The foundry every AI chip is built on
The only pure-play foundry at the leading edge at scale — it manufactures the chips it does not design for NVIDIA, AMD, Broadcom and Apple alike, and owns the CoWoS packaging that governs how many AI accelerators ship. The chokepoint the entire complex depends on.
| Date |
|---|
| Type |
|---|
| What happened |
|---|
| Source |
|---|
| 2026-09-05 | editorial note | Capex figure revised: $52 → $846.76BCapex moved from $52 (deep-dive-2026-07-24-q3-guide-verification.md) to $846.76B (deep-dive-2026-09-05-refresh.md). | dossier |
| 2026-09-05 | editorial note | Revenue figure revised: 40% → $1,270.38 billionRevenue moved from 40% (deep-dive-2026-07-24-q3-guide-verification.md) to $1,270.38 billion (deep-dive-2026-09-05-refresh.md). | dossier |
| 2026-09-05 | editorial note | Verdict changed: (no verdict) → The Q2 beat was not a peak — July revenue accelerated to +44.7% YoY, the board appropriated a single-tranche US$29.4B of capacity capital, and H1 capex now clears the …Before (deep-dive-2026-07-24-q3-guide-verification.md): (no verdict) After (deep-dive-2026-09-05-refresh.md): The Q2 beat was not a peak — July revenue accelerated to +44.7% YoY, the board appropriated a single-tranche US$29.4B of capacity capital, and H1 capex now clears the primary bar at US$26.8B. The thesis is unchanged and better evidenced; the only thing that moved against it is the tape, which is ~2% below the 2026-07-09 price after a beat. | dossier |
| 2026-07-24 | editorial note | Capex figure revised: $15.70B → $52Capex moved from $15.70B (deep-dive-2026-07-16-reaction.md) to $52 (deep-dive-2026-07-24-q3-guide-verification.md). | dossier |
| 2026-07-24 | editorial note | Margin figure revised: 67.7% → 65%Margin moved from 67.7% (deep-dive-2026-07-16-reaction.md) to 65% (deep-dive-2026-07-24-q3-guide-verification.md). | dossier |
| 2026-07-24 | editorial note | Revenue figure revised: $40.20B → 40%Revenue moved from $40.20B (deep-dive-2026-07-16-reaction.md) to 40% (deep-dive-2026-07-24-q3-guide-verification.md). | dossier |
| 2026-07-16 | editorial note | Capex figure revised: $52 → $15.70BCapex moved from $52 (deep-dive-2026-07-14.md) to $15.70B (deep-dive-2026-07-16-reaction.md). | dossier |
| 2026-07-16 | editorial note | Revenue figure revised: $39.6B → $40.20BRevenue moved from $39.6B (deep-dive-2026-07-14.md) to $40.20B (deep-dive-2026-07-16-reaction.md). | dossier |
| 2026-07-14 | editorial note | Capex figure revised: $40.9B → $52Capex moved from $40.9B (deep-dive-2026-07-10.md) to $52 (deep-dive-2026-07-14.md). | dossier |
| 2026-07-14 | editorial note | Revenue figure revised: $121,423M → $39.6BRevenue moved from $121,423M (deep-dive-2026-07-10.md) to $39.6B (deep-dive-2026-07-14.md). | dossier |
| 2026-07-14 | editorial note | Verdict changed: The AI complex's single foundry + CoWoS chokepoint just printed a hard 20-F — FY2025 rev $121.4B, 59.9% GM, 50.8% op margin, $54.1B net income, $55B net cash — and the filing DE-RISK…Before (deep-dive-2026-07-10.md): The AI complex's single foundry + CoWoS chokepoint just printed a hard 20-F — FY2025 rev $121.4B, 59.9% GM, 50.8% op margin, $54.1B net income, $55B net cash — and the filing DE-RISKS the two loudest bear points: top-2 customers are only 36% of revenue (not 40%+), and its first overseas fab (Arizona) already turned a $514M profit. The only first-order risk stays exogenous (Taiwan), and the Jan-2026 US–Taiwan deal quietly converts TSMC's Arizona spend into tariff immunity. Thesis intact; BULLISH / HIGH / 1Y+ — with one honest downgrade to the "cheapest AI name" line (NVIDIA has de-rated to multiple parity). After (deep-dive-2026-07-14.md): (no verdict) | dossier |
