Margins Still at Record Highs, and the First Turning Signal Has Appeared
2026-11-06 · @Mythogen Engine
AI in the Financial Statements, Part 5; also Part 5 of the Memory and AI Compute Cycle series
Nanya Technology's gross margin in Q3 2026 was 85.7%, and Winbond's 70.6%. Both are higher again than in Q2, and far above any peak either company reached in the past thirteen years.
What happens after the peak? The last two cycles left an answer. Four observations, in exactly the same order: revenue growth year on year peaks first; one to three quarters later gross margin peaks; two quarters after that, revenue growth year on year turns negative.
As of Q3 2026, the first step of that sequence has appeared at Nanya: year-on-year revenue growth peaked in Q2, and quarter-on-quarter growth has slowed for two quarters in a row. Winbond's quarter-on-quarter growth is slowing too. Both companies' gross margins are still rising; the second step, the gross-margin peak, has not appeared.
Expansion has also moved from budget to spending: Nanya's capital spending in Q3 alone was about NT$20.8 billion, roughly five times Q2. In the last cycle, both companies' expansion peaked only after gross margins had peaked, and the new capacity ended up arriving in the downturn.
This article answers one question only: following the order of the last two cycles, which step have Taiwan's DRAM makers reached?
Step one: how high this time
| Company | 2017–18 peak | 2021–22 peak | Q2 2026 | Q3 2026 |
|---|---|---|---|---|
| Nanya Technology | 58.9% (Q3 2018) | 49.4% (Q4 2021) | 79.5% | 85.7% |
| Winbond | 38.9% (Q2 2018) | 48.6% (Q1 2022) | 66.2% | 70.6% |
A year ago Nanya was still losing money: its gross margin in Q2 2025 was −20.6%. Within four quarters it swung by a hundred percentage points.
Quarterly gross margin = gross profit (net) ÷ operating revenue. Figures up to Q2 2026 come from quarterly reports on the Market Observation Post System (MOPS), calculated quarter by quarter in a database I built myself; Q3 figures come from the companies' earnings calls, with the quarterly reports due in mid-November.
Step two: how the last two cycles turned
| Cycle | Company | ① Revenue growth (YoY) peaks | ② Gross margin peaks | ③ Revenue growth (YoY) turns negative | Gross margin after the peak |
|---|---|---|---|---|---|
| 2017–18 | Nanya Technology | Q2 2018 (+94.8%) | Q3 2018 (58.9%) | Q1 2019 | 25.8% five quarters later |
| 2017–18 | Winbond | Q4 2017 (+22.6%) | Q2 2018 (38.9%) | Q4 2018 | 22.1% six quarters later |
| 2021–22 | Nanya Technology | Q3 2021 (+55.6%) | Q4 2021 (49.4%) | Q2 2022 | −25.2% seven quarters later |
| 2021–22 | Winbond | Q2 2021 (+97.9%) | Q1 2022 (48.6%) | Q3 2022 | 27.7% four quarters later |
Across four observations, the order of the three signals had no exceptions:
- ① to ②: the quarter of fastest revenue growth came one to three quarters before the gross-margin peak. Revenue was still rising, only more slowly.
- ② to ③: two quarters after the gross-margin peak, revenue growth year on year turned negative. All four times, two quarters.
- After ③: within four to seven quarters, gross margin fell by 17 to 75 percentage points.
What makes this sequence useful: the gross-margin peak can only be confirmed afterwards, but a slowdown in revenue growth can be seen as it happens. On the last two cycles' evidence, it is the earliest turning signal in the financial statements.
In May, in HBM Is Not the Next TSMC (in Chinese), I argued that memory is a cyclical standard product bound by JEDEC standards. That was a structural inference; the four observations above are the evidence thirteen years of financial statements give it.
