TSMC Q3 2026: $46.7B Revenue, HPC at 68%, and a $52.3B to $54.3B Q4 Guide

TSMC reported its third quarter of 2026 on 15 October, and the AI build-out showed up in every line. Revenue was US$46.72 billion, up 41.2% year on year in dollar terms, with a 66.4% gross margin and net income up 68.7%, according to the company's 6-K filing. High-performance computing, the platform that holds AI accelerators, was 68% of revenue. Guidance for the fourth quarter is US$52.3 billion to US$54.3 billion. C.C. Wei's message on the call was blunt: "we still think the demand is higher than the supply."

What happened

The quarter ended 30 September. In New Taiwan dollars revenue rose 50.9% year on year and 17.6% sequentially; in US dollars it rose 16.2% on the quarter. Operating margin was 58.7%. Advanced nodes (7nm and below) made up 79% of wafer revenue, with 3nm alone at 33% and 2nm, in its initial ramp, at 8%.

On the earnings call, management said HPC revenue grew 22% sequentially to 68% of the total. TSMC raised its full-year 2026 growth expectation to around 44% in US dollars and reiterated a capital budget of US$60 billion to US$64 billion, per Futurum's summary and the call transcript. Wei said "we did not see any sign of slowdown" and declined to give a new figure for TSMC's long-term AI accelerator growth, which it had previously put in the mid-to-high 50s percent a year for 2024 to 2029. His comment: "It probably will be higher."

On packaging, Wei called CoWoS "still the number one choice for our customers", said TSMC is in production at 5x reticle size and working on 9x and 14x, and acknowledged competition. N2 family capacity is expected to more than double in 2027.

The numbers

Why it matters

TSMC is the closest thing the AI industry has to a meter. Nvidia, AMD, Google's TPUs, Amazon's Trainium and most custom accelerators all go through the same fabs and the same CoWoS packaging lines. When TSMC says HPC is 68% of revenue and demand still exceeds supply, it is confirming that the orders behind every hyperscaler capex number are real enough to have been placed with the foundry.

The Q4 guide is the more telling figure. At the midpoint, US$53.3 billion would be about 14% above the third quarter, and that is a quarter in which AMD expects its Data Center segment to grow about 50% sequentially on MI450 and Helios, per its prepared remarks, and Nvidia has begun production shipments of Vera Rubin, per its 10-Q. They are all drawing on the same capacity.

The node mix also matters for anyone predicting prices. 3nm is now TSMC's largest node by revenue, and 2nm is already 8% in its initial ramp. New accelerators on these nodes carry wafer costs that are not coming down, and Wei's comment that N2 capacity will more than double in 2027 is a statement about where next year's GPUs will be made.

What sceptics say

Every supplier near the top of a cycle says demand exceeds supply; that is what a cycle looks like from the inside. The useful question is whether the end customers can pay, and TSMC's numbers cannot answer that: they show chips being ordered, not models being monetised.

Concentration is the second risk. With HPC at 68% of revenue, TSMC's results now depend on a handful of customers whose own demand rests on a small number of very large buyers, some of them paid partly in equity, as AMD's warrant charges show. If those customers cut orders, a foundry with a US$60 billion to US$64 billion capital budget has a lot of fixed cost to carry.

Margins are also starting to give a little. TSMC expects overseas fabs to dilute gross margin by about 2 points this year and 3 to 4 points later, and the 2nm ramp is dilutive in the short term. A 65% to 67% Q4 margin guide is still extraordinary, but the direction is worth watching.

What it means for engineering and DevOps teams

My view: plan 2027 as another supply-constrained year for top-end accelerators. If the foundry, the packaging lines and the two biggest GPU vendors all describe demand above supply, spot availability of the newest parts will be thin and committed contracts will be priced accordingly. Teams that need specific capacity next year should be talking to providers now, not in March.

The flip side is that older generations get cheaper as new ones ramp, and most inference workloads do not need the newest part. Measuring your real throughput per dollar, as described in my guide to benchmarking LLM inference, is how you find out whether a previous-generation GPU at a lower hourly price beats a scarce new one. And since power is the other constraint every supplier mentions, the arithmetic in AI data centre power is worth rereading before you sign anything multi-year.

FAQ

What was TSMC's revenue in Q3 2026?

US$46.72 billion (NT$1,494.24 billion), up 41.2% year on year in US dollars and 16.2% on the previous quarter. Gross margin was 66.4%.

How much of TSMC's revenue comes from AI?

TSMC does not report an AI line. High-performance computing, which includes AI accelerators and data centre CPUs, was 68% of Q3 2026 revenue.

What is TSMC's Q4 2026 guidance?

Revenue of US$52.3 billion to US$54.3 billion, gross margin of 65% to 67% and operating margin of 56.5% to 58.5%, based on an exchange rate of 31.8 NT dollars per US dollar.

Related: AI Power in 2026: 485 TWh, a 0.3 GW Stargate, a 116 GW Turbine Backlog, and 1,800 MW That Dropped Off the Grid in Seconds, How to Benchmark LLM Inference: TTFT, Throughput and Goodput (2026) and GPU Clouds in 2026: H100s at $2.59 an Hour, $104B of CoreWeave Backlog on $35B of Debt, and a Six-Year Bet on Silicon.