| 2026-07-10 | editorial note | Capex figure revised: $52 → $40.9BCapex moved from $52 (deep-dive-2026-06-09.md) to $40.9B (deep-dive-2026-07-10.md). | dossier |
| 2026-07-10 | editorial note | Margin figure revised: 66.2% → 59.9%Margin moved from 66.2% (deep-dive-2026-06-09.md) to 59.9% (deep-dive-2026-07-10.md). | dossier |
| 2026-07-10 | editorial note | Revenue figure revised: $35.90B → $121,423MRevenue moved from $35.90B (deep-dive-2026-06-09.md) to $121,423M (deep-dive-2026-07-10.md). | dossier |
| 2026-07-10 | editorial note | Verdict changed: The chokepoint that captures the foundry + CoWoS-packaging rent for the ENTIRE AI complex (NVIDIA, AMD, Broadcom, Apple all depend on it), trading at ~22× forward — BELOW every custo…Before (deep-dive-2026-06-09.md): The chokepoint that captures the foundry + CoWoS-packaging rent for the ENTIRE AI complex (NVIDIA, AMD, Broadcom, Apple all depend on it), trading at ~22× forward — BELOW every customer it supplies — with the only real risk exogenous (Taiwan geopolitics, the source of the discount). The cleanest, highest-quality, lowest-multiple way to own AI infrastructure. BULLISH / HIGH / 1Y+. After (deep-dive-2026-07-10.md): The AI complex's single foundry + CoWoS chokepoint just printed a hard 20-F — FY2025 rev $121.4B, 59.9% GM, 50.8% op margin, $54.1B net income, $55B net cash — and the filing DE-RISKS the two loudest bear points: top-2 customers are only 36% of revenue (not 40%+), and its first overseas fab (Arizona) already turned a $514M profit. The only first-order risk stays exogenous (Taiwan), and the Jan-2026 US–Taiwan deal quietly converts TSMC's Arizona spend into tariff immunity. Thesis intact; BULLISH / HIGH / 1Y+ — with one honest downgrade to the "cheapest AI name" line (NVIDIA has de-rated to multiple parity). | dossier |
The verdict
The Q2 beat was not a peak — July revenue accelerated to +44.7% YoY, the board appropriated a single-tranche US$29.4B of capacity capital, and H1 capex now clears the primary bar at US$26.8B. The thesis is unchanged and better evidenced; the only thing that moved against it is the tape, which is ~2% below the 2026-07-09 price after a beat.
Primary sources
SEC filings
Source documents — open to read in full
July revenue accelerated, hard. NT$467.58B, +5.6% MoM and +44.7% YoY; Jan–Jul cumulative
NT$2,872.06B, +37.0% YoY [primary: SEC EDGAR 6-K, accession 0001046179-26-000471, filed
2026-08-10, "TSMC July 2026 Revenue Report"]. July alone ran +10.4% above the Q2 monthly
average of NT$423.46B.
The board appropriated US$29,442.50 million of capacity capital in one resolution — the
first hard, TSMC-filed Capital expenditureMoney spent on long-lived things — buildings, machines, servers — rather than on running costs. dollar figure this shelf has ever held. Purposes stated verbatim:
"1) Installation and upgrade of advanced technology capacity; 2) Installation and upgrade of
advanced packaging, mature and/or specialty technology capacity; 3) Fab construction, and
installation of fab facility systems" [primary: 6-K accession 0001046179-26-000536, filed
2026-08-11, "TSMC Board of Directors Meeting Resolutions"]. This is an appropriation, not FY
capex guidance — see the provenance warning in §"The Capital expenditureMoney spent on long-lived things — buildings, machines, servers — rather than on running costs. figure, finally, and what it is not".
