When Will AI Capex Peak? What Chip Stock Prices Imply, November 2026

All figures below are publicly reported market data or sell-side consensus estimates, not company-reported results unless labeled “actual” or “company guidance.” Snapshot date is 9 November 2026 U.S. close (Korean ordinary shares: 9 November 2026 KRX close). Korean won figures are converted at 1 USD = 1,352.90 KRW (9 November 2026 reference rate). Forward P/E definitions differ by vendor (next fiscal year vs next-twelve-months); where sources disagree, the range is shown.

Dated snapshot table

Company (listing)	Price (9 Nov 2026)	YTD	Forward P/E	Consensus EPS (public estimates)	Coverage / flags
Micron (MU, USD)	$733.15	+157.1%	4.15	FY2027 (Aug): $176.50 (36); FY2028 (Aug): $207.76 (36). FY2026 actual used as base: $75.52	Thick through FY2028. FY2029+ exists but only 11 analysts on Seeking Alpha and the number is paywalled — treat as thin
SK Hynix (000660.KS)	$839.68 (₩1,136,000)	+74.5%	2.19–2.27	CY2026: $265.11 (₩358,672, 38); CY2027: $389.65 (₩527,157, 38)	2026–27 thick. No free 2028 street print found. FnGuide’s lower 2026–28 set looks stale vs current Yahoo
Samsung Electronics (005930.KS)	$140.36 (₩189,900)	+61.6%	2.67	CY2026: $34.11 (₩46,148, 28); CY2027: $56.17 (₩75,994, 33)	2026–27 usable. 2028 consensus not in free Yahoo. Do not use single-broker models as “the Street”
Nvidia (NVDA, USD)	$178.67	−4.0%	11.2–14.8	FY2027 (Jan): $9.31 (51); FY2028 (Jan): $15.91 (52)	Thick. Forward P/E depends on year: $178.67 / $9.31 = 19.2x current FY; / $15.91 = 11.2x next FY. StockAnalysis prints 14.75 (likely NTM)
Broadcom (AVGO, USD)	$319.48	−7.7%	16.4–18.4	FY2026 (Oct/Nov): $11.67 (48); FY2027: $19.47 (48)	FY2026–27 thick. No free annual FY2028 consensus. Quarterly averages sum to about $30.2, but each quarter has only 11–21 analysts — thin. Management has said FY2028 EPS should exceed $30 (company guide, not consensus)
TSMC ADR (TSM, USD)	$388.62	+27.9%	16.6	CY2026: $17.50 (11); CY2027: $23.39 (13). A TradingView-sourced feed (8 Nov) shows CY2028 $29.45	EPS coverage is thin (11–13) even though revenue coverage is ~40. Yahoo revenue estimates are in TWD; ADR EPS is in USD. 2028 is a secondary feed, not Yahoo
Celestica (CLS, USD)	$222.92	−24.6%	10.0–12.3	CY2026: $11.56 (21). CY2027 annual consensus blank on Yahoo	2027 street EPS is missing/thin. Company guidance (not consensus): $21.50 adjusted EPS on $40B revenue. GuruFocus forward P/E 10.04 vs StockAnalysis 12.26
Amazon (AMZN, USD)	$266.93	+15.6%	24.5–27.0	CY2026: $12.80 (53); CY2027: $10.67 (57). Markets Insider extends CY2028 $14.63, CY2029 $18.10, CY2030 $22.53 (analyst counts not shown)	2026–27 thick. 2026 GAAP is inflated by investment gains (Q2 surprise +215%). Normalized 2028 cited elsewhere near $13.37, below the GAAP Markets Insider figure. 2028+ should be treated as thinner
Alphabet (GOOGL, USD)	$288.15	−7.8%	14.6–20.9	Yahoo GAAP: CY2026 $19.53 (45); CY2027 $14.95 (51). A separate Nov feed shows CY2026 avg $11.54 (43), which looks normalized	Do not mix GAAP and normalized. Q2 2026 actual $9.11 vs ~$2.90 expected is a one-time spike. MarketScreener’s 2026–28 EPS page is dated 31 Aug 2026 and is stale. Forward P/E swings with the EPS definition
Microsoft (MSFT, USD)	$545.18	+12.7% price / +13.4% total	22.5–26.3	FY2027 (Jun): $20.01 (32); FY2028 (Jun): $24.20 (33). Markets Insider: FY2029 $29.61, FY2030 $34.18, FY2031 $46.31	FY2027–28 solid (30+). FY2029–31 analyst counts not disclosed in the free table — treat outer years as thinner
Meta (META, USD)	$641.37	−2.6%	19.5	CY2026: $29.12 (55); CY2027: $33.97 (56)	Thick through 2027. No free annual 2028 EPS on Yahoo. Recent quarters missed (Q2 −14%, Q3 −29%), and the 2026 estimate has been cut from ~$31.8 ninety days ago

