Dylan Patel (SemiAnalysis): Chips — Not Power — Are the Binding AI Compute Constraint Through 2027, Contrarian to Data Center Narrative

Dwarkesh Podcast·medium signal

In a deep-dive on the Dwarkesh Podcast, SemiAnalysis founder Dylan Patel argues the biggest AI scaling bottleneck has shifted back from power/data centers to semiconductor logic and memory production. Evidence: TSMC N3 capacity is 'very tight' through 2027, NVIDIA has overtaken Apple as TSMC's largest customer at 19% of revenue, and all three major memory vendors (SK Hynix, Samsung, Micron) are completely sold out through 2026. This is a contrarian take against the dominant 'power is the bottleneck' narrative — and explains why hyperscaler capex alone won't solve compute shortages.