Step three: which side of the peak the expansion lands on
The strength of expansion is measured by capital spending divided by depreciation (trailing four quarters): above 1 is expansion, below 1 is not even replacing wear.
| Cycle | Company | Expansion ratio peaks | Relative to gross-margin peak | Gross margin that quarter |
|---|---|---|---|---|
| 2017–18 | Nanya Technology | Q2 2017 (5.94) | Five quarters earlier | 43.9%, still rising |
| 2017–18 | Winbond | Q1 2018 (2.87) | One quarter earlier | 37.7%, still rising |
| 2021–22 | Nanya Technology | Q1 2023 (1.70) | Five quarters later | −8.6% |
| 2021–22 | Winbond | Q4 2022 (4.59) | Three quarters later | 37.5%, revenue down 26% year on year |
The two cycles ended in opposite ways. In 2017–18, expansion ran ahead of the peak. In 2021–22, expansion landed after the peak: gross margins were already falling, yet both companies kept increasing capital spending for three to five more quarters. In the quarter Nanya expanded hardest, its gross margin was negative.
The money for the last expansion was spent in the downturn, and the capacity came online afterwards. The damage in a commodity cycle usually lands in exactly this gap.
In "HBM Is Not the Next TSMC" I also wrote about the fate of memory makers: when prices soar they inevitably expand, and the capacity comes online 18 to 24 months later. The two companies' 2021–22 numbers are the trace that mechanism left in Taiwan's financial statements.
Where things stand today
| Company | Revenue growth (YoY) | Revenue growth (QoQ) | Gross margin | 2026 capital spending |
|---|---|---|---|---|
| Nanya Technology | 684.2% → 611.7% | 68.2% → 61.9% | 79.5% → 85.7% | Budget NT$69.7 billion; NT$27.6 billion spent in the first three quarters |
| Winbond | 184.7% → 278.2% | 56.4% → 37.6% | 66.2% → 70.6% | Full-year estimate NT$39.5 billion; NT$23.4 billion spent in the first three quarters |
Left of the arrow is Q2 2026, right is Q3. Q3 figures come from the companies' earnings calls (Nanya: Statementdog; Winbond: UAnalyze) and will be checked against the quarterly reports when they are published in mid-November.
Following the order of the last two cycles, the first turning signal has appeared at Nanya: year-on-year revenue growth peaked in Q2. Winbond's year-on-year growth is still rising, so by definition it has not reached step one, but its quarter-on-quarter growth has slowed as well. Both companies' gross margins are still at record highs; the second step, the gross-margin peak, has not appeared.
Monthly revenue shows it more clearly. In the first month of each quarter, the new quarter's contract prices start to be booked: Nanya's July revenue was NT$43.87 billion, up 49.3% from June; its October revenue was NT$46.48 billion, up only 3.1% from September. The price increase in Q4 is clearly smaller than in Q3. Winbond said at its earnings call that non-AI demand is weak in Q4 and that DRAM prices still have room to rise, but possibly by less than in Q3 (UAnalyze).
Part of the fall in Nanya's year-on-year growth is a base effect (see below). But quarter-on-quarter and month-on-month growth are slowing at the same time, which cannot all be put down to the base.
In May, in HBM Is Not the Next TSMC (in Chinese), I wrote down a pattern: memory stocks usually start falling one to two quarters before the fundamentals. Korean stocks peaked in June and fell sharply in July; in early November they fell again: on 5 November Micron dropped 17.46% in a day (Wealth Express), and on 6 November the Korean market fell as much as about 8.8% after the open, triggering a circuit breaker (Commercial Times). The fundamentals were still rising through Q3, only more slowly. So far, the order of the two fits that pattern, but has not proved it.
By that pattern, the first step should show up in the Q3 or Q4 2026 financial statements. Nanya's Q3 figures already fit. The next thing to watch is gross margin: in the last two cycles, gross margin peaked one to three quarters after the first step.
Expansion has moved from budget to spending. Nanya's actual capital spending in 2025 was NT$13.4 billion; this year's budget went from about NT$50 billion at the start of the year to NT$69.7 billion in August, to speed up construction of the new fab and pay equipment deposits early (Economic Daily News). It had spent NT$27.6 billion by the end of the third quarter (SETN); subtracting the NT$6.8 billion in its first-half filings, Q3 alone was about NT$20.8 billion. Winbond planned NT$42.1 billion of capital spending for the year at the start of 2026 (Statementdog); its latest estimate is NT$39.5 billion, with NT$23.4 billion spent in the first three quarters (Business Today); on 3 November its board approved a further NT$73.16 billion of capital spending to build out Module B at its Kaohsiung fab, preparing capacity for 2027 and beyond (Statementdog). Through Q2, the trailing four-quarter expansion ratio was still below 1; Q3's spending will push it up, and the exact figure has to wait for the quarterly reports.
