TSMC Risk (Stratechery Article 1-26-2026)
Ben Thompson argues the main TSMC risk for AI is not just Taiwan geopolitics but TSMC’s conservative capacity expansion: Amazon, Microsoft, Google, and Meta all said in late-2025 earnings that AI demand still exceeds supply, and TSMC CEO C.C. Wei explicitly said the bottleneck is "silicon from TSMC," not power, cooling, racks, or turbines.
- TSMC’s underinvestment predates the current shortage: after a large 2021 capex increase tied to COVID shortages and 5G expectations, annual capex was essentially flat and even declined year-over-year in 2023 and 2024, despite ChatGPT’s November 2022 launch triggering a hyperscaler AI spending surge.
- TSMC is now raising capex, but on a lag that matters for AI buildout timing: 2025 capex rose 37% to $41 billion, and 2026 guidance is $52-56 billion (about $54 billion midpoint, up roughly 32%), yet Wei said new fabs take 2-3 years to build, meaning little effect in 2026, only some effect in 2027, and significant capacity mainly in 2028-2029.
- Wei framed TSMC’s stance as rational foundry risk management: because foundries are overwhelmingly capex-driven businesses, overbuilding can turn high-margin revenue into years of depreciation, weak pricing, and potentially "a big disaster," so TSMC is effectively offloading demand-risk onto hyperscalers and fabless chip firms.
Ben Thompson’s January 26, 2026 Stratechery essay reframes “TSMC risk” away from the familiar Taiwan-invasion narrative and toward a more immediate economic bottleneck: TSMC’s monopoly position and conservative capital allocation are constraining AI infrastructure growth. He ties together hyperscaler earnings commentary from Amazon, Microsoft, Google, and Meta, all of which reported that AI demand still exceeds supply, with TSMC CEO C.C. Wei’s own admission that the limiting factor is chip output rather than electricity, cooling, or other data-center inputs. That matters because it suggests the practical bottleneck in the AI buildout is at the foundry layer, not the grid or server deployment layer many observers focus on.
The article’s central evidence is the mismatch between hyperscaler capex acceleration after ChatGPT’s November 2022 debut and TSMC’s comparatively flat capex in 2023-2024. TSMC is now responding—2025 capex rose to $41 billion and 2026 guidance is $52-56 billion—but Wei says fabs take two to three years to come online, so meaningful relief arrives mostly in 2028-2029. Thompson argues this is rational from TSMC’s perspective because foundries bear enormous fixed-cost risk, but irrational for the ecosystem because it shifts the downside onto customers that lose revenue when demand cannot be served. His strategic conclusion is that hyperscalers and fabless chip firms should absorb the pain of qualifying Samsung or Intel as real alternatives; only credible foundry competition, he argues, will force more aggregate capacity investment and reduce both economic and geopolitical concentration risk.