Sources for the table: Yahoo Finance quotes and analysis pages, StockAnalysis, MarketScreener, MetaCap, financecharts / Investment Visual, GuruFocus, Markets Insider, and the 9 November USD/KRW reference rate.

How to read the multiples

The memory names are the valuation outlier, and the mechanism is the estimate revision, not a cheap multiple on last year’s earnings. Micron’s published FY2027 consensus is $176.50 versus a FY2026 base of $75.52, so a 4.15x “forward” P/E is price divided by an earnings year that is more than double the year just reported. SK Hynix and Samsung sit in the same setup: Yahoo’s next-year EPS is 47% and 65% above the current-year estimate, and both trade at about 2–3x that next year. Those multiples only look extreme if the 2027 numbers survive. They are consensus, not guidance, and memory consensus has been revised up sharply in the last 90 days (Micron’s FY2027 estimate moved from about $154 ninety days ago to $176.50).

Hyperscaler and foundry multiples are ordinary once the year label is fixed. Nvidia at $178.67 is 19x the FY2027 consensus of $9.31 and 11x FY2028’s $15.91; vendors printing 11x, 15x, or 25x are not disagreeing about the stock so much as about which earnings year “forward” means. Broadcom’s 16–18x band is price over FY2027’s $19.47. TSMC’s 16.6x matches $388.62 / $23.39. Amazon and Alphabet look expensive on forward P/E only if 2027 is used after a 2026 year stuffed with mark-to-market gains: Amazon’s published consensus falls from $12.80 to $10.67, and Alphabet’s Yahoo GAAP consensus falls from $19.53 to $14.95.

Fiscal year and currency traps

• Micron: fiscal year ends late August. “2027” on Yahoo is the year ending August 2027, not calendar 2027.
• Nvidia: fiscal year ends late January. Yahoo “current year (2027)” is the year ending January 2027. The next earnings date on the quote page is 17 November 2026.
• Broadcom: fiscal year ends early November. Yahoo “current year (2026)” is the year ending around November 2026; earnings date shown as 9 December 2026.
• Microsoft: fiscal year ends 30 June. Yahoo “current year (2027)” is FY2027, already underway (Q1 FY2027 has been reported).
• SK Hynix and Samsung: calendar year, December year-end, KRW on the primary listing. EPS above is converted; the raw consensus is in won. Prices moved again on 10 November while the KRX was open (Samsung was around ₩197,500 intraday versus the ₩189,900 prior close), so a same-day comparison with U.S. closes will not match.
• TSMC: figures are for the NYSE ADR in USD. Yahoo’s revenue consensus on the same page is in TWD (2026 avg TWD 5.58 trillion, 40 analysts) while EPS is in USD. Do not divide a USD price by a TWD EPS.
• Celestica: calendar year, USD, NYSE listing of a Canadian company. Adjusted EPS is what the Street and the company guide; GAAP and adjusted diverge.

Where consensus is thin, stale, or not really consensus

Three gaps matter more than the rest.

TSMC’s EPS sample is small. Eleven analysts on 2026 EPS and thirteen on 2027, against about forty on revenue. The USD EPS number is usable but easier to move than Micron’s or Nvidia’s. The $29.45 2028 figure is from a TradingView consensus feed dated 8 November, not from the Yahoo table, and Seeking Alpha shows a 2028 column with only 10 analysts and the value behind a paywall.

Celestica’s 2027 number in circulation is company guidance, not a street average. Yahoo’s annual 2027 EPS cell is blank. Management has guided $21.50 adjusted EPS and $40 billion of revenue. Using that as “consensus” overstates how settled the Street is, especially after the stock’s −24.6% YTD and a 2026 free-cash-flow guide that was cut.

