Summary

  • Taiwan Semiconductor Manufacturing Company, Ltd. reports revenue, margin and guidance with precision, but the materials reviewed contain no company-sourced figure for advanced packaging capacity, utilisation, lead time or allocation — the step that most plausibly limits how many AI accelerators can be completed.
  • The CoWoS capacity series now used to plan around that step (roughly 35,000 wafers per month at end-2024, roughly 75,000 at end-2025, a 125,000 to 130,000 target for end-2026) originates with TrendForce and reached readers through trade press, not through a company filing.
  • A dated company announcement of a Chiayi Science Park packaging facility and a trade report on Kumamoto packaging work describe intended capability. Neither establishes installed, qualified or available throughput.
  • On this record advanced packaging is a candidate constraint, not a disclosed bottleneck. The disclosure gap is the finding, and it is the thing worth pricing.

What the company does publish

Start with what is not in dispute. The company's own third-quarter 2026 results page reports net revenue of US$46.72 billion against prior guidance of US$44.6 to 45.8 billion, a gross margin of 66.4% and an operating margin of 58.7%. It guides the fourth quarter to US$52.3 to 54.3 billion of revenue, with gross margin of 65.0% to 67.0% and operating margin of 56.5% to 58.5% (TSMC 2026 Q3 results).

These are good numbers, and they are good numbers of a particular kind. They describe the commercial outcome of a quarter that has already been produced and sold. They are the residue of a manufacturing system, not a description of it. A revenue line tells you what customers paid. A gross margin tells you the spread between that revenue and the cost of producing it. Neither tells you which production step was the binding one, and neither tells you how much headroom exists at that step for the quarter after next.

This is not a criticism of the disclosure. It is what an income statement is for. A foundry reports by technology node because that is how it prices and how its customers buy; it reports by platform because that is how demand is organised. The same page, as read, carried no packaging capacity figure, no packaging utilisation figure, no split of capital expenditure between wafer fabrication and packaging, and no largest-customer percentage of net revenue. That last omission matters for a different reason and I will return to it.

The company's investor-relations site does maintain an archive of annual reports and Form 20-F filings (TSMC annual reports), and those filings are locatable through the U.S. Securities and Exchange Commission's EDGAR system under CIK 1046179 (SEC EDGAR, CIK 1046179). Both were reached. Both returned index pages rather than filing bodies. The archive exists; the passages that would settle the packaging question were not obtained as text. That is a boundary of this reading, stated plainly, and not a claim about what the filings contain.

Four kinds of claim, kept apart

Most of the confusion in this subject comes from a single mechanical failure: collapsing claims of different provenance into one sentence, so that an estimate acquires the authority of a filing by losing a clause. Four kinds of claim are in circulation, and they are not interchangeable.

Company-reported results and guidance. Dated, issuer-sourced, produced under disclosure obligations. The revenue, margin and guidance figures above are of this kind. They are the strongest material available, and they say nothing about packaging.

Third-party capacity estimates relayed by trade press. Quantified, specific, and originating outside the company. The CoWoS series is of this kind. Trade press reported on 11 July 2026 that TrendForce estimates CoWoS capacity at roughly 35,000 wafers per month at end-2024, roughly 75,000 at end-2025, and a target of 125,000 to 130,000 by end-2026, describing growth of roughly 80% a year (TechTimes, 11 July 2026). Note precisely what that sentence attributes, and to whom. The number belongs to TrendForce. The transmission belongs to the trade outlet. Neither belongs to the company.

Trade-press-attributed executive remarks. The same report attributes to CEO C.C. Wei, speaking at the 4 June 2026 annual shareholders meeting, the statement that CoWoS capacity is extremely tight and sold out through 2026. This is a real signal and a weak instrument. It is a quotation at one remove, no primary transcript or recording was obtained, and the remark is qualitative: it carries no number. "Sold out" describes bookings against a period, not throughput achieved within it. A year can be sold out and still ship below plan.

Dated company announcements. A press notice establishes that a company said it would build something, on a date. It does not establish that the thing is built, qualified, yielding or incremental to an existing estimate. Announcements are commitments, and commitments are not capacity.

To these four I add a fifth, which is mine and should be labelled as such: inference about mechanism. When I say that a packaging line is a physically separate asset from a wafer fab, with its own tools, its own qualification cycle and its own yield ramp, that is a description of how this class of manufacturing works, not a disclosure by anyone. It is the kind of claim that is usually true and always worth separating from the others.

