TSMC Q3 Earnings: Record AI Quarter Tests Capacity Limits

A record that was already announced

Taiwan Semiconductor Manufacturing Co. reports third-quarter results on October 15, but most of the headline number is already known. On October 8 the company disclosed record July-to-September revenue of NT$1.49 trillion, roughly $46.71 billion. That is about 50% higher than a year earlier and above the NT$1.46 trillion that an LSEG SmartEstimate drawn from 19 analysts had expected.12

The open question is profit. Forecasts cluster around a gain of about two-thirds. One LSEG SmartEstimate puts net profit near NT$740.8 billion, up 64% year over year.1 Another projection puts it at NT$746.3 billion, up 65%, or about $23.4 billion.2 The gap between the two is small. Either figure would extend what is shaping up as a sixth consecutive record quarter.1

Why the market watches TSMC before Nvidia

TSMC fabricates the physical silicon for Nvidia and AMD. That makes its results an early signal for AI hardware demand. Investors get it weeks before Nvidia reports in November.12 The signal so far points to demand outrunning supply. TSMC's advanced CoWoS packaging, a bottleneck for high-end AI accelerators, is reportedly booked solid into 2027.1

This quarter extends a pattern from the summer. In July, TSMC reported record second-quarter revenue of $40.2 billion. It raised its full-year growth forecast to slightly more than 40%, up from its earlier call of 30%-plus.3 Chair C.C. Wei said then that the company's "conviction in the multi-year AI megatrend" remained very high.3 Wall Street agrees for now. Analysts give TSMC's U.S.-listed shares a Strong Buy consensus with an average price target of $883.2

The capex figures don't line up

Capital spending is where the numbers disagree. One account says 2026 capex is on track for $56 billion.1 A report on an earlier earnings call describes Wei raising annual spending from $52 billion to $56 billion to add N3 capacity.5 TSMC's July update went further, guiding to $60 billion to $64 billion for 2026. That is at least $4 billion above its previous forecast.3 A Taiwanese equipment supplier likewise refers to TSMC's AI-driven capex as roughly $60 billion.4

The likeliest explanation is that the figures come from different points in the year. The $56 billion figure appears to have been superseded by the July increase. The direction is not in doubt, though. TSMC has raised its spending plans more than once in 2026, and each raise has been tied to AI demand.

Arizona: ambitious, but not quick

Arizona is central to the expansion. Alongside the July guidance, TSMC announced an additional $100 billion investment in the state.3 Wei has said the second Arizona fab will make 3-nanometer chips. Construction of that building is complete, but volume production is not slated until the second half of 2027.5

That timeline matters. The U.S. footprint is growing, but output from it trails demand being felt right now. Analysts disagree on what comes next. Morgan Stanley points to space constraints in Arizona and argues TSMC may need to move quickly on Texas. Morningstar sees Texas as a longer-term goal.2 Either way, new U.S. capacity is a 2027-and-beyond story. That is roughly the same horizon as the current CoWoS backlog.15

Spending ripples through the supply chain

The effects already show up among TSMC's suppliers. Tachia Yung Ho Machine Industry says its cleanroom business is running near 100% utilization because of TSMC's higher capex.4 Its third-quarter revenue rose 54.6% year over year to NT$530 million, its strongest quarter in nearly two years. Its subsidiary Okawa Research expects secondary hookup engineering work to grow several-fold next year.4 Figures like these are small next to TSMC's, but they show the build-out is being contracted and paid for, not just announced.

The reading

Thursday's report will very likely confirm another record. The more important parts are the profit margin and the tone of TSMC's guidance. Those will indicate whether the AI build-out is still accelerating or merely holding steady.

The main constraint on TSMC is capacity, not demand. Arizona headlines and rising capex show the company's response. The 2027 production dates and packaging backlog show how long that response will take. For Nvidia and its peers, the practical limit on AI chip supply over the next year or so is how fast TSMC can build, more than how much customers want to buy.