One Industry Code, Two Opposite Moves
2026-11-06 · @Mythogen Engine
AI in the Financial Statements, Part 1
On 5 November 2026, the Philadelphia Semiconductor Index fell 12% in a single day and TSMC's ADR dropped more than 10%. On the 6th, Taiwan's market fell as much as 3,737 points after the open, the largest intraday point drop on record (Economic Daily News, Yahoo News). The market is asking: will the AI build-out slow down?
This series does not guess where prices go. It only looks at what the financial statements have already recorded. Part 1 draws a map: in Taiwan, who is expanding, and who is pulling back.
One table, two camps
"Semiconductors are expanding" is the line heard most often this year. It is only half right.
Industry code 24 at the Taiwan Stock Exchange is "Semiconductor". For each of the 96 listed companies under this code with usable data, I calculated the ratio of capital expenditure to depreciation and followed it over the last eight quarters.
Under the same code, two groups are doing opposite things. TSMC and the packaging and test houses kept adding, eight quarters without turning back. Mature-node foundries and silicon-wafer makers kept pulling back, also eight quarters without turning back. The memory makers sit in between: pulling back over the past year, turning around in the latest quarter.
This article does one thing: lay the table out and say clearly what it can and cannot see.
What the ratio measures
Capital expenditure is the plant and equipment bought this year. Depreciation is the value that equipment bought in the past used up this year. Dividing one by the other asks a plain question: is the company spending fast enough to keep up with wear?
\text{Expansion ratio} = \frac{\text{Purchases of property, plant and equipment (trailing four quarters)}}{\text{Depreciation expense (trailing four quarters)}}
There are only three readings:
- Above 1: spending more than is used up; capacity is expanding.
- About 1: replacing what wears out; capacity is flat.
- Below 1: not even replacing wear; capacity is shrinking in relative terms.
I use the trailing four quarters (TTM) because capital spending lands in lumps: it is booked in the quarter the equipment arrives, so single-quarter figures jump around. TTM shows direction, the single quarter shows turning points; the two should be read separately.
The ratio only means something for asset-heavy companies: foundries, memory, packaging and test, silicon wafers. Chip-design houses have no fabs; their capital spending is mostly offices or R&D equipment, and a high ratio says nothing about capacity.
A K shape
Pick out the asset-heavy companies and the eight-quarter trends split into two arms: the upper arm rising for eight straight quarters, the lower arm falling for eight.
| Company | Segment | TTM Q3 2024 | TTM Q2 2026 | Q2 2026 single quarter | Direction |
|---|---|---|---|---|---|
| ASE Technology Holding | Packaging and test | 1.08 | 3.03 | 4.28 | Up eight quarters, still accelerating |
| Sigurd Microelectronics | Packaging and test | 0.90 | 2.29 | 4.41 | Jumped in the latest quarter |
| KYEC | Testing | 1.33 | 3.22 | 2.59 | Peaked at 4.31 in Q4 2025, slowing |
| Powertech Technology | Packaging and test | 0.79 | 2.15 | 1.38 | Peaked in Q1 2026, slowing |
| TSMC | Leading-edge foundry | 1.21 | 2.19 | 2.53 | Up eight quarters, still accelerating |
| UMC | Mature-node foundry | 2.12 | 0.79 | 0.57 | Down eight quarters |
| PSMC | Mature-node foundry | 3.26 | 0.59 | 0.69 | Down eight quarters |
| Taiwan Mask | Photomasks | 1.64 | 0.44 | 0.22 | Down five quarters since peaking in Q1 2025 |
| Formosa Sumco Technology | Silicon wafers | 5.45 | 0.59 | 0.55 | Down eight quarters |
The upper arm is the part of the chain AI uses directly. TSMC's capital spending in Q2 2026 was NT$496 billion, matching the company's own figure (TSMC 2Q26 presentation); its full-year budget has been raised three times this year, to US$60–64 billion (Tom's Hardware). ASE raised its full-year capital spending from US$7 billion at the start of the year to US$10.5 billion, of which US$4 billion is buildings and facilities (MoneyDJ, vocus call notes).
