Dwarkesh Podcast: Dylan Patel (SemiAnalysis) on the three bottlenecks to scaling AI compute#
| URL | https://www.dwarkesh.com/p/dylan-patel |
| Published | 2026-03-13 |
| Read | 2026-09-26 (show page and transcript excerpts) |
| Kind | Podcast; spoken numbers are looser than written research |
Summary (our words)#
- The three bottlenecks: logic, memory and power. The semiconductor supply chain has the longest lead times.
- When each binds: memory in 2026–27; ASML's EUV tool output becomes the top constraint by about 2030. About 70 tools a year now, about 80–100 by 2030; about 3.5 tools per gigawatt.
- Memory prices: keep rising through 2027–28; new fabs arrive late 2027 or 2028.
- Nvidia: will hold 70%+ of N3 in 2027.
- Power: "will not be a problem" in the US.
- Old GPUs: an H100 is worth more today than three years ago, because better models extract more value per hour.
- Hopper economics: long-term rentals at $2–2.40/hr against about $1.40/hr cost on five-year depreciation.
- Labs: Anthropic added $4B of run-rate in January and $6B in February; OpenAI goes from about 2 GW at end 2025 to about 10 GW by end 2027.
- The risk he names: that returns on the infrastructure turn out middling.
Claims#
| ID | Claim | Evidence |
|---|---|---|
| SA-dwarkesh-1 | Memory prices keep rising through 2027–28; new fabs late 2027 or 2028 | Transcript |
| SA-dwarkesh-2 | Nvidia holds 70%+ of N3 wafer capacity in 2027 | Transcript |
| SA-dwarkesh-3 | Smartphones "might drop to 500–600 million this year" from 1.4 billion | Transcript |
| SA-dwarkesh-4 | An H100 is worth more today than three years ago | Transcript |
| SA-dwarkesh-5 | Power won't constrain US scaling | "Scaling power in the US will not be a problem." |
| SA-dwarkesh-6 | The downside risk is middling returns on infrastructure, not a physical limit | Transcript |
| SA-dwarkesh-7 | Hopper long-term rentals are $2–2.40/hr against about $1.40/hr cost on five-year depreciation | Transcript |
Our notes#
- SA-dwarkesh-3 doesn't hold. IDC's worst case is −16.7%, about 1.0 billion units. Treat his podcast numbers as tone.
- SemiAnalysis sells to the buildout. Weight their supply-chain facts above their demand views.