Dylan Patel (SemiAnalysis) Deep Dive: Memory Is Now the Core AI Compute Bottleneck at 30% of CapEx

Dwarkesh Podcast·high signal

On the Dwarkesh Podcast (175K views), SemiAnalysis CEO Dylan Patel details three sequential AI compute bottlenecks — logic (TSMC N3), memory (HBM), and power — where bottlenecks rotate roughly annually. The key 2026 finding: memory will absorb 30% of major tech CapEx because long-context inference's KV Cache has 'completely ignited' HBM demand, and HBM physically requires 4x the wafer area of DDR memory. Nvidia locked up TSMC N3 allocation early — explaining why H100 pricing is higher today than three years ago despite more supply.