Nanya's new fab is expected to start tool installation in early 2027 and reach first-phase capacity only in the first half of 2028 (SinoPac Richclub). So the real question is not "will they expand?" but: does the gross-margin peak come before or after 2027? If before, it is the 2021–22 script all over again.
Where this conclusion could be wrong
Only the four things that would overturn the conclusion above.
- The sample is small. Two companies, two cycles; but the two are peers in synchronised cycles, so there are really only two independent samples. "Two quarters apart" is a hypothesis, not a law.
- Base effects. This cycle is a rebound from a loss-making trough, and the base rose steeply from Q3 2025. Part of the fall in Nanya's Q3 year-on-year growth is mechanical, and it cannot on its own be taken as a peak; that is why this article also watches quarter-on-quarter and month-on-month growth. If Q4 quarter-on-quarter growth reaccelerates, the first step was a misreading.
- An accounting break. From 2019, depreciation of leased assets enters the denominator but their acquisition does not enter the numerator (IFRS 16). The absolute expansion ratios of the two cycles cannot be compared directly, but the order of events is unaffected.
- Process transitions. Nanya's single-quarter expansion ratio reached 10.15 in Q4 2016, possibly related to a process transition, not yet verified. If so, "expansion ran ahead of the peak" in 2017–18 may have been a technology change, not a bet on the cycle.
Receipts
Both predictions below have a check date. When the results are out, I will check them with the same data and definitions, and if I am wrong I will say so.
| # | Prediction | Basis | Check date |
|---|---|---|---|
| 1 | At least one of Nanya and Winbond has a lower Q4 2026 gross margin than in Q3 (85.7%, 70.6%) | In the last two cycles, gross margin peaked one to three quarters after the first step; Winbond said Q4 price increases may be smaller than Q3's, and Nanya's October revenue was up only 3.1% from September | February 2027 (Q4 earnings calls) |
| 2 | Winbond's Q4 2026 quarter-on-quarter revenue growth is below 20% (Q3: 37.6%) | Winbond said non-AI demand is weak in Q4 and price increases are narrowing | By 10 January 2027 (December revenue release) |
If number 1 fails, the distance from the first step to the gross-margin peak is longer this cycle than in the last two.
Quarterly reports take time; monthly revenue does not. Taiwan-listed companies publish revenue every month, one to two months ahead of the quarterly reports. Following the order of the last two cycles, that is where the earliest signal hides.
Data and method
Data up to Q2 2026 come from the financial statements on MOPS from Q1 2013 onwards, with the original files kept and calculated quarter by quarter in a database I built myself. Q3 2026 revenue, gross margin and capital spending come from company earnings calls and news reports, and monthly revenue from MOPS; when the quarterly reports are published in mid-November, they will be recalculated with the same definitions.
- Gross margin: gross profit (loss), net ÷ total operating revenue, single quarter. Q4 is derived as the full year minus the first three quarters.
- Revenue growth (YoY): single-quarter revenue against the same quarter a year earlier.
- Expansion ratio: "purchases of property, plant and equipment" ÷ "depreciation expense" from the cash flow statement, trailing four quarters. Cash flow statements are year-to-date totals; single quarters are derived by subtracting adjacent quarters.
- Account choice: which accounting line to use was decided by me, one by one.
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. The query design, outside fact-checking and 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 the industry cycle, not investment advice, and not a reason to buy or sell any stock. I 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
- 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 (this article)
📚 The Memory and AI Compute Cycle series (parts 1–4 in Chinese)
- HBM Is Not the Next TSMC: The Structural Gap Between Memory and Logic Chips
- The Truth About 900 Layers: When Your Switch and Steam Deck Pay for AI
- I Asked AI Whether Memory Stocks Are Still Worth Buying
- Honesty Is a Paid Feature: How AI Pricing Decides Who Gets the Truth
- Margins Still at Record Highs, and the First Turning Signal Has Appeared
@Mythogen Engine