TSMC Raises Its Spending Bet Again as AI Demand Outruns Its Own Forecasts
A 77 per cent jump in profit, a higher capex range and a sales outlook above 40 per cent growth say the foundry still sees no end to the queue.
TSMC reported second-quarter results on 16 July that would be startling from almost any other company and have become routine from this one. Revenue was NT$1.27 trillion, or US$40.2 billion, up 36 per cent on a year earlier in local currency and 12 per cent on the previous quarter. Net income rose 77.4 per cent to NT$706.6 billion, and gross margin came in at 67.7 per cent, above the top of the range the company had guided. The quarter, though, was less important than what management said about the rest of the year. Capital spending for 2026 is now expected to land between US$60 billion and US$64 billion, up from US$52–56 billion in April, and full-year revenue growth is now pitched at more than 40 per cent in dollar terms, up from more than 30 per cent three months ago.
What the node mix tells you
The wafer revenue split is the most useful single table TSMC publishes. In the second quarter, 3-nanometre accounted for 30 per cent of wafer revenue, 5-nanometre for 33 per cent and 7-nanometre for 11 per cent, so 77 per cent of the business now comes from what the company calls advanced technologies. The new entry is 2-nanometre at 3 per cent. That is a small number, but the chief financial officer, Wendell Huang, used the word "steep" to describe the ramp expected in the third quarter.
It is worth noticing that 5-nanometre is still the biggest bucket. Leading-edge nodes keep earning long after the next one arrives, and nodes whose tools are further along their depreciation schedules tend to be very profitable while customers are still queuing for the capacity. That goes a long way to explaining margins like these.
The third-quarter guidance is revenue of US$44.6–45.8 billion and gross margin of 65–67 per cent. The midpoint implies another 12 per cent sequential rise in sales and a slight step down in margin. The release does not itemise why, but ramping a new node is expensive in its early quarters, and the 2-nanometre ramp is now under way.
The demand argument, and how much of it to believe
Chairman and chief executive C.C. Wei told analysts that AI-related chip demand is growing faster than the mid-to-high 50 per cent five-year compound rate the company set out earlier in the year, though he declined to give a new figure. Asked how far to trust that demand, he described a mix of bottom-up and top-down assessments, including checks on the progress and location of customers' AI data-centre builds so that TSMC's chips do not end up sitting in inventory, and said he expects demand to stay very strong through 2029 or 2030. Huang added that capital spending over the next three years will be significantly higher than over the past three.
Two caveats are worth holding on to. First, the reading that the extra spending reflects durable demand rather than a cyclical spike is TSMC's characterisation; the customer forecasts behind it are not public. Second, the risk here is asymmetric. Equipment ordered this year produces wafers for years, and the depreciation arrives whether or not the demand does. If AI infrastructure spending plateaus in 2028, it is TSMC's balance sheet that carries the tools.
That said, the foundry has a better view of the pipeline than anyone. When the company that sees every major accelerator design before it ships raises its spending range by roughly 15 per cent mid-year, the simplest reading is that its customers are still asking for more capacity than it can supply.
What to watch
- The 2-nanometre share in the third quarter. Management has promised a steep ramp; how fast that 3 per cent climbs is the clearest public test of whether it is on schedule.
- Where the capex goes. TSMC says 70–80 per cent is for advanced process technology, around 10 per cent for specialty processes and 10–20 per cent for advanced packaging, testing, masks and other work. Advanced packaging such as CoWoS is one of the engines of AI demand, so the size of that last slice matters.
- Margin under pressure. A 65–67 per cent guide is still extraordinary, but any further dip would be the first sign that the cost of expansion is catching up with pricing power.
For builders, the practical consequence is unglamorous. Leading-edge capacity is being allocated to whoever pays most, and right now that is AI. Expect that to show up in the price of everything else built on the same wafers, from phones to laptops, for at least another year.
Sources