Alphabet and Amazon 2026 EPS are not clean run-rates. Both had enormous positive surprises in mid-2026 (Alphabet Q2 actual $9.11 vs $2.90 expected; Amazon Q2 $5.75 vs $1.82). Yahoo’s 2026 averages embed those gains, which is why 2027 consensus is lower. A November feed that shows Alphabet 2026 EPS near $11.54 is a different definition. MarketScreener’s Alphabet forecast page, last stamped 31 August 2026, should not be used as current.

Also thin or not independently confirmed in free sources: Micron FY2029 (11 analysts, number not public on the page reviewed); Broadcom FY2028 as an annual figure; SK Hynix and Samsung 2028; Meta 2028; Microsoft FY2029–FY2031 analyst counts.

What the cross-section implies

The tape on 9 November split the AI complex in half. Memory and the foundry were still up huge year to date — Micron +157%, SK Hynix +75%, Samsung +62%, TSMC +28% — while the designers and the buyers were flat to down: Nvidia −4%, Broadcom −8%, Celestica −25%, Alphabet −8%, Meta −3%, with Microsoft and Amazon only modestly positive. That is not a uniform “AI multiple.” It is the market paying up for the bottleneck (HBM, DRAM, leading-edge wafers) and marking down names where 2026 earnings already include one-time gains or where the next leg of EPS depends on capex that has not yet been earned.

For anyone underwriting these names off this table, the comparable number is next fiscal year’s consensus EPS divided into the 9 November close, with the year-end month written down, and with Amazon and Alphabet restated on a normalized basis before any multiple is compared with Micron or Nvidia. A 4x Micron forward P/E and a 20x Alphabet forward P/E are not the same kind of statistic.

---

Recent Findings Supplement (November 2026)

As of early-to-mid November 2026 (data primarily from Nov 9–10, 2026 closes or latest available quotes on Yahoo Finance, MarketWatch, and related analyst aggregates), semiconductor and AI-related stocks show significant volatility and elevated valuations driven by AI demand, with updated consensus EPS estimates reflecting strong growth through 2028.[1][2][3]

Data below draws exclusively from sources published after May 10, 2026. All figures are labeled as publicly reported analyst consensus estimates or market data (e.g., from Yahoo Finance, MarketWatch, or company aggregates); later-year estimates (2028+) often have thinner coverage or wider ranges. Fiscal years vary (e.g., MU fiscal ends August/September; NVDA ends January). Currencies noted where non-USD. YTD changes are approximate based on reported performance vs. S&P 500 benchmarks in sources. Forward P/E is typically next-12-months or fiscal-year forward where specified.

Key recent context (post-May 2026 updates): Analyst EPS revisions have been sharply upward for memory/AI chip firms amid record revenues and margins (e.g., MU Q4 FY2026 results showed massive YoY jumps). Tech giants (AMZN, GOOGL, MSFT, META) have more stable but still growing estimates tied to cloud/AI capex. Consensus for 2028+ is often less dense or shows moderating growth rates.[4][5]

Semiconductor/Memory Companies (MU, SK Hynix, Samsung, NVDA, AVGO, TSM, CLS)

• Micron (MU, USD): Latest price ~$690–$795 range (e.g., ~$733 close Nov 9 per MarketWatch; volatile with sessions around $795 earlier in Nov). YTD return ~142% (as of ~Nov 6). Forward P/E ~4.8–5.1x (deep discount to growth). Consensus EPS: FY2026 ~73.85 (largely actual/reported); FY2027 ~175.61 (range ~140–219); FY2028 ~211.13 (range ~146–336); FY2029 ~217.90. Strong upward revisions post-Q4 results; fiscal year ends ~Aug/Sep.[1][5]
• SK Hynix (000660.KS, KRW; ADR SKHY in USD): Latest price ~1,202,000 KRW (~Nov 10). Forward P/E ~2.3–3.6x. Consensus EPS (KRW): 2026 ~308,665 (high growth from prior ~59k); 2027 ~434,696–547k range across sources; 2028 ~433k–570k. Massive YoY jumps expected; strong buy consensus with high targets.[6][7]
• Samsung Electronics (005930.KS, KRW): Latest price ~198,200 KRW (recent session). Consensus EPS (KRW, per Yahoo aggregates): 2026 ~47,639; 2027 ~72k–77k. Revenue estimates also sharply higher (2026 ~700T+ KRW). HBM/memory strength cited in updates.[8]
• Nvidia (NVDA, USD): Latest price ~$178.67 (Nov 9 close). YTD ~4%. Forward P/E ~12–21x (sources vary; e.g., 11.81x or ~21.5x in recent commentary). Consensus EPS: Current year (FY2026/27) ~4.69; next year (FY2027/28) ~7.46; TTM ~7.57–7.91. Pre-earnings (Nov 17) updates note potential beats; forward multiples compressed recently.[2][9]
• Broadcom (AVGO, USD): Latest price $319.48 (Nov 9 close). YTD ~7% or slightly negative per some reports (~7.18% in one overview). Forward P/E ~16.4–18.4x. Consensus EPS: 2026 ~11.68; 2027 ~19.32; 2028 ~30.66. Strong analyst targets (~$530 avg).[10][11]
• TSMC (TSM, USD ADR): Latest price ~$388.73 (Nov 9). Consensus EPS (USD): 2026 ~17.07–17.18; 2027 ~22.44–23.12; 2028 ~29.45. Revenue growth projected 40%+ in 2026–27 per some models; Bernstein raised targets citing ASIC/CPU demand.[3][12]
• Celestica (CLS, USD): Limited specific post-May 2026 consensus details in results; generally tracks EMS peers with AI/server exposure. Recent analyst focus on supply chain beneficiaries implies upward EPS pressure, but coverage thinner than peers—flag as potentially stale or sparse for 2027–28.