H1 2026 capex now clears the primary bar: NT$846.76B ≈ US$26.80B, against H1 2025's
NT$628.05B ≈ US$19.88B — +34.8% YoY [primary: 6-K accession 0001046179-26-000541, filed
2026-08-14, Exhibit 99.1 "Consolidated Financial Statements for the Six Months Ended June 30,
2026 … pursuant to Taiwan-IFRSs", consolidated statements of cash flows, "Acquisitions of:
Property, plant and equipment"; USD at the Q2 actual average NT$31.60/US$1.00 from the 2Q26
Management Report]. This retires a gap that had been carried unverified across four consecutive
passes (07-10, 07-14, 07-16, 07-24).
A new joint venture with Sony Semiconductor Solutions. "Advanced Vision Semiconductor
Manufacturing Corporation," Koshi City, Kumamoto — next-generation smartphone image sensors,
volume production expected 2029. Sony contributes ~¥465B and is sole controlling
shareholder (the JV consolidates into Sony, not TSMC); TSMC contributes ~¥282B in cash,
phased on demand; further capacity investment is premised on Japanese government support
[primary: 6-K accession 0001046179-26-000539, filed 2026-08-11, joint Sony/TSMC release; the
¥282B subscription cap was separately approved as board resolution 4 in accession
0001046179-26-000536].
Q2 dividend NT$7.0/share, and a first: a USD election for foreign holders. Record date
2026-12-16, ex-date 2026-12-10 (same ex-date for ADSs), payable 2027-01-07. "[P]ursuant to the
amended regulations of the Taiwan Financial Supervisory Commission, TSMC will, starting with
this dividend distribution, provide foreign shareholders … with the option to receive cash
dividends in U.S. dollars" [primary: accession 0001046179-26-000536, resolution 2]. Separately,
the Q1'26 dividend was mechanically adjusted to NT$7.00000137 per share after restricted-stock
reclamation [primary: 6-K accession 0001046179-26-000552, filed 2026-09-01].
Balance sheet at 2026-06-30, primary and stronger: cash and equivalents NT$3,134.22B ≈
US$99.2B; bonds payable gross NT$980.36B ≈ US$31.0B (domestic NT$554.84B + overseas
NT$427.70B − NT$2.17B discounts); long-term bank loans NT$49.23B; total equity NT$6,474.47B on
total assets NT$9,375.65B. Net cash ≈ NT$2,104.6B ≈ US$66.6B [primary: accession
0001046179-26-000541 Ex-99.1, balance sheet + Note 17; net-cash arithmetic is
`` over primary inputs]. H1 operating cash flow NT$1,482.34B ≈ US$46.9B, +32.0% YoY;
implied H1 Free cash flowCash left after paying to run and maintain the business. Unlike profit, it is hard to flatter with accounting choices. ≈ US$20.1B.
The structural thesis is unchanged and better evidenced. The foundry + advanced-packaging chokepoint, the customer concentration, the geopolitical tail — nothing in the window touched them. What the window did was convert two long-standing summary-derived numbers (capex level, balance-sheet strength) into primary ones, and add one genuinely new fact (the ¥282B Sony JV) that is small relative to the base but tells you TSMC is now willing to put capital into a JV it does not control.
Nothing cracked. No regulatory action, no management change, no customer loss, no guidance cut in the window.
The Q2 2026 print itself (2026-07-16) is fully analysed in the previous dossier and is
not restated here. Three things have been added to it since:
(a) The board formally approved the quarter. "Approved the 2026 second quarter Business Report
and Financial Statements. Second quarter consolidated revenue was NT$1,270.38 billion and net income
was NT$706.56 billion, with diluted earnings per share of NT$27.25" [primary: accession
0001046179-26-000536]. This is an independent third confirmation of the three headline figures —
they now agree across the 07-16 earnings release, the IR management report, and the board minute.
