Peak Memory? Stress-testing the top-of-cycle call · Equity Pulse
Faircurve Equity Pulse thematic note: testing the memory / semiconductor peak-demand hypothesis against this cycle and four prior chip cycles.
By Faircurve Research
Memory margins and memory shares both stand at multiples of every prior record. The demand data say the shortage is real and sold out into 2027; the valuation and price data say the equities are already discounting the top. Those are two different questions — and the honest answer differs for each.
§ 01 — The Backdrop
i.A sell-off that asked the question out loud
On 13 July the semiconductor complex fell hard. Korean equities had their worst session since the 2008 crisis, and the weakness carried into the US chip names. It landed within days of SK Hynix's blockbuster New York listing — priced at $158.14 on 10 July and raising roughly US$26–28bn, among the largest share sales on record — a reminder that supply of paper tends to arrive when enthusiasm is highest.
Micron, the cleanest listed proxy for the memory cycle, closed at $937, down 4.3% on the day and about 23% below its 25 June record of $1,213.56. Its market value is now near US$1.06tn. The question on every desk: has memory demand peaked? We test that claim directly — against this cycle's own fundamentals and four prior semiconductor cycles — and separate the question of a demand peak from the question of an equity peak.
The hypothesis, increasingly voiced in the press, that memory and semiconductor demand is at or near a cyclical peak and turns down soon — and, if so, which companies carry the most risk and how today's valuations compare with prior peaks.
§ 02 — The Cycle In One Company
ii.Micron's margin signature
Memory is the most cyclical corner of technology, and Micron's gross margin traces that cycle almost perfectly. Over a quarter-century it has peaked near 46% (FY2000), 59% (FY2018) and 45% (FY2022) — and on each occasion collapsed to negative territory within one to two years. The median annual gross margin across 2000–2025 is just 23%.
Against that history, the current cycle is not merely strong — it is vertical. Micron's quarterly gross margin has run from −33% in the February-2023 quarter to +85% in the May-2026 quarter, with quarterly revenue rising from $3.7bn to $41.5bn and quarterly earnings from −$2.13 to +$25.04 a share. The last three quarters alone stepped 56% → 74% → 85%.
Every prior gross-margin peak reverted to a loss within one to two years. The current quarter sits at about twice the highest peak Micron has ever printed. Extreme readings do not, by themselves, date the turn — but they define how far there is to fall if history's mean-reversion holds.
§ 03 — What Previous Peaks Looked Like
iii.The anatomy of a memory top
When memory turns, the equity moves are violent. Measured on closing prices, Micron fell 93% in the dot-com bust, 88% into the 2008 low, 54% in the 2018 memory downturn and 50% in 2021–23. The broad chip index (VanEck Semiconductor ETF, SMH) is milder but still fell 28–83% across the same episodes. Recoveries are slow: Micron's 2000 peak was not regained for roughly nineteen years.
| Episode | MU peak→trough | Depth | Months down | Months to recover |
|---|---|---|---|---|
| Dot-com 2000–03 | $96.6 → $6.8 | −93.0% | 31 | 227 |
| GFC 2007–09 | $14.2 → $1.7 | −88.1% | 22 | 55 |
| Memory 2018 | $62.6 → $29.0 | −53.7% | 7 | 23 |
| Memory 2021–23 | $97.4 → $48.9 | −49.8% | 8 | 17 |
The amplitude of memory drawdowns dwarfs the broad index. In this part of the market, position sizing and entry point matter more than being directionally right — a 50% fall has twice preceded a full recovery, but only after a long wait.
§ 04 — Is This A Top?
iv.The evidence, weighed both ways
A rigorous read of the data does not give a one-word answer, because the classic demand-peak signal and the valuation-peak signal point in opposite directions.
Against an imminent demand peak
- No inventory glut. Micron's inventory is flat at about $8.5bn even as revenue quadrupled; days-of-inventory have fallen (FY2023 ~180 → FY2025 ~136). Prior tops in 2018 and 2022 were preceded by rising inventory. This, the single most reliable cycle tell, is not flashing.
- Order books sold out. Press reporting has HBM capacity committed through 2027 and new fabs not operating before then; IDC frames the shortage as a “permanent reallocation” of capacity, not a normal glut-and-correction.
