"So we don't suddenly increase our price by which I like to have 4x or 5x." That was C.C. Wei on July 16, explaining why TSMC would not chase the 86% gross margins he openly envies in memory. The restraint sounds like discipline. It is actually a symptom.
TSMC's second-quarter numbers were the best in its history: $40.2 billion in revenue, 67.7% gross margin, 60.3% operating margin, all landing at or above the top of guidance. HPC grew 20% sequentially to 66% of revenue. N3 capacity was so tight that customers paid rush-order fees to jump the queue, lifting average selling prices on their own.
Read only that, and TSMC looks like a monopoly compounding. Read one line further into the same quarter, and the picture fractures.
Read only that, and TSMC looks like a monopoly compounding. Read one line further into the same quarter, and the picture fractures.
While 3nm customers bid up rush fees, Counterpoint projects 6/7nm utilization drifting to 80-90% over the next two or three quarters. And the mature nodes below it are being walked off the field entirely — "gradually exiting the general purpose logic market, with the unserved demand spilling over to UMC and Vanguard."
The market reads these as good news and bad news. They are the same news.
Last week we tracked Google's decision to commit more than 3 million TPUs to Intel for 2028, with Nvidia testing Intel's process for its Feynman GPUs — a diversification story that looked like Intel finally clawing back relevance. This week the demand-side driver comes into focus, and it isn't Intel winning. It's TSMC's own success eating its middle.
Start with the single most telling number in the evidence. Axis Intelligence Research calculates TSMC's revenue per active customer at $229.3 million in 2025, up from $94.0 million in 2020 — a 143.8% jump. Over the same five years, the customer count grew under 5%.
That is not a market-share story. TSMC is serving roughly the same customers at 2.4 times the average wallet.
The mechanism is AI repricing. High-performance computing was 40% of revenue in 2022; it is 66% now. The same handful of buyers — Nvidia, Apple, AMD, Broadcom, and a cluster of custom-accelerator designers — simply started spending vastly more per name.
This is where the two data points collapse into one. When your fastest-growing revenue comes from a fixed roster paying more, and that revenue concentrates in leading-edge nodes, the leading edge floods and the legacy tiers starve — not from weak demand elsewhere, but because the demand event itself is node-specific and customer-specific.
The geography follows the wallets. North America was 56% of revenue at the start of 2020; it hit 78% in the second quarter of 2026. TSMC's own filings put U.S.-headquartered customers at 75.64% of first-half 2026 revenue, against 6.85% from Taiwan and 6.76% from China.
For every NT$100 TSMC earns, roughly NT$76 now comes from customers headquartered in one country. China's slice fell from 22% to 6% as export controls tightened. The diversification that once defined the foundry model has quietly inverted.
So TSMC now runs a business with two structural problems stacked on top of each other. Its growth is concentrated in one national customer base. And within that base, it is concentrated in a single workload — AI accelerators on the most advanced nodes — that it physically cannot supply enough of.
Wei has said as much: chip supply will struggle to keep pace with AI demand for years, and even the US expansion won't fully close the gap. Nvidia's Jensen Huang says he is now TSMC's largest customer, ahead of Apple. Both the leading-edge wafer lines and the CoWoS packaging lines are running full.
That unmet demand does not evaporate. It routes.
It routes to Intel, which is why Google — after months of testing Intel's EMIB packaging — has already committed over 3 million TPUs for 2028 production, against a projected 2027-2028 TPU output north of 6 million units. It routes to Nvidia, testing whether Intel's process suits the four-die Feynman package.
This is the crucial reframing of last week's story. Intel is not being chosen because its 18A process beat TSMC. It is being chosen because TSMC's success created an overflow that has to land somewhere.
The evidence on Intel's own standing is sober about this. Counterpoint notes 18A utilization "is not yet stable," that Intel has pivoted to a customer-commitment-driven model precisely because it can no longer afford speculative capacity, and that its history is littered with volume-production delays. Intel is not the winner here. It is the relief valve.
Which is exactly what makes the second-sourcing fragile rather than curative. The whole thread we're following — does Intel diversification convert to repeat orders, or revert once TSMC's 2028 capacity arrives — hinges on whether these are strategic wins or spillover.
The evidence points to spillover. TSMC has pulled forward groundbreaking on Arizona Fab21 P3, Hsinchu Fab20 P3, and Kaohsiung Fab22 P4 into the second half of 2026. A14 hits volume production in 2028. When that capacity lands, the overflow that fed Intel could simply flow back.
Meanwhile the legacy tier keeps hollowing. Existing N5 lines are converting to N3. Part of Fab14's mature 12-inch capacity is being repurposed for silicon interposers to feed advanced packaging. Mature nodes are being handed to UMC and VIS by design, not by defeat.
TSMC is actively cannibalizing its own middle to feed the top. That is rational — the top pays rush fees and 66% margins. But it means the diversified base that once cushioned the company through downturns like the 2023 inventory correction is being deliberately dismantled.
Which brings Wei's jealousy back into focus. He won't raise prices 4x or 5x because he wants margins "enough for our long-term sustaining expansion" for both TSMC and its customers. But the restraint also reflects dependence: when 75% of your revenue comes from a handful of American buyers whose overflow you're already routing to a competitor, you cannot squeeze them the way a memory oligopoly squeezes a commodity.
Instead TSMC is negotiating a broad 10-15% price increase for 2027, per Insightology — enough to offset N2 dilution and the four-to-five-times cost of overseas fabs, not enough to test how far it can push customers who now have a second source.
That is the market pricing the fragility. A foundry whose leading edge can't keep up and whose legacy tier is going idle has traded diversified resilience for a single-geography, single-workload bet.
The 75% and the softening 6/7nm are not two stories. They are the two ends of the same crack — and the second-sourcing to Intel is the market's acknowledgment that the crack is real, not its repair.