(b) The reviewed six-month statements landed (2026-08-14) and carry the cash-flow and balance-sheet detail the press release never had. Ingested above. The Taiwan-IFRS six-month statements are the only document on this shelf that states capex as a cash outflow.
(c) The Q3 print is not the next catalyst — the August monthly report is. As of 2026-09-05 the
August 2026 revenue 6-K has not been filed [primary-negative: EDGAR submissions index for CIK
0001046179 read 2026-09-05; the most recent filings of any type are the 2026-09-01 dividend
adjustment and a run of Form 4s]. TSMC's monthly cadence puts it around 2026-09-10. The
earnings-manual.json note for TSMC already flags this: the monthly print "is a higher-frequency
catalyst than the quarterly print and is not modelled anywhere in this system." It still isn't.
transcripts/ remains empty. No call occurred in the window (the next is the Q3 call, expected
2026-10-15), so there is nothing new to ingest — but the Q1'26, Q4'25 and Q3'25 backfill that
the previous dossier listed as an open item is still outstanding, and the deep-dive skill's own
rule is explicit that a refresh which adds zero primary material to a shallow shelf is a refresh
done wrong. This pass added six filings and zero transcripts. Flagged, not fixed.
| Catalyst | Date | Status | Provenance |
|---|---|---|---|
| August 2026 monthly revenue 6-K | ~2026-09-10 | NEXT — not yet filed | primary-negative: EDGAR index, 2026-09-05 |
| September monthly revenue 6-K | ~2026-10-09 | pending | pattern |
| Q3 2026 earnings + Q4 guide | 2026-10-15 (estimated) | pending | research/calendars/earnings-manual.json, firmness estimated, NOT verified against investor.tsmc.com |
| 2027 price increases take effect | 2027-01 | reported, not filed | see below |
| Q1'26 dividend paid | 2026-10-08 | scheduled | primary: accession 0001046179-26-000552 |
| Q2'26 dividend ex-date / record / pay | 2026-12-10 / 12-16 / 2027-01-07 | scheduled | primary: accession 0001046179-26-000536 |
| Sony JV close (regulatory approvals) | not stated | pending | primary: accession 0001046179-26-000539 |
Price: $427.98 last, 2026-09-05. Prior close $417.01 — the name moved +2.63% on the day of this refresh, on no TSMC filing (nothing was filed 2026-09-04 or 09-05 except Form 4s). 52-week high $478.89 was set within the last 13 weeks, i.e. post-print; the stock has given back ~11% from it.
The Q3 revenue guide, converted to the monthly cadence — a check Connor can run on 2026-09-10 in
one line. The guide is $44.6–45.8B at TSMC's own stated FX assumption of NT$32/US$1.00 [primary:
accession 0001046179-26-000451, EX-99.1, verified twice already on this shelf]. That is
NT$1,427.2–1,465.6B for Q3. July printed NT$467.58B. So August + September must total
NT$959.6–998.0B — a monthly average of NT$479.8–499.0B, i.e. +2.6% to +6.7% above July's level
. An August print below
~NT$480B does not by itself miss the guide (September is seasonally the stronger month) but it puts
the low end of the band in play. This is a mechanical decomposition, not a forecast.
The 07-24 verification note's central finding was that the $52–56B FY2026 capex guide had never appeared in any TSMC-filed document, having been carried from earnings-call aggregators across three passes. That finding still stands. The two capex numbers this pass adds do not confirm it, and conflating them would be exactly the error the 07-24 note was written to prevent:
0001046179-26-000545, filed 2026-08-25], which
is what makes the August figure a discrete tranche rather than a running total. It is not a
full-year capex guide and must never be labelled as one.Status of the $52–56B figure: still summary-derived, still uncorroborated by any TSMC-filed document, now in its fifth consecutive pass. The correct next move is the Q3 earnings call transcript, where the number is actually spoken.
