Which Node Runs at TSMC Fab 12, 14 and 15? Four Public Answers, One Filing

Ask a simple question: which process node runs at TSMC's Fab 12 in Hsinchu, Fab 14 in Tainan, and Fab 15 in Taichung? These are three of the company's largest 300 mm sites, and the answer decides how much TSMC capacity you book to 7 nm, 5 nm and 3 nm when you build a supply model.

The public record gives you four answers. They do not agree with each other. And none of them agrees with the one document that is legally required to say.

This is the first of a monthly series in which we take a number the industry circulates, go to the primary source, and publish what we find, including what it says about our own data. This month the data that was furthest off was ours.

Four public answers

Three trackers, three different stories, and ours was the one that placed all three fabs at the leading edge. Our registry carried the attribution under TSMC's name with a confidence label of "Confirmed," and for Fab 15 it added a note that "Fab 15B has been fully converted to N3." That note cited nothing. It was added in an April refresh and never re-sourced.

What the filing says

TSMC's annual report on Form 20-F, Item 4, carries a table of every wafer fab in operation with a column headed "the most advanced technology for volume production," footnoted "in nanometers, as of year-end." For fiscal 2025, filed with the SEC on 16 April 2026, the three fabs read 40, 16 and 7 [1].

The obvious objection is that the column might be stale, a line nobody at TSMC updates. So we pulled the same table from the last eight filings [2].

The column is maintained. Fab 18 was updated the year N3 arrived. Fab 20, 21 and 22 were added the year they started. Fab 12 was walked through 7, 5 and 3 as each node was brought up there, then restated to 40 in the FY2024 filing and held there. Fab 14 has never been listed below 16 nm, and Fab 15 never below 7 nm, in eight years of filings.

That last sentence is the finding. A fab that TSMC has described as a 7 nm fab in every annual report since 2018 was carrying 3 nm capacity in our model, and 20 nm capacity in someone else's.

What it does to the totals

We publish TSMC capacity two ways: a node-level allocation record with a stated 2026 target, and a fab-by-fab registry that rolls up to each node. The two are supposed to reconcile. Rebuilding the roll-up with the filed ceilings, and leaving every fab's total capacity untouched, changes the picture more than we expected.

Three things fall out of that table.

The 3 nm roll-up reconciles exactly the moment Fab 15's unsourced N3 slice is removed. The residual 10% that our reconciliation test had been tolerating was entirely an attribution TSMC does not recognise.

The 5 nm "perfect match" was an artifact. Eighty of the 230 kWSPM we had booked to 5 nm sat at Fab 12, 14 and 15, where TSMC says no 5 nm runs in volume. What remains at Fab 18 and Fab 21 is 150. If the 230 target is right, the gap is real capacity somewhere we have not sourced, most plausibly under-counted phases at Fab 18. We have not filled it in, because filling it in would be inventing per-fab data to make a total add up, which is how the original error was made.

The 7 nm roll-up flips from a 48% overshoot to a 24% shortfall. Fab 12 and Fab 14 do not run 7 nm in volume, so what remains is Fab 15 alone at 130 kWSPM, below the 208–239 kWSPM that TSMC's own 2Q26 revenue mix implies at normal utilisation. So either Fab 15's total is well above 130, or 7 nm also runs somewhere with a lower ceiling. Both are possible. Neither is sourced.

What we changed, and what the filing cannot settle

On 12 September we corrected the attribution to the filing and left the totals alone. Fab 12 is now a 40 nm fab in our registry, Fab 14 a 16 nm fab, Fab 15 a 7 nm fab, each stamped "Projected" rather than "Confirmed," because the node ceiling is TSMC's but the capacity is our estimate. The node records for 3 nm, 5 nm, 7 nm and mature were re-pointed to the fabs that actually carry them. Twenty-nine correction entries went onto the public corrections trail with the filing as the cited source, and the affected capacity snapshots were unpublished and re-derived rather than edited, so the old values remain visible as superseded [6]. The reconciliation notes on each node record now say where the roll-up diverges from the target and why [7].

The filing settles less than it looks like it settles, and we should be exact about that.

It gives a ceiling, not a mix. "Most advanced technology for volume production" means Fab 15 runs 7 nm and nothing more advanced; it does not say how much of Fab 15 is 7 nm rather than 28 nm. TSMC has not published per-fab capacity since its FY2014 filing, so every per-fab number in circulation, ours included, is an estimate.

It does not explain Fab 12. Public build history has 28 nm at Phases 5 and 6 and 16 nm at Phase 7, yet two consecutive filings say 40, and the executive biographies in the same document still name a Fab 12B operations head. TSMC evidently does not count what the advanced half does as volume production. We record 40 as filed and say that we do not know why.

It describes year-end. A May 2026 report that Fab 15A is being converted from 28/22 nm to 4 nm [5] may well be right; if so, it will appear in the FY2026 filing, and we will update the ceiling when it does, not before.

Our best reading is that the column is doing exactly what its footnote says and nothing more. Fab 12 is where TSMC has historically brought up each new node before moving volume to a GIGAFAB, and the 7, 5, 3 sequence in the FY2018 to FY2023 filings tracks that role rather than the site's installed capacity. Once N3 volume moved to Fab 18, what remained at Fab 12 in "volume production" by TSMC's own threshold was the mature half of the site, and the column was written down to match. On that reading the 40 is a definition, not a downgrade, and the 16 nm and 28 nm phases are real but below whatever bar TSMC applies to the word "volume". It is a reading, not a finding. The FY2026 filing is the first test: if Fab 12 stays at 40 while Fab 20 and 22 carry 2, the definition story holds. The second test is cheaper and available now: TSMC's own fab descriptions, and any reported equipment moves at Fab 12B, would show whether the advanced lines are still there or have been re-tooled.

If you model TSMC supply

The practical lesson is not "trust TSMC's table." It is that the node attribution of individual fabs is the least-sourced number in most supply models, and the one with the largest leverage on node-level totals. A single unsourced note moved our 3 nm figure by 10% and our 7 nm figure by 72 percentage points. Before you rely on any fab-level node split, ask where it came from. If the answer is "everyone says so," it came from nowhere.

From the buying side the useful question is not "how much 3 nm does TSMC have" but "how much 3 nm is at the one site where 3 nm actually runs." On the filing that is Fab 18, and only Fab 18, until the FY2026 report says otherwise. A buyer who had Fab 15 down as a 3 nm site was carrying a second-source assumption that did not exist, and any contingency built on it was a contingency on paper. The same goes for the 80 kWSPM of 5 nm that our model, and probably others, had parked at fabs where TSMC says no 5 nm runs in volume: if that capacity is real it sits somewhere unsourced, and if it is not, the 5 nm market is tighter than the headline totals suggest. The practical change is small and cheap. Ask which fab and which phase an allocation commitment is written against. Read the counterparty's own filing before the meeting, not the tracker that summarises it. And treat a node that lives at one site as single-site risk, however large that site is.

Next month: Kumamoto. TSMC's filing lists Fab 23 at 28 nm; our registry, and most public descriptions, carry 12/16 nm there. The same column, the same question.