The lower arm is the part AI barely uses. UMC and PSMC make mature-node chips; Formosa Sumco makes silicon wafers. All three ratios fell from above 2 to around 0.6, not even covering replacement.
The upper arm is starting to split as well. ASE, Sigurd and TSMC still have single quarters above TTM, meaning they are accelerating; KYEC and Powertech have single quarters below TTM, meaning they are slowing after a peak. Same "packaging and test", same quarter, two speeds.
The accurate version of "semiconductors are expanding" is: the layer AI uses is expanding; the rest is shrinking. An industry average squeezes the two arms into one deceptively gentle line.
Outside code 24, the companies assembling AI servers have the same shape, only steeper. Wistron's TTM rose from 1.14 in Q3 2024 to 3.31 in Q2 2026, Wiwynn's from 2.47 to 4.55, Quanta's from 1.05 to 2.29; Quanta's Q2 2026 single quarter reached 4.35. The upper arm is not just chips; it is the whole AI supply chain.
How much the market had already paid before the drop
Capital spending has a K shape. Valuations do not. On 3 November 2026 (three trading days before the Taiwan sell-off, and the latest valuation data I have), the price-to-book ratios of companies on both arms were almost all higher than on 90% of the days in their own history. The expanders were expensive, and so were the ones pulling back.
| Arm | Company | P/B on 3 Nov | Own historical median | Historical percentile on 3 Nov |
|---|---|---|---|---|
| Upper | TSMC | 9.25 | 3.70 | 96.8 |
| Upper | ASE Technology Holding | 6.96 | 1.81 | 95.7 |
| Upper | Sigurd | 3.62 | 1.37 | 97.3 |
| Upper | KYEC | 4.95 | 1.38 | 95.4 |
| Upper | Powertech | 3.10 | 1.77 | 83.3 |
| Lower | UMC | 2.74 | 0.94 | 94.3 |
| Lower | Formosa Sumco | 5.04 | 2.42 | 93.4 |
| Lower | PSMC | 2.20 | 1.34 | 72.1 |
| Lower | Taiwan Mask | 2.18 | 1.24 | 72.4 |
| Memory | Winbond | 3.17 | 0.89 | 95.8 |
| Memory | Nanya Technology | 4.40 | 1.48 | 85.3 |
| Memory | Macronix | 2.39 | 1.20 | 89.6 |
P/B is the figure the Taiwan Stock Exchange publishes daily; history runs from September 2005 (ASE Technology Holding from 2018 and PSMC from its 2021 listing). A percentile of 96.8 means that on 96.8% of recorded trading days the stock was cheaper than on 3 November.
In other words, the market did not distinguish which arm was expanding and which was pulling back; it pushed the whole semiconductor board to historic highs together. The drop on 5 and 6 November hit a board that was already at a historic extreme. A high P/B can mean the market expects higher future earnings; it does not mean the stock must fall. The table only shows that the expectation had been paid for in advance.
Wiwynn is left out: its P/B on 3 November was the lowest in its own history, the exact opposite of the other AI server makers. A change in share capital or par value may have broken the exchange's series; I do not use it until that is checked.
Memory: three readings of the same table
On TTM alone, Taiwan's memory makers are all pulling back: in Q2 2026 Winbond was at 0.97, Nanya Technology 0.90, Macronix 0.23, all below 1. That was the conclusion I wrote down the first time I looked at the table. It was wrong.
| Company | TTM trough | TTM Q2 2026 | Q2 2026 single quarter | Full-year budget |
|---|---|---|---|---|
| Winbond | 0.53 (Q4 2025) | 0.97 | 1.73 | NT$42.1 billion at start of year; latest estimate NT$39.5 billion |
| Nanya Technology | 0.75 (Q1 2026) | 0.90 | 1.40 | Raised from NT$52.0 billion to NT$69.7 billion |
| Macronix | Still falling | 0.23 | 0.38 | — |
First reading, TTM: pulling back. The four-quarter total is still below depreciation, because three of the four quarters are trough quarters.
