TSMC Beat Its Guidance and Says AI Demand Still Outruns Capacity

If you want one number that tells you whether the AI build-out is still accelerating, TSMC's quarterly report is a good place to look. Nearly every leading AI accelerator is made in its fabs. On October 15, it reported a quarter that beat its own guidance and said demand is still running ahead of what it can build.

The quarter

Third-quarter revenue was $46.72 billion, above guidance of $44.6 billion to $45.8 billion. That was up 16.2% from the second quarter and 41.2% from a year earlier. Earnings per share rose 68.7% year on year to NT$29.43.

The profit lines were nearly as striking. Gross margin came in at 66.4%, inside the 65% to 67% guidance range, and operating margin at 58.7%, above guidance. More than half of every dollar of revenue became net profit: the net margin was 51.1%.

Cash generation was even stronger. Operating cash flow reached NT$919.77 billion, up from NT$783.36 billion in the second quarter, and free cash flow was NT$476.56 billion.

Where the money comes from

TSMC does not report a single "AI" line, but its platform split leaves little doubt. High-performance computing, which includes AI accelerators, server and PC processors and networking chips, was 68% of revenue and grew 22% quarter on quarter. Smartphones, once the company's anchor, were 21%.

By process node, the leading edge is now the business. 3-nanometre chips were 33% of wafer revenue and 5-nanometre were 28%. 2-nanometre, still in early ramp, was already 8%. Altogether, 7nm and below made up 79% of wafer revenue.

The volume side tells you this is pricing and mix, not just more wafers. TSMC shipped 4,476 thousand 12-inch-equivalent wafers in the quarter, up only 3.2% from the second quarter, while revenue rose 16.2%. Each wafer is worth more because more of them are leading-edge, and leading-edge wafers carry the large AI dies and advanced packaging that customers are queuing for. Revenue from 7nm and below approached NT$1 trillion in the quarter, against roughly NT$350 billion in the first quarter of 2024. Inventory turnover also tightened, to 82 days from 87, another sign that output is leaving the building about as fast as it is made.

Why the stock fell anyway

Shares closed down 3.4% at $430 after the report and slipped further after hours. The reason sits in one line: gross margin fell 130 basis points from the second quarter's 67.7%.

Two things are pulling on it. The first is the 2-nanometre ramp. New nodes are expensive while yields climb and fixed costs are spread across few wafers, so every quarter of ramp dilutes margin before the node starts paying for itself. The second is overseas manufacturing. TSMC said its fabs outside Taiwan currently dilute margin by about 2%, a drag it expects to grow to 3% to 4% in coming years as those sites scale.

Neither is a demand problem. Both are costs of growing faster, in more places, at the most advanced nodes. Investors priced them anyway.

The outlook

Management raised the bar again. Fourth-quarter revenue guidance is $52.3 billion to $54.3 billion, roughly 58% above a year earlier at the midpoint. TSMC now expects full-year 2026 revenue growth of about 44% in US dollar terms. Its capital budget for the year stays at $60 billion to $64 billion.

On the roadmap, the company expects 2-nanometre capacity to more than double in 2027, with its A14 node entering risk production in 2027 and volume production in 2028.

The most important comments were qualitative. Chief executive C.C. Wei described demand as "extremely robust." Management said AI accelerator growth will "probably" come in above the mid-to-high-50s percent compound annual rate it had previously guided for 2024 to 2029, and that underlying demand for accelerators, networking chips and advanced packaging remains above TSMC's current production capacity.

Reading it against the AI-capex debate

This quarter lands in the middle of a noisy argument about whether AI spending has peaked. Hyperscaler capital budgets, credit spreads on GPU-backed debt and the valuations of AI labs have all been read as signs of strain.

TSMC sits upstream of all of it, and its signal is unambiguous for now: orders for leading-edge AI silicon are still growing, and the binding constraint is how fast it can add capacity, not whether customers will show up. That does not prove every downstream investment will pay off. It does mean the companies buying chips are still buying more of them.

The takeaway

TSMC delivered a beat, a raise and a warning about margins in a single report, and the market chose to listen to the warning. That is a reasonable reaction for a stock that has already priced in a great deal of growth.

But the core message is the one investors should keep: the world's dominant AI chip foundry says it cannot yet build as much as its customers want, and it expects to sell about 44% more this year than last. Whatever happens to AI valuations, the physical build-out is still speeding up.