- Demand is structural. AI data-centre build-out, not the PC/handset replacement cycle, is the driver; long-term agreements are replacing volatile spot pricing, improving visibility.
- Earnings still accelerating. Gross margin rose in each of the last three quarters; on trailing earnings the P/E is only ~21x, ~9x on the annualised run-rate.
For a late-cycle equity top
- Margins at ~2× the record. An 85% gross margin has no precedent; mean reversion is the iron law of memory, and the distance back to a 23% median is vast.
- Valuation offers no cyclical cushion. Micron trades at a record 10.5× book and 11.7× sales — versus roughly 2× book at the 2018 peak. The usual “cheap at the top” margin of safety on assets is gone.
- The advance was parabolic. Micron rose ~1,900% from its 2023 low to the June high; such slopes rarely resolve gently.
- The de-rating has begun. The whole complex peaked within days in late June and has already fallen 12–26% (KLAC −26%, Micron −23%). Euphoric commentary and a record-size IPO are familiar late-cycle company.
On the fundamentals, the memory demand peak is not yet visible: the inventory signal that marked every prior top is absent and capacity is committed into 2027. But the equity sits at a valuation and price extreme that has always mean-reverted, and it has already started to de-rate. The two need not coincide — in 2018 the shares topped and fell 54% while revenue and margins were still near record.
§ 05 — Risk, Return & The Shape Of The Advance
v.How the money was made — and how it moves together
From the 2023–25 cycle troughs, every name in the complex multiplied. Micron leads at roughly +1,445% to 13 July; the widely quoted “~1,400%” figure in fact measures from Micron's July-2023 low, not from 2025. Even the diversified sector ETF is up ~326%.
Over five years the risk-adjusted picture is more sober. Micron compounded at 64% a year but with 55% volatility and a 58% peak-to-trough fall along the way — a Sharpe near 1.1 and a Calmar near 1.1. Names with steadier demand (AVGO, KLAC) delivered comparable Sharpe with materially less pain.
| Name | CAGR | Vol | Max DD | Sharpe | Calmar |
|---|---|---|---|---|---|
| Micron (MU) | 64.3% | 54.6% | −57.8% | 1.10 | 1.11 |
| NVIDIA (NVDA) | 59.4% | 51.5% | −66.4% | 1.07 | 0.89 |
| Broadcom (AVGO) | 51.5% | 43.5% | −41.5% | 1.09 | 1.24 |
| KLA (KLAC) | 48.4% | 45.8% | −40.9% | 0.96 | 1.18 |
| AMD | 43.1% | 55.9% | −65.4% | 0.69 | 0.66 |
| Lam Research (LRCX) | 39.8% | 48.0% | −56.8% | 0.74 | 0.70 |
| SMH (sector ETF) | 35.2% | 36.1% | −45.3% | 0.86 | 0.78 |
| Applied Materials (AMAT) | 33.5% | 45.8% | −55.4% | 0.64 | 0.60 |
| TSMC (TSM) | 27.7% | 38.0% | −57.1% | 0.61 | 0.48 |
| ASML | 19.2% | 43.1% | −57.4% | 0.35 | 0.34 |
§ 06 — Valuation Versus Prior Cycles
vi.The peak-earnings trap
Memory's cruel trick is that it looks cheapest at the top. Because earnings peak with the cycle, the P/E troughs just as risk is greatest: Micron traded near 5× earnings at the 2018 peak and screens at ~21× today (~9× on the annualised run-rate). But earnings-based cheapness is only real if the earnings are durable. On measures that do not flatter a margin peak — sales and book value — Micron is at the richest multiples in its history: 11.7× sales and 10.5× book, against roughly 2× book at the 2018 top. The equipment makers are dearer still, at 50–62× earnings and up to 50× book.
| Name | P/E | P/B | P/S | EV/EBITDA | Gross mgn |
|---|---|---|---|---|---|
| Micron (MU) | 20.9 | 10.5 | 11.7 | 23.1 | 72.6% |
| NVIDIA (NVDA) | 31.0 | 25.3 | 19.4 | 25.5 | 74.1% |
| Broadcom (AVGO) | 62.0 | 20.8 | 24.2 | 44.6 | 67.0% |
| TSMC (TSM) | 32.5 | 10.7 | 15.3 | 20.6 | 61.9% |
| ASML | 59.4 | 28.4 | 17.6 | 45.0 | 52.6% |
| Lam Research (LRCX) | 62.0 | 39.0 | 19.0 | 51.5 | 50.0% |
| Applied Materials (AMAT) | 53.8 | 19.1 | 15.7 | 41.0 | 49.0% |
| KLA (KLAC) | 62.5 | 49.9 | 22.2 | 48.6 | 61.4% |
Trailing-twelve-month multiples at 13 Jul 2026 prices. Micron's low P/E reflects peak-cycle earnings; its P/B and P/S are at record highs.
