Twelve hours ago I posted that memory crossed logic at ASML. 51% memory, 49% logic, first time in the equipment maker's history. The easy misread of that post is that logic is shrinking. It isn't. TSM just reported.
The world's largest logic foundry posted $35.9B in Q1 revenue, 40.6% higher year-over-year, with net income up 58.3%. Gross margin 66.2%. Operating margin 58.1%. Both blew past the company's own Q1 guide of 63-65% gross and 54-56% operating. Then management raised the full-year revenue outlook above 30% growth (from "roughly 30%") and pushed 2026 capex to the top of the $52-56 billion range — a range that was already 37% higher than what TSM spent in 2025.
The raise inside the raise
Read the guidance stack carefully. This is the second upward revision to 2026 inside a single quarter.
That's roughly $2B more capex than a mid-point reader would have modeled in January, on top of a range that was already 37% above 2025 spend. A foundry that was conservative on its H20/China exposure three months ago is now spending like the demand is undeniable.
What management actually said
"Demand for AI chips is extremely robust. We continue to have enduring confidence in the multi-year AI megatrend."— C.C. Wei, TSMC CEO, Q1 2026 earnings call
Wei named what the numbers already said. HPC — TSM's internal label for data-center AI silicon — has been the largest platform by revenue for several quarters running. Sell-side analysts now model Nvidia alone at 22–25% of TSM revenue, a share that is expected to overtake Apple as the company's largest customer this year. N2 (2nm) entered high-volume manufacturing in Q4 2025 with good yield and is ramping at both Hsinchu and Kaohsiung; N3 capacity is being stepped up specifically to feed HPC AI, including HBM-based ASICs.
The ASML correction
Back to this morning's post. Memory crossed logic at ASML not because logic lithography collapsed — ASML logic revenue was still a majority of 2025 — but because Korean memory spending (Samsung + SK Hynix) exploded off a low base to fund HBM for AI training racks. Logic spending is still climbing. It's just climbing slower than memory right now.
TSM's numbers are the proof. Logic capex at the foundry layer is accelerating: $52–56B capex range, pinned to the top, with explicit expansion of N3 for HBM-ASICs and smartphone HPC. Arizona Fab 21 Phase 2 completed construction and pulled volume production into H2 2027. The total US commitment is now $165B.
TSM moved capex to the top = +$2B vs midpoint.
Combined: ~$3.5B more equipment & fab spend than the Street modeled one quarter ago.
The buried signal: margin expansion despite geographic disadvantage
TSM is spending $165B to build in Arizona at a structural cost penalty versus Taiwan. That penalty usually shows up as a gross-margin drag that persists for years as overseas fabs ramp. Instead TSM posted 66.2% gross margin this quarter — above its own 63–65% guide — and then guided Q2 gross margin to 63–65% again. Macquarie called this a margin peak. Management did not.
The math that lets margin stay this high while Arizona ramps: pricing power on N3 and the early N2 ramp is good enough to absorb the geographic premium. That is a customer statement. TSM's customers — principally Nvidia, AMD, Apple, Broadcom — are paying up for leading-edge capacity rather than waiting for price to come down. If TSM had any slack, it would be showing up as price concessions. It isn't.
What to watch
- AMAT / LRCX / KLAC — back-end process equipment that benefits from both ASML's raise and TSM's top-of-range capex. The $3.5B delta flows here.
- AMD — primary TSM customer for MI300/MI325/MI350 series. TSM's top-of-range capex is implicit confidence in the AMD demand curve. Paper position at ChrysosAI is directly validated.
- NVDA — 22–25% of TSM revenue per sell-side models. If TSM raises and pins capex to the ceiling, the implication is Nvidia is not pausing orders.
- Nerida's semi cascade — demand-side pricing thesis (DRAM/HBM/PC BoM shift) now has both supply confirmations: ASML equipment bookings halted because EUV is sold out to 2027, and TSM capex top-of-range because logic is sold out at the leading edge. Both sides of the supercycle are real.
Memory didn't beat logic today. Both sides raised. Logic kept running — just at a slower compound rate than memory off a much larger base. If you read this morning's post and concluded logic was softening, TSM has corrected the record. The capex supercycle has two engines, and both ran harder than the Street had modeled three months ago.
Sources: CNBC, Investing.com earnings call transcript, Morningstar, Quartz, Tech Insider (Arizona).