Deep Dive — TSMC's record quarter and the repriced order book

TSMC posted the best quarter in its history on Thursday and guided the next one higher, and its American depositary receipts fell anyway. Third-quarter revenue came in at $46.72 billion, up 41.2% year over year and above the top of the company's own guidance range. Net profit rose 68.7% to NT$763.06 billion. The fourth-quarter guide of $52.3 billion to $54.3 billion implies roughly 58% growth at the midpoint against a year earlier. The stock closed at $430.00 the day before the release, down 3.4%, and slid another 2% in after-hours trading as the call went on.

That gap — between what the largest and most leveraged supplier in the AI supply chain reported and what investors were willing to pay for it — is the actual story of this week. Not because TSMC said anything wrong. Because the market has started asking a question TSMC cannot answer from inside a fab.

What the quarter actually said

The numbers, in order of how much they reveal.

High-performance computing was 68% of platform revenue, up 22% quarter over quarter, against 21% for smartphones. AI is no longer a growth segment attached to TSMC's business; it is the base of it. Three-nanometer wafers were 33% of wafer revenue, five-nanometer 28%, and two-nanometer — in its first quarters of volume production — already 8%. Advanced nodes at seven nanometers and below accounted for 79% of wafer revenue.

Operating margin landed at 58.7%, above the top of the company's own 56% to 58% guidance range. That is pricing power stated as a number, and it is the line that matters most for the read-through: a foundry running at the top of its margin band is a foundry whose customers are not negotiating hard. Return on equity was 45.5%. Cash and marketable securities reached NT$4 trillion, about $126 billion. Inventory days fell five to 82, which is the one metric in the report that argues against the inventory-stuffing thesis — you do not ship 16% more wafers quarter over quarter and shorten your inventory days if the demand is being manufactured by your customers' finance departments.

Then there is what the company said about what comes next. Full-year 2026 revenue growth is now expected around 44% in dollar terms, up from at least 40% guided in July. The 2026 capital budget was reaffirmed at $60 billion to $64 billion, which is 13% to 14% of revenue. TSMC says capital spending from 2026 through 2028 will be significantly higher than the three years before it, and it is building nearly 20 fabs at once: 13 leading-edge and advanced-packaging sites in Taiwan, five to six overseas. N2 family capacity more than doubles by 2027. A14 risk production is on track for 2027 with volume in 2028. Advanced packaging is in real-world production at five times reticle size, with nine-times and fourteen-times versions in development — C.C. Wei's claim on the call was that "I don't believe that anybody can do that today except Taiwan Semiconductor."

The thing the market repriced instead

Nothing above is a warning. The warnings were elsewhere, and they arrived in the same 48 hours.

Barclays cut Dell, HPE and NetApp to hold from buy on "enterprise AI timing" and argued that the strong enterprise server revenue the market has been rewarding looks like memory pricing passed through to customers rather than an AI refresh. Dell fell 11%; Oracle touched down 33% for the year. CoreWeave dropped more than 15% with its bond yields near 15% — a level that reads as distress, not growth. Anthropic told pre-IPO investors its third-quarter revenue reached $16.8 billion, up from roughly $11.5 billion in the second quarter, which is about 45% sequential growth against the more than 140% it posted the quarter before, from $4.7 billion in the first quarter to $11.6 billion in the second.

So on the same morning the AI trade's most capital-intensive supplier raised its outlook and said demand exceeded supply, three separate things priced its customers' credit, its customers' revenue growth, and its customers' hardware spending as deteriorating. The supply chain's actual results and the market's expectations have decoupled, and the decoupling has a direction.

Two clocks

The mechanical reason is that TSMC and the public market are reading different dates off the same calendar, and TSMC's date is earlier.

A foundry books revenue when wafers ship, and wafers ship six to nine months after the order that paid for the capacity. A record quarter in October is a report on what the order book looked like in the spring, before Barclays wrote its note, before Anthropic's deceleration was disclosed, and before anyone had a bond yield to argue about. This is not a subtle point and it is not new — it is just the thing that gets forgotten every cycle. When a supplier tells you its backlog is full, it is telling you about decisions its customers already made. It is not telling you about decisions they have not made yet.

The forward-looking part of the report is therefore not the revenue line. It is the capital budget, and the capital budget is where TSMC makes a bet, not a measurement. Reaffirming $60 billion to $64 billion while simultaneously warning that overseas fab margin dilution widens from about 2% this year to 3% to 4% in the next few years is a company saying it will spend more money to earn a thinner margin, on the strength of demand it has assessed rather than demand it has received. That is the sentence to test against the next two quarters of hyperscaler capex guidance. The wafer revenue will be fine either way; it is already committed.

The tell: TSMC is checking the data centers itself

Here is the most interesting thing said on Thursday, and it did not come with a number attached.

TSMC is now sending its own teams to inspect the physical progress of data center construction and the power supply situation in the United States, rather than relying only on what its customers forecast. The feedback, management said, is "so far so good" — the cloud providers have been well prepared. Wei described the company's capacity planning on an earlier call with a line worth carrying: customers give him their demand numbers and he believes they try to tell the truth, but "all the truth together is not a truth." Mark that word, he said.