Second reading, single quarter: turning around. In Q2 2026 Winbond spent 1.73 times its depreciation in a single quarter, Nanya 1.40 times. TTM is climbing off the bottom, just a few quarters behind.
Third reading, budget: surging. Financial statements lag company decisions. Nanya's actual capital spending in 2025 was NT$13.4 billion; at its January earnings call it set the 2026 budget at about NT$50 billion (SinoPac Richclub), with a board cap of NT$52 billion; in August it raised the cap to NT$69.7 billion to speed up construction of its new 5A fab and pay equipment deposits early (Economic Daily News). Winbond planned NT$42.1 billion of capital spending for 2026 at the start of the year (Statementdog); its latest estimate, given at the November earnings call, is NT$39.5 billion (Business Today).
Nanya's budget path deserves a look on its own. Its 2025 earnings calls had guided 2026 capital spending at about NT$16–20 billion (vocus). After DRAM prices surged, the budget was raised again and again within a year and ended at three and a half times the top of the original guidance. This is not a fab scheduled years ago and built to plan; the budget is following the price.
Cycle, or a new order?
In mid-August, Sandisk's investor day put forward another view: long-term contracts lock in customers, suppliers keep their discipline, and memory is no longer a commodity cycle. Nanya has made a matching move: this year it brought in Kioxia, SanDisk, Solidigm and Cisco as subscribers to a private placement, tying in long-term customers (Economic Daily News).
The table has room for both views. It can establish only one thing: Taiwan's second-tier DRAM makers are not reining in capacity; they are adding capacity while prices and margins surge, and Nanya's 5A fab is the largest single investment in Taiwan's DRAM industry in nearly twenty years. The 5A fab is expected to start tool installation in early 2027 and reach first-phase capacity only in the first half of 2028 (SinoPac Richclub). Today's decision becomes capacity two years from now. The damage in a commodity cycle has always landed in that gap.
Macronix is the exception: its ratio fell from 1.32 to 0.23 and has not turned back. All three are called "memory", but they make different products and sit at different points in the cycle.
The upper arm: where the money comes from
The upper arm of the K is ultimately paid for by a handful of US cloud giants. This year, the way they pay changed.
- FactSet estimates that new borrowing as a share of these companies' capital spending rose from 9% in fiscal 2024 to 32% by mid-2026 (as cited by mlq.ai).
- Order books for their new bonds fell from about five times covered in February to less than twice in July (Fortune); in September, credit spreads for AI-related issuers were about 115 basis points, against about 78 for investment grade in general (nai500).
- Over the same period, the 30-year US Treasury yield closed at 5.61% on 1 October (FRED).
Third-quarter results, released at the end of October, pushed further in the same direction. Amazon's free cash flow was negative US$42.4 billion, yet it held full-year capital spending at US$220 billion (CNBC); Meta's free cash flow for the quarter was negative US$2.1 billion (Vantage). None of the four cut spending for lack of money; Microsoft trimmed its full-year forecast from US$175 billion to US$170 billion, citing delivery delays at chip suppliers (CNBC).
The Taiwan end is in different shape, and it can be calculated directly from the cash flow statements.
| Trailing four quarters (Q3 2025–Q2 2026), NT$ | TSMC | ASE Technology Holding |
|---|---|---|
| Net cash from operating activities | 2,634.7 billion | 168.8 billion |
| Purchases of property, plant and equipment | 1,491.1 billion | 211.1 billion |
| Free cash flow | 1,143.6 billion | −42.3 billion |
| Cash dividends paid | 531.6 billion | — |
| Net increase in long-term borrowings and bonds | 8.6 billion | 51.4 billion |
After paying for capital spending and dividends, TSMC still had about NT$611.9 billion left over the four quarters; its new borrowing equals only 0.3% of operating cash flow. ASE is the opposite: in three of the four quarters its capital spending exceeded operating cash flow, and the gap was filled with debt. Its net increase in long-term borrowings and bonds, NT$51.4 billion, is larger than its free-cash-flow gap.
