Sources
Dylan Patel on Dwarkesh: 30% of $600B AI Capex Flowing to Memory — ASML Is the Ultimate Bottleneck
SemiAnalysis CEO Dylan Patel's deep dive on AI compute bottlenecks (208K views, trending) covers three constraints: logic, memory, and power. The headline finding: roughly 30% of Big Tech's projected $600B 2026 capex is flowing to memory, driven by long-context inference models requiring massive KV cache. HBM occupies 4x the wafer area of standard DDR. Patel argues power is solvable (gas turbines, fuel cells) but ASML's production ceiling of fewer than 100 EUV machines/year becomes the ultimate constraint post-2028.
↳ Follow the thread