Hyperscalers (AMZN, GOOGL, MSFT, META)

These show steadier (but still AI-boosted) growth. Specific latest prices/YTD/forward P/E/EPS not detailed in the most recent targeted results above, but general patterns from cross-referenced coverage indicate forward P/Es in the 15–28x range, with EPS estimates for 2026–2028 reflecting continued cloud/AI revenue expansion (e.g., double-digit growth tapering in outer years). Consensus tends to be denser for these names through 2027 but thinner for 2028+. Check latest Yahoo/MarketWatch for exact Nov 2026 quotes (e.g., AMZN/MSFT/META typically trade with moderate YTD gains vs. semis).[13]

Notes on data rigor and limitations: All values are aggregates from public sources (Yahoo Finance analysis pages, MarketWatch estimates, Reuters/Bloomberg summaries referenced in articles) as of Nov 2026 publications. EPS consensus often shows recent upward revisions (e.g., MU 2027 estimates rising month-over-month). Later years (2028+) frequently have fewer analysts or wider high/low ranges—flagged as thinner/staler where applicable. Exchange rates apply for KRW figures (convert at prevailing rates for USD equivalents). Fiscal/calendar year differences noted; always verify latest on IR pages or terminals for real-time updates. No single source had complete data for all 11 tickers in one view.[1][8]

This represents the most recent publicly available consensus snapshots; markets move quickly, and pre-earnings periods (e.g., NVDA Nov 17) can shift estimates rapidly.

The equity table isn’t in the question, so I’ll pull current prices, consensus earnings, and through-cycle valuation ranges for the names the prompt itself flags—memory plus Nvidia, Broadcom, and TSMC—and build the reverse-valuation grid from that.
Prices and street EPS are in hand. Next I’ll pin down fiscal-year estimates and historical peak-to-trough earnings so the reverse-valuation grid is anchored, not guessed.
The equity table was not in the prompt. The framework below is built on the four names the question itself uses as the valuation set — Micron for memory, plus Nvidia, Broadcom, and TSMC — with SK hynix as a memory cross-check. Prices are the November 9, 2026 closes. Consensus EPS vintages differ by source and by a few weeks; where they disagree, the range is shown rather than averaged away.

What the price is actually paying for

Divide the price by the through-cycle P/E you are willing to pay for earnings that survive the cycle. That dollar figure is the only EPS the price treats as durable. Everything above it is what the market assumes will be given back, and the year you think earnings will have fallen to that level is the trough year already in the price.

On November 9 that residual was about $73 for Micron at 10×, $8–9 for Nvidia at 20–22×, $18 for Broadcom at 18×, and $22 for TSMC at 18×. Set against the street, those residuals say four different things. Micron’s price is roughly 10× the year it just reported, so the entire guided ramp is treated as transient. Nvidia and Broadcom are priced as if next fiscal year is the sustainable level and the year after that is a peak. TSMC is priced as if 2027 earnings are already the through-cycle number, with only a shallow fade after that.

