Dylan Patel (SemiAnalysis): The Three Bottlenecks to Scaling AI Compute — Logic, Memory, and Power

Dwarkesh Patel·high signal

Dwarkesh Patel interviews SemiAnalysis founder Dylan Patel (182K views) on the three fundamental bottlenecks to scaling AI compute: logic (chip design limits), memory (bandwidth constraints), and power (energy infrastructure). Patel walks through the economics of AI labs vs. hyperscalers, why memory bandwidth may be the single biggest constraint, and the implications for the next generation of training runs. Essential context for understanding why compute scaling isn't just about building more data centers.