Why Hardware-Software Co-Design Is AI's Real 100x: Dylan Patel of SemiAnalysis
Sequoia Capital
Jun 30, 2026
Dylan Patel argues that the most significant gains in AI performance arise from co-optimizing hardware, software kernels, and model architecture simultaneously. This deep integration, rather than isolated hardware improvements, creates 100x efficiency leaps. Consequently, the industry is shifting toward specialized, heterogeneous compute stacks rather than relying on one-size-fits-all general-purpose solutions.
Key insight: The most efficient AI deployments today achieve performance gains not just from faster chips, but from 'co-design' where the model's expert structure and memory access patterns are custom-built for specific hardware architectures like Hopper or Blackwell.
