TSMC reported third-quarter revenue of NT$1.49 trillion, roughly $48.4 billion, up 51% year over year — beating the NT$1.46 trillion analysts had expected, according to Bloomberg. The figure, disclosed October 8, lands amid persistent warnings that AI spending has outrun reality. The silicon underneath that spending tells a different story.

September alone generated NT$511.86 billion, up 54.6% from a year earlier and just short of August's all-time monthly record of NT$514.8 billion. Nvidia and Apple remain TSMC's two largest customers, and neither has shown any sign of pulling back orders.

Where the AI buildout is actually solid

For months, warnings about an AI bubble have piled up alongside genuine supply strain in memory chips, and the two narratives have blurred together. TSMC's quarter separates them. The logic chips doing the actual computing are not short of buyers. The memory sitting next to them is a different matter.

Samsung and SK Hynix have spent 2026 warning that DRAM and NAND supply is running short into 2027. TrendForce expects conventional DRAM contract prices to rise 10% to 15% sequentially in the fourth quarter, with NAND flash up 15% to 20%, after much sharper increases earlier in the year. Industry reports indicate major memory makers' DRAM and HBM capacity for 2027 is already booked.

SK Group chairman Chey Tae-won has said the squeeze could run through 2030, as Samsung, SK Hynix and Micron shift more output toward high-bandwidth memory for AI accelerators at the expense of conventional supply.

Why TSMC's capacity problem is a backlog, not a crisis

TSMC operates on a different axis. It builds the logic dies — the processors themselves — not the memory stacked beside them. Its capacity situation looks less like a shortage and more like a queue of customers willing to pay more for a place in line.

Industry reports say TSMC's most advanced capacity is heavily booked into 2028, with early 2-nanometer wafers priced around $30,000. Apple has reportedly secured roughly half of early 2nm output for chips destined for the iPhone 18 generation. Nvidia has taken about 60% of TSMC's CoWoS advanced packaging capacity — the lines that bind compute dies to high-bandwidth memory for AI accelerators — and has overtaken Apple as TSMC's single largest customer.

The contrast in business models is stark:

• Memory makers chased HBM margins and left conventional DRAM and NAND supply short.
• TSMC raises leading-edge prices because it can.

What does the pricing power signal?

That pricing power is the core of the story. TSMC controls roughly 72.5% of the global foundry market as of the second quarter. Its 5nm, 4nm and 3nm lines are fully booked, giving it no reason to compete on price even as questions swirl about whether AI capital spending is sustainable.

A chipmaker worried about a bubble in its customers' spending would defend share, not push leading-edge prices higher. TSMC did the opposite. That behavior signals it sees years of booked demand ahead, not a slowdown.

None of this settles the broader argument over whether hyperscalers are overbuilding data centers, or whether the financing behind that buildout is sound. Those are real questions, and TSMC's results do not answer them.

What the quarter does answer is narrower and more useful: the physical chips that run AI workloads are still selling as fast as TSMC can make them, at higher prices than last year, to the same two customers that have driven this cycle since it started.

TSMC will release fuller Q3 detail and its Q4 outlook at an investor call on October 15. Until then, the September numbers stand as the clearest read yet that whatever is happening to AI valuations on paper, the hardware underneath them has not slowed down.