Understand what that means. The most sophisticated capacity planner in the semiconductor industry, working with the largest customers in the world, with multi-year product roadmaps in hand and visibility no outside analyst has, has concluded that customer forecasts are not sufficient inputs. So it is dispatching people to look at whether the buildings exist and whether the electricity has been arranged.

This is either the strongest available evidence that the AI buildout is real, or the most revealing admission in the sector. Both readings are supported. If you can verify a data center with your own eyes, the demand is not a story. But the fact that verification is necessary means the company has decided that its own customers cannot reliably tell it what they will need — and TSMC is the party that has to live with the answer for a decade, because a fab built for demand that never arrives cannot be un-built.

The direction of the inspection is the interesting part. TSMC is checking power availability, the same constraint we worked through in Power, not GPUs, now sets the pace of AI. A foundry that has to independently verify grid capacity before committing capital has effectively confirmed that the binding constraint on its order book sits outside the semiconductor industry entirely. Wei's capacity planning answer on the call listed the inputs: cloud capex, data centers, power supply, "and also other components." That is a chipmaker describing its demand curve in terms of other industries' engineering schedules.

If you want the mechanics of why the packaging line matters as much as the wafer line here, the reticle limit and hybrid-bonding arithmetic are in AI 101 — What is a chiplet? — the short version is that the constraint on how many AI chips the world gets has moved from how small transistors are to how tightly dies and memory stacks can be assembled.

Where the margin goes

The margin math is worth spelling out, because it is the one part of the story where TSMC's own guidance is bearish.

Gross margin fell 130 basis points sequentially to 66.4%, which the company attributed to the initial ramp of two-nanometer production. Fourth-quarter gross margin is guided down another 40 basis points to about 66%, with the same cause plus dilution from overseas fabs. Overseas dilution is currently around 2% for the full year 2026 and is expected to widen to 3% to 4% over the next few years as more non-Taiwan fabs ramp. Wei's framing was that there is no ceiling on TSMC's profitability, but that a dramatic increase will not happen overnight.

Read the two halves of that together and you get the tension inside the company. Arizona's second fab is now scheduled for volume production in the third quarter of 2027, slightly ahead of the previous plan. A second three-nanometer fab in Tainan starts volume production in the first half of 2028, on top of one already slated for the first half of 2027. Japan's second fab moves to three-nanometer for 2028. Every one of those announcements is capacity the market wants and margin the market does not. The company is trading profitability for resilience, deliberately, and has now put a number on the trade: one to two points of gross margin, permanently, in exchange for not having its entire leading edge in one earthquake zone with a geopolitical question attached.

What the skeptics have right

The bear case does not require TSMC to be lying, and it is stronger when it doesn't accuse anyone of that.

It says that a record quarter this deep into a cycle is normal, that the cadence has been six consecutive records, and that the marginal buyer of AI infrastructure has shifted from equity to credit — where the price signals are faster and less forgiving. CoreWeave's bonds telling you something at 15% that its equity told you at a lower number six months ago is not a contradiction; it is the same information arriving through a channel that reprices daily rather than quarterly. Anthropic growing 45% sequentially off an $11.5 billion base is an extraordinary business and a deceleration at the same time, and both facts matter to a company about to ask public investors for a valuation built on the extrapolation, not the level.

The strongest form of the skeptical case is about the lag itself. If TSMC's revenue is a spring order book reported in autumn, then the record quarter is not evidence about the future at all — it is evidence about a past that has already been priced. The only forward-looking numbers in the release are the capital budget and the Q4 guide, and the capital budget is a judgment call made by people who just told you they cannot fully trust their own customers' forecasts. That is not a bearish argument. It is an argument that the report contains less information than its headline suggests, in both directions.

What cuts against the pessimists: the operating margin beat, the inventory days coming down, and the fact that HPC grew 22% sequentially on a base that was already 66% of revenue — that is not a segment being propped up. And the demand mix is broadening in a way that matters. On the call, analysts pressed on whether AI accelerator demand is shifting toward agentic workloads, which pulls in CPUs and networking switches alongside GPUs. If that shift is real, TSMC's addressable content per AI dollar goes up, not down, because agentic inference is less concentrated in one accelerator and more spread across the rack.

What to watch

Three things will settle this, and none of them are TSMC's next print.

First, hyperscaler capex guidance for 2027 in the coming earnings round. That is the input TSMC's $60 billion to $64 billion is a bet on, and it is the number that converts or kills the "demand is higher than supply" claim.

Second, whether overseas fab dilution stays at the bottom of the 3% to 4% range or the top. That range is a permanent tax on gross margin that TSMC has chosen to accept, and the market has not fully repriced for it.

Third, CoreWeave's debt. If AI infrastructure financing reprices further in credit before it reprices in TSMC's order book, the gap between the two clocks widens, and the supplier's record quarters stop being read as good news at all. We covered the memory-shortage pricing story behind the same repricing wave in TSMC's sixth straight record quarter — 58.7% margin, 68% from HPC.

The honest read of Thursday is narrower than either camp wants. TSMC executed flawlessly on demand it had already booked, told investors it sees no slowdown, and watched its stock fall — and the fall was rational, because the company's own numbers describe the spring while everyone else is trying to price 2028.

If a record quarter with a raised outlook can't move the stock up, what would it take — and is that a broken market or a correct one? Tell us in the comments.