Three capital-allocation decisions in a single board meeting, all primary:
Two bond tranches were issued in July: NT$14.0B at 2.03% (5-year) and NT$4.5B at 2.10%
(10-year), bullet repayment, annual interest [primary: accession 0001046179-26-000545]. A company
with US$66.6B of net cash issuing 10-year paper at 2.10% is funding the capex wave at a cost of
capital well below its own returns; that is a choice, not a necessity.
Bull, restated on this window's evidence: revenue growth is accelerating, not decelerating —
July's +44.7% YoY is above the Jan–Jul +37.0% cumulative, which is above the FY guide of "slightly
above 40%" in USD [primary: the July 6-K; the FY guide from accession 0001046179-26-000451
EX-99.2 "Future Outlook"]. The company is funding a US$29.4B capacity tranche out of US$46.9B of
half-year operating cash flow while holding US$66.6B net cash and issuing 10-year debt at 2.10%.
Reported 2027 price increases of 5–10% across advanced and mature nodes, with a further 10–15% on
above-forecast orders, land on a book that is already sold out.
Bear, restated: the price says the market has already paid for this. The stock is below its pre-print level after a three-for-three beat and ~11% off a high set inside the window — the classic shape of a name where the good news is in the number and the risk is in the multiple. Capex is compounding: +34.8% YoY in H1 with a US$29.4B tranche appropriated on top, which is a depreciation wave arriving in 2027–28 whether or not AI demand does. The Sony JV is a small dollar amount but it is capital going into a structure TSMC does not control, on a 2029 payback, partly dependent on a foreign government's subsidy decision. And the Taiwan tail is untouched by any of this.
Devil's advocate — the strongest argument against the bull case, stated properly: Every number in this refresh is a supply number. Revenue, capex, appropriations, capacity — TSMC is telling you how much it is building and shipping. Nothing in the window is a demand number from an end market. The customer concentration means one hyperscaler capex decision moves the whole book, and the one place demand durability actually gets discussed — the earnings call — is precisely the document this shelf has never had. A refresh built entirely on filings is structurally blind to the question that decides the thesis. That is not a reason to doubt the print; it is a reason to weight the Q3 call heavily and to notice that four passes have now flagged the missing transcripts.
the previous dossier.the previous dossier. Nothing in the window
disclosed a supplier or chokepoint change.the previous dossier.the previous dossier. Monthly 6-Ks carry no segment split.n/a rather than a carried-forward table.0001046179-26-000545]. Nothing to flag.model.xlsx does not exist on this shelf
and the CSVs are header-only, so there is no workbook to rebuild and read back; re-forecasting from
the prior dossier's estimates without that spine would be stacked on.
See Open items.the previous dossier, with three additions
in the position seed below.Flagged separately so they are never mistaken for filings:
transcripts/ is empty, so structurally unreachable from this shelf.estimated from
research/calendars/earnings-manual.json, itself explicitly unverified against
investor.tsmc.com's financial calendar.Every dossier we have written on TSMC, newest first.
The Q2 beat was not a peak — July revenue accelerated to +44.7% YoY, the board appropriated a single-tranche US$29.4B of capacity capital, and H1 cape…
The AI complex's single foundry + CoWoS chokepoint just printed a hard 20-F
The chokepoint that captures the foundry + CoWoS-packaging rent for the ENTIRE AI complex (NVIDIA, AMD, Broadcom, Apple all depend on it), trading at…
Covered in the Knowledge Base
Hardware — Memory & Compute Architecture
| Industry | Semiconductors |
| Founded | 1987 |
| Website | Visit TSMC |
Where TSMC sits against the other names we cover on this beat. Each line is that company’s verdict, not a summary of it.
The Q2 print did the one thing the bear case could not survive
Cash $4.7B
The thesis got WIDER and the price got WORSE.
Cash $19.6B
The de-rate the June dossier warned about arrived early and for the exact reason it named
Cash $1.4B
LONG — MEDIUM (conviction DOWN one notch from the prior MEDIUM-HIGH, on wider risk, not a worse business).
Cash $25.0B
The circularity flag stopped being a footnote and became the tape
Cash $13.2B