The identity, with no discounting and no credit for interim cash:

sustainable EPS = price / through-cycle P/E

decline already in the price = 1 − (price / through-cycle P/E) / street peak EPS

Discounting is added only in the peak-year grid, because an undiscounted identity does not change with the calendar. When a discount is applied, it is a harsh clock: fair value at the peak equals through-cycle P/E × trough fraction × peak EPS, discounted at 10% a year, with zero credit for dividends or free cash flow earned on the way. That understates fair value. A name that looks slightly rich on this test can still be fair once interim cash is counted. A name that looks cheap on it is cheap even if you ignore the cash.

Assumptions, stated before the grid

Through-cycle P/E is the multiple applied to post-fall earnings, not to the peak.

• Memory: 8×, 10×, 12×, with 15× as the top of the range you suggested. Micron’s own history says why the low end exists. Diluted EPS went from $11.51 in fiscal 2018 to $2.37 in fiscal 2020, about 21% of the peak, and from $7.75 in fiscal 2022 to a loss of $5.34 in fiscal 2023. [1] Upturn multiples on peak earnings have historically sat around 4–7×, which is a peak multiple, not a through-cycle one. [2]
• Nvidia: 20×, 22×, 25×. The five-year median trailing P/E is about 56×, and one broker’s five-year average is 33.6×, but both are hypergrowth artifacts. [3] [4] A through-cycle multiple here means the multiple once growth has normalized, closer to a premium industrial than to the last five years.
• Broadcom: 16×, 18×, 20×. The ten-year median forward P/E is 15×; a five-year average cited in broker work is 20.5×. [5] [4]
• TSMC: 16×, 18×, 20×. A recent trailing history of the forward multiple puts the five-year average near 18×. [6]

Trough as a percent of peak is scenario, not fact.

• Memory base case 35% of peak EPS still earns money. That is harsher than fiscal 2019 (about 48% of the 2018 peak) and kinder than fiscal 2020 (about 21%) or fiscal 2023 (below zero). [1] The grid also shows 50%, 40%, 25%, and 10%.
• Nvidia and Broadcom base case 50–55%. A capex pause, not a DRAM price collapse.
• TSMC base case 70–75%. In the last down year, 2023 ADR EPS of $5.19 was about 79% of 2022’s $6.57. [7] Foundry earnings fade; they do not invert the way memory earnings do.

Peak year is an assumption you choose. The street, as of early November, is still raising outer-year estimates. It is not modeling a peak. [8] [9] [10]

The snapshot the grid is built on

Name	Nov 9, 2026 close	Earnings anchor	Source of EPS	P/E on that anchor
Nvidia	$178.67	FY2027E $9.30; FY2028E $15.80	FactSet-style consensus, early November; cross-checks at 19.2× and 11.3×	19.2× / 11.3×
Broadcom	$319.48	FY2027E $19.47; FY2028E $30.37	Seeking Alpha consensus, Nov 6; screen NTM implies ~$19.3	16.4× / 10.5×
TSMC ADR	$388.62	2027E $22.2; 2028E $28.2	FactSet local medians of NT$145.69 and NT$185.61 on Oct 15, ×5, ÷ 32.882 TWD per USD on Nov 9. Seeking Alpha’s Oct 29 2027 figure is $23.39; one data vendor has 2028 at $29.82	17.5× / 13.8×
Micron	$733.15	FY2026A ~$75.4; FY2027E $176.50; FY2028E $207.76	FY2026 basic EPS $75.38; Seeking Alpha Nov 6 for the outer years (FY2027 range $161–$215). A Nov 9 screen forward P/E of 3.47 implies NTM EPS near $211, not $177	9.7× trailing / 4.2× / 3.5×
SK hynix ADR	$134.38	TTM $19.41; forward ~$37.0	Price ÷ published forward P/E of 3.63	6.9× / 3.6×

Prices: Nvidia, Broadcom, TSMC, Micron, and SK hynix all printed on November 9 across market-data pages. [11] [12] [13] Nvidia’s $9.30 / $15.80 path is the early-November consensus and matches a published FY2027 estimate multiple of 19.2× and FY2028 of 11.3× on a $179 handle. [8] [3] Broadcom’s $19.47 / $30.37 path is a single consensus page as of November 6, when the stock was $294; the EPS figures are estimates, not a function of that price. [10] TSMC’s local medians are FactSet as of October 15; the FX rate is the Taipei interbank close on November 9. [14] [15] Using Seeking Alpha’s higher 2027 ADR estimate of $23.39 instead of $22.2 moves the multiple by about one turn, not the conclusion. [16] Micron’s fiscal 2026 EPS and the $176.50 / $207.76 path are the two load-bearing memory numbers; the screen’s ~$211 NTM figure is the disagreement to keep in view. [17] [9] [18]

No public, unpaywalled consensus EPS was found for 2029 or 2031. Those columns are implied peaks only. Inventing a street number there would fake a comparison.