What the capacity series measures

The TrendForce series is quoted in wafers per month. That unit deserves attention, because at least three different quantities get discussed as "capacity" and they are not the same thing.

Wafer starts at a node measure how many wafers begin fabrication. That is the number a fab manager schedules and the number that appears when a new fab is announced. Packaged units measure how many finished assemblies leave a packaging and test line. Those are different counts, related by yield and by how many die go into each package. And CoWoS-class capacity, as estimated, measures a specific intermediate capability — the interposer and bonding throughput that sits between the two.

An 80% annual growth rate is fast in absolute terms. Precisely because it is fast, the 125,000 to 130,000 figure for end-2026 should be read as a plan with a date attached, not as an observation. Plans of that shape depend on tool delivery, cleanroom availability, substrate supply, customer qualification schedules and hiring, and they slip for ordinary reasons that have nothing to do with anyone's intentions.

There is a further gap. TrendForce's methodology, denominator, geographic scope and product definition were not obtained. It is therefore not established whether the series describes this company specifically, the contract-packaging market as a whole, or a defined subset of CoWoS variants. A reader who treats the number as a company figure inherits an error that no subsequent arithmetic can repair. A reader who treats it as a market estimate, dated and attributed, is using it correctly.

The capital-expenditure figure has the same shape of problem. Trade press reports a 2026 capital budget of US$52 to 56 billion, roughly 37% above 2025 levels, with management indicating an intent to spend at the high end of the range. That is a large number and a real commitment. It is also aggregate. A dollar of capex does not say which step it buys, and a budget line that funds wafer capacity, packaging capacity, land, buildings and tool prepayments in one total cannot be decomposed by the reader.

The inference that a rising capex budget must be relieving the packaging constraint is plausible, unfalsifiable from this evidence, and therefore not usable for scheduling.

The mix disclosures have the same character. Trade press reports first-quarter 2026 figures of 74% of wafer revenue from 7nm and below, and 61% of total revenue from high-performance computing. Both are useful. Neither is a packaging measure. The first says where revenue comes from by transistor generation; the second says who buys. Packaging sits across both, and a shift in either can move the packaging load in ways the mix numbers do not reveal.

The step between die and system

Strip the subject back to its mechanics and the reason for the attention becomes clear. A leading-edge logic die is not a product. It becomes one after it is bonded to memory and an interposer or substrate, encapsulated, tested and binned. Advanced packaging is that step, and it is where a wafer becomes something a server can accept.

The structural point is simple and does not depend on any disputed number: a packaging line is a separate physical asset from a wafer fab. It has different tools — bonders, inspection, molding, test — a different cleanroom envelope, a different qualification cycle with each customer, and its own yield curve during ramp. A wafer fab coming online does not by itself add packaging throughput, and a packaging line coming online does not by itself add die supply. They are two gates in series, and the system's output is set by whichever is narrower.

That is why the disclosure question has commercial weight. If both gates were reported, a buyer could compare them and see which one binds. Because only one is reported with any precision, the comparison is made outside the company, by parties who must estimate one side of it. The estimate may be excellent. It is still an estimate, and its error is invisible to the person using it.

Chiayi and Kumamoto: intent, not output

Two dated items are usually offered as evidence that the packaging gate is being widened. Both are real. Neither says what it is asked to say.

On 18 March 2024 the company announced an advanced packaging facility in Chiayi Science Park to support CoWoS demand (TSMC announcement, 18 March 2024). That is a fact about a decision, dated and attributable. It is not a throughput figure, a tool move-in date, a qualification date or a start of volume production. Two and a half years have passed since the announcement; whether the site is now producing, qualifying or still fitting out is not established by anything in the record reviewed.

On 7 August 2026 trade press reported that the company is preparing to add 3nm production in Kumamoto and is deepening local advanced packaging capability in Japan, alongside chip design and next-generation image sensor work (Digitimes, 7 August 2026). Read the verbs. "Preparing to add" and "deepening" describe activity and intent. They do not describe installed capacity, and they do not say whether any packaging work in Japan is incremental to the TrendForce series or already assumed inside it.

The second item is more interesting than the first, for a reason that has nothing to do with its numbers. Wafer capacity is dispersing. New fabs outside Taiwan are a documented, much-discussed programme. If packaging remains concentrated while wafer fabrication disperses, then a die fabricated in Arizona or Kumamoto may still cross the Pacific before it becomes a shippable accelerator. That is a conditional statement, not a finding, because the record reviewed does not establish where the packaging happens for output from those fabs.

It is, however, the specific question a supply-chain planner should be asking, and it is answerable by disclosure in a way that a capacity estimate is not.