The memory makers stand on TSMC's side. Winbond's capital spending in Q3 was about NT$15.3 billion, three times Q2, and its free cash flow in the same quarter was still NT$29.9 billion (TechNews); it is expanding with money it earned itself, on the back of memory prices. If prices turn, that cash flow shrinks first, and the construction already under way will not stop with it.
At its 29 October earnings call, ASE raised this year's capital spending a third time, to US$11.5 billion, with US$8 billion already spent in the first three quarters (Anue); analysts expect its free cash flow could still be negative in 2027 (China Times).
Yet at the open on 6 November, both opened limit-down (Economic Daily News). TSMC, which does not need to borrow, and ASE, which borrows to expand, were not treated differently by the market. This is the same thing as the valuations in the previous section: both arms were expensive together, so both were sold together.
Note: calculated from the consolidated cash flow statements on the Market Observation Post System (MOPS); single-quarter figures for Q2 to Q4 are derived by subtracting year-to-date totals. Net increase = proceeds from long-term borrowings plus bonds issued, minus repayments of long-term borrowings and bonds; short-term borrowings excluded. ASE pays its dividend in Q3, and the database cannot derive the Q3 2025 single-quarter figure for that line (the accounts in adjacent quarters do not match), so it is left blank.
In other words, the upper arm's expansion is built on a demand chain that relies more and more on borrowing, at rising borrowing costs. This table cannot see the top of that chain, only the equipment arriving at the bottom. If money upstream gets more expensive, the first signal on the Taiwan end will probably be single-quarter ratios dropping below TTM. KYEC and Powertech already have that shape, but for now there is no evidence linking the two.
What this table cannot see
A table is only as credible as it is clear about its own blind spots. Each of the six below could change the reading above.
| Blind spot | Effect | Example |
|---|---|---|
| Listed companies only | Over-the-counter companies are excluded, so the mature-node and silicon-wafer picture lacks key names | Vanguard International Semiconductor, GlobalWafers, Phison, WIN Semiconductors |
| Equipment prepayments are not in the numerator | Expansion paid as deposits before equipment arrives is understated | Nanya said in August it was paying equipment deposits early, yet that line is empty in the table; WinWay's and PlayNitride's prepayments are several times their quarterly capital spending |
| Ratios for chip-design houses do not mean capacity | No fabs; capital spending is mostly offices or R&D equipment | Elan 10.88, Global Unichip 4.51 |
| Newly listed companies have too small a base | Depreciation is tiny and the ratio is distorted | Hon Precision 16.26 (only five of eight quarters of data), Song Sheng 15.92 (four quarters) |
| Capital spending includes buildings and land | Buying buildings is not adding capacity | Of ASE's US$10.5 billion budget this year, US$4 billion is buildings and facilities |
| The IFRS 16 break | From 2019, depreciation of leased assets enters the denominator but their acquisition does not enter the numerator | Ratios before and after 2019 cannot be compared directly |
Two more points. First, single-quarter ratios are noisy; the timing of one equipment delivery can double a quarter, so any "turning point" needs at least one more quarter. Second, a few companies file standalone rather than consolidated statements, for example Formosa Advanced Technologies, Phoenix Silicon International and VisEra, so cross-company comparisons are not perfectly like for like.
This article excludes companies with fewer than eight quarters of data and those with quarterly depreciation below NT$100 million.
