“TSMC Earnings, The TSMC Brake Revisited, Why AI Needs Foundry Competition” — Ben Thompson
Why this is in the vault
Explains the silicon supply bottleneck constraining AI scaling — TSMC is the brake on the entire AI buildout, and competition from Intel/Samsung is the only way to release it.
The core argument
TSMC’s Q4 2025 earnings confirmed Thompson’s “TSMC Brake” thesis. CEO C.C. Wei admitted that TSMC underinvested in prior years and current demand far exceeds supply. Despite boosting 2026 capex to $52-56B (from $41B), new fabs take 2-3 years to build, so supply relief won’t arrive until 2028-2029. Cloud providers told Wei that silicon from TSMC — not power — is the real bottleneck.
The key strategic argument: hyperscalers should invest in bringing Intel and Samsung online as credible foundry alternatives, not to replace TSMC but to shift risk. Currently TSMC is risk-averse (afraid of overbuilding into a bubble), which means it systematically underinvests relative to demand. Only competitive pressure from viable alternatives would push TSMC to invest more aggressively. Wei acknowledged being “nervous” about whether AI demand is real despite evidence that it is.
Mapping against Ray Data Co
No direct connection to RDCO’s workstreams. Background context for understanding AI compute cost trajectories and availability constraints.
Related
- 2026-01-07-stratechery-nvidia-ces-vera-rubin