“TSMC Risk” — Ben Thompson
Why this is in the vault
Definitive articulation of why TSMC’s monopoly is the binding constraint on AI scaling, and why foundry competition is the only solution — important macro context for AI infrastructure.
The core argument
The TSMC risk isn’t primarily about China invading Taiwan (though Thompson covers that). The real risk is economic: TSMC’s rational conservatism in capex is already costing hyperscalers billions in foregone revenue, and will cost exponentially more by 2028. Every major hyperscaler (Amazon, Microsoft, Google, Meta) reported demand exceeding supply last quarter. TSMC’s CEO admitted silicon is the bottleneck, not power.
TSMC’s capex was flat/declining in 2023-2024 despite ChatGPT launching in late 2022 — they didn’t buy the AI hype. Even with 2026 capex up to $52-56B, new fabs take 2-3 years, so relief comes only in 2028-2029. Thompson argues hyperscalers need to invest in making Intel and Samsung viable alternatives — not to replace TSMC, but to create competitive pressure that makes TSMC invest more aggressively. The risk of working with non-TSMC foundries (lower quality, worse service) is real but smaller than the risk of insufficient total industry capacity.
On geopolitics: Thompson favors keeping China dependent on TSMC and US chips (including allowing H200 sales) rather than cutting them off, which would incentivize China to destroy TSMC.
Mapping against Ray Data Co
No direct connection. Useful macro context for AI compute availability and cost planning.
Related
- 2026-01-21-stratechery-tsmc-earnings-foundry-competition