§ 07 — Who Is Most Exposed
vii.The order of the fall, if it comes
Not every chip name carries the same memory risk. Measuring each stock's sensitivity to Micron's daily moves over the past two years sorts the complex cleanly. The memory-levered equipment makers — Lam, Applied Materials, KLA — show the highest beta to Micron and the richest earnings multiples, a poor combination into a possible turn. Diversified logic and foundry (NVIDIA, TSMC) and custom-silicon Broadcom sit lower: their demand is broader and less tied to the memory price.
In order of directness of exposure: the pure memory makers first — Micron, SK Hynix and Samsung, whose economics are the DRAM/HBM price; then the memory-weighted equipment names, which sell into memory capex; then the broader logic and foundry names, which would feel sentiment and any AI-spending wobble but are not price-takers on memory. Micron is most correlated with Lam (0.77) and the sector (0.79), and least with NVIDIA (0.53).
§ 08 — Positioning
viii.From owning the shortage to being paid to carry the risk
The evidence supports a specific, two-part conclusion rather than a slogan. On demand, the peak is not yet in the data: inventories are lean and falling, order books are committed into 2027, and the shortage looks structural rather than speculative. An investor calling an imminent demand collapse is fighting the fundamentals. On the equities, however, the asymmetry has clearly deteriorated: margins and valuations sit at multiples of every prior record, the advance was near-vertical, and the complex has already begun to de-rate from its late-June high. History says the share price tops well before the income statement — so “the stock has peaked” and “demand has peaked” can both be answered honestly, and differently, at the same moment.
Our reading: this is a late-cycle setup in which the reward for carrying peak-memory risk is shrinking. We would be more selective from here — favouring lower-beta, demand-side exposure (diversified logic and foundry) over the highest-beta, richest-multiple memory-capex names, and treating the 2027 supply wave, with inventories and days-of-inventory as the tell, as the fundamental signal to watch. The shortage has been the trade; being paid less and less to own it is the risk.
Not “sell, demand is collapsing.” Rather: the demand data are still firm, but the valuation and price data have removed the cyclical margin of safety. Reduce beta to the memory price, keep exposure to the demand it serves, and watch inventories for the first real crack.
Methodology & definitions
All figures are drawn live from the data source noted below and independently re-computed; none are from recollection. Returns are price-only on split-adjusted daily closes. CAGR = (end/start)^(1/years)−1 over Jul-2021–Jul-2026. Volatility = standard deviation of daily log returns × √252. Max drawdown = largest peak-to-trough decline on closing prices. Sharpe = (CAGR − 4.3% risk-free)/volatility; Calmar = CAGR/|max drawdown|. Beta = covariance of daily returns with Micron ÷ variance of Micron, trailing two years. Gross margin = gross profit ÷ revenue from reported income statements; P/E, P/B, P/S, EV/EBITDA are trailing-twelve-month multiples at 13 Jul 2026 prices. Cycle troughs/peaks were identified programmatically from the closing-price series. Market-colour claims (the 13 July sell-off, the SK Hynix listing, analyst commentary, IDC's characterisation) are from press reporting; figures reported by different outlets are given as a range. Note: certain reported EBITDA/D&A fields for Micron's most recent quarters were internally inconsistent and were excluded; the revenue, gross-profit, operating-income, net-income and EPS lines reconcile and were used.
Data. All market and fundamental data sourced from Financial Modeling Prep (financialmodelingprep.com), retrieved 14 July 2026, and independently recomputed.
Published by Faircurve Research. Informational and educational only; not investment advice, and not a recommendation to buy or sell any security. Figures are point-in-time and will change. Do your own research. An interactive Monte-Carlo fair-value analysis is available at faircurve.io.