAI 量子位 —
original