Three numbers to check next quarter
This article was finalised on 6 November 2026, with financial data through Q2 2026; some companies already disclosed Q3 figures at their October earnings calls, noted below. When the Q3 2026 filings are out, I will come back with the same table and check. Three expectations that can be proven wrong:
- Can Nanya and Winbond use up their full-year budgets? Both disclosed Q3 capital spending at their earnings calls: Nanya about NT$27.6 billion for the first three quarters (Pocket), Winbond NT$23.4 billion (TechNews). Subtracting the first-half figures in the filings, NT$6.8 billion and NT$8.1 billion, gives Q3 single quarters of about NT$20.8 billion and NT$15.3 billion, five times and three times Q2 (call figures, to be checked against the filings). To use up their budgets, in Q4 Nanya would need to spend about another NT$42.1 billion to reach its NT$69.7 billion cap, and Winbond about another NT$16.1 billion to reach its NT$39.5 billion full-year estimate. Falling short means the budget is running ahead of construction. Check date: Q4 2026 filings (March 2027).
- KYEC's and Powertech's single-quarter ratios will stay below TTM. If they bounce back above TTM, the Q2 slowdown was just noise from equipment delivery timing.
- TSMC's and ASE's single-quarter ratios will stay above TTM. Both have raised their full-year budgets; if their Q3 single quarters fall instead, the gap between budget and actual spending needs a second look. ASE raised its budget a third time at the end of October, consistent with this, but it still needs the filings to confirm.
I have made the memory argument in earlier articles: memory is a standard product, suppliers can be swapped, and it cannot escape the cycle. This table does not rubber-stamp that argument. It records only one thing: in the best of times, Taiwan's DRAM makers chose to add. Commodity cycles have never been seeded in the worst of times.
Data and method
The ratios in this article come from a financial-statement database I built myself. The raw data are the financial statements on MOPS from Q1 2013 to Q2 2026, covering 1,095 listed companies, with every original file kept as is.
- Numerator: "Purchases of property, plant and equipment" in the cash flow statement.
- Denominator: "Depreciation expense" in the cash flow statement.
- Single quarters: Taiwan's cash flow statements are year-to-date totals; single-quarter figures are derived by subtracting adjacent quarters. Where adjacent reports use different bases (consolidated or standalone), no subtraction is made and the quarter is left blank.
- TTM: the sum of the latest four single quarters.
- Account mapping: which raw accounts count as capital spending and depreciation was decided by me, one by one, not left to a program to guess.
Spot checks: TSMC's Q2 2026 capital spending of NT$496 billion matches its earnings presentation; ASE's single quarter of about NT$81.5 billion is in line with the quarterly average of its US$10.5 billion full-year budget.
Cross-check: comparing TSMC's 2024 annual report filed with the US SEC against the sum of four quarters on MOPS, the differences in both capital spending and depreciation are less than NT$100,000, and the ratio is 1.4627 in both.
Valuations: the Taiwan Stock Exchange's daily "P/E ratio, dividend yield and P/B ratio by stock", every trading day from 2 September 2005 to 3 November 2026, kept as is. This article uses only the exchange's published figures, with no recalculation.
AI disclosure
The database architecture and specifications were settled in discussion between me and Claude; the code was implemented to spec by Antigravity (Gemini 3.8 Flash), with every step reviewed by me. Number checking, outside fact-checking and the writing of this article were assisted by Claude Opus 5.5; the arguments, choices and conclusions are my own.
Disclaimer
This is an analysis of industry structure, not investment advice, and not a reason to buy or sell any stock. I am not an investment analyst and will not tell anyone what to buy or sell.
This site has no advertising budget and no paid promotion. If this article was useful to you, one share is one vote.
AI in the Financial Statements
- One Industry Code, Two Opposite Moves (this article)
- Change One Estimate, Earn Billions More
- Where Did the Costs Go?
- The Payer Starts Collecting
- Margins Still at Record Highs, and the First Turning Signal Has Appeared