What would move this from inference to measurement

It is worth being concrete about what would settle the question, because the list is short and every item on it is ordinary disclosure rather than a demand for a trade secret.

A packaging capacity figure with a date and a unit. A utilisation figure, or a statement of constraint by name. A capital-expenditure split that separates wafer fabrication from advanced packaging. A lead time, a backlog figure, or an explicit allocation rule for constrained output. A qualified start-of-volume date for Chiayi. A statement of which named sites perform advanced packaging and at what scope. TrendForce's methodology, denominator and product definition. Any customer commitment structure — prepayment, take-or-pay, reservation — disclosed by either side.

Not all of these are equally likely, and none is owed to a journalist. The point is that the current uncertainty is not a mystery. It is a gap in a specific, enumerable set of disclosures, and it could close.

One clarification about the shape of the gap. It is tempting to say that packaging is sold bundled with wafer fabrication and therefore invisible in revenue. That may be true. It is not established by the materials reviewed, and I am not asserting it. What is established is narrower and sufficient: the mix disclosures in circulation measure node and platform, and neither is a packaging measure.

Who bears the cost of an unmeasured step

Here the analysis has to stay conditional, because the evidence does not support more.

Suppose a datacenter operator or server vendor commits downstream capital — a power contract, a building schedule, a launch date, a customer commitment of its own — against an accelerator delivery date. Suppose that date was modelled from wafer capacity, node mix or aggregate capex rather than from packaging throughput. Then the exposure is timing and option value: capital that sits idle or delayed, expediting expense, a missed window that cannot be recovered by working harder later.

Exposure is not loss. A buyer with a conservative buffer and a diversified supplier set may absorb a packaging constraint without ever recording an incident. A buyer who assumed the reported numbers described the whole system may not. The difference between those two buyers is not intelligence; it is whether they treated an estimate as an estimate.

What about the supplier? Control of a scarce, unreported step plausibly confers informational and bargaining advantage: the party that can see the queue can price and allocate against it. But advantage is not the same as cost transfer. Establishing that buyers bear the cost would require evidence of commitment terms, allocation rules, price premiums or realised delays, and none of that was obtained. On this record, cost incidence is undetermined. Saying so is more useful than picking a villain.

The disclosure gap as the finding

The honest conclusion is narrower than the headline the subject usually attracts, and more useful.

It is not established that advanced packaging is the binding constraint on AI accelerator output. It is not established that it is not. What is established is this: the step that most plausibly gates that output is not quantified in the materials reviewed, while the figures most widely used to plan around it come from a third party, are dated, are attributed, and carry a methodology nobody outside that third party has inspected.

That is a fact about the information supply chain, and it has a practical consequence. An organisation that schedules capital against accelerator delivery should know which of its inputs is measured and which is estimated, and should hold a buffer against the estimated ones. The buffer is not pessimism. It is the price of an unmeasured step, and it is cheaper than the alternative.

There is also a reason for restraint in the other direction. Operators disclose by segment and platform because that is how they are managed and how they report; a packaging line is a process step inside a segment, not a business unit. Its absence from the disclosure taxonomy is unremarkable on its own terms. It becomes notable only because the step has acquired systemic importance that the taxonomy was not built to express. That is a mismatch between an old reporting structure and a new dependency, not evidence of concealment, and the distinction matters if the goal is to get the disclosure rather than to score a point.

Evidence boundaries

Recorded as of 17 October 2026, so that later readers can see what this piece did and did not have.

  • No verbatim text from the company's Form 20-F, quarterly management report or earnings transcript quantifying packaging capacity, utilisation, lead time or allocation was obtained. The investor-relations annual-report index and the SEC EDGAR filing page were reached as index pages only.
  • No capital-expenditure split between wafer fabrication and advanced packaging was obtained. Aggregate capex does not identify the step it funds.
  • No largest-customer percentage of net revenue was read verbatim from the filed Form 20-F. A hedged secondary summary circulated; it was not relied on.
  • TrendForce's methodology, denominator, geographic scope and product definition were not obtained, so the series' scope — company-specific, market-wide or a defined CoWoS subset — is unknown.
  • No site-level status for Chiayi or Kumamoto was obtained. Operational, qualified, volume-producing and incremental are all unestablished.
  • No buyer commitment, take-or-pay term, penalty, price premium or realised delay evidence was obtained. Cost incidence is undetermined.
  • The statement attributed to the chief executive at the 4 June 2026 annual shareholders meeting is available through trade-press attribution only.

Sources