CUDA software optimization reportedly reached 2x in a month with AI assistance

SemiAnalysis says Nvidia engineers at the PyTorch Conference described faster CUDA software optimization with frontier AI models. The figures they cited were 5x from October 2025 to August 2026 with expert human optimization, 2x from September to October 2026 with AI assistance, and 1.5x in October with AI orchestration and MegaKernels.