How far earnings can fall, and still justify today’s price

This is the undiscounted identity. “Justified” means price = through-cycle P/E × whatever EPS is left. If the cell is negative, even a zero decline does not cover the price at that multiple — the peak has to come in above the street, or the multiple has to be higher.

Micron, from the street peak

Through-cycle P/E	EPS the price treats as durable	Fall from FY2027E $177	Fall from FY2028E $208	Fall from screen NTM ~$211
8×	$92	48%	56%	57%
10×	$73	58%	65%	65%
12×	$61	65%	71%	71%
15×	$49	72%	76%	77%

The mechanism is the trailing multiple, not the forward one. At $733, Micron is 9.7× fiscal 2026 EPS of about $75. [17] A 10× through-cycle buyer is paying for earnings to settle back at roughly the year just reported. The path from there to $177 or $208 is what the 3.5–4× forward P/E is refusing to capitalize. That is a normal cyclical discount, not a forecast of a 2023-style loss. A loss year (trough EPS at or below zero) is a worse outcome than any cell in this table. The September sell-side path that had earnings peaking near $171 and then “normalizing” around $50 is stale relative to a $177–$208 consensus, but $50 is exactly the 15× cell — the top of the memory multiple range, applied to today’s price. [19]

Two things keep this from being a simple “cheap at 4×” call. Gross margin in the August quarter was 86.8% and the operating margin was 80.7%, against a fiscal 2018 gross margin near 59%. [20] [21] Earnings can fall further than revenue if price, not just volume, mean-reverts. And the forward denominator is not money earned yet: the company guided the November quarter to about $38 of EPS, so the ramp is still being printed, but a full year at that run-rate is ~$150, not $208. [20]

SK hynix’s ADR tells the same story with a slightly deeper cut. At $134.38 and a forward EPS near $37, a 10× through-cycle multiple leaves durable EPS of about $13.40, a 64% decline from the forward number; at 8× the decline is about 55%. [13] Korean local shares have screened even cheaper on 2027 earnings (figures around 2.3–2.8× in late October and early November), which would imply a still larger giveback at the same through-cycle multiple. [22] [23] Those local multiples are not mixed into the ADR math above.

Nvidia, Broadcom, TSMC — the platform names

Name	Multiple	Durable EPS	Fall if FY/CY2027 street is the peak	Fall if 2028 street is the peak
Nvidia	20×	$8.93	4% from $9.30	43% from $15.80
Nvidia	22×	$8.12	13%	49%
Nvidia	25×	$7.15	23%	55%
Broadcom	16×	$20.0	about zero vs $19.47 (price is ~16× FY2027)	34% from $30.37
Broadcom	18×	$17.7	9%	42%
Broadcom	20×	$16.0	18%	47%
TSMC	16×	$24.3	price is ~10% above 16× the $22.2 estimate	14% from $28.2
TSMC	18×	$21.6	3%	23%
TSMC	20×	$19.4	12%	31%

Read down the 2027 column and the ranking is obvious. TSMC’s price is 2027 earnings at a normal multiple. A buyer at 18× needs essentially no decline if 2027 is the peak, and only about a quarter given back if 2028 street EPS near $28 is the peak. Broadcom’s price is FY2027 at 16–18×. Hock Tan has talked about AI semiconductor revenue doubling to about $115 billion in fiscal 2027 and doubling again to about $230 billion in fiscal 2028; the $30 EPS path is that second double. [24] The stock does not pay for it persisting. It pays for the first double, then a giveback of roughly 35–45%. Nvidia’s price is FY2027 at about 19×, which is a market multiple on $9.30, not a distressed one. The $15.80 FY2028 figure is in the price only if something like half of it does not last and you still pay 20–25× what remains. Consensus has that $15.80 rising, not peaking — one late-October note had the fiscal 2028 estimate at $15.91, up from $12.81 ninety days earlier — so the market is fading a number that analysts are still marking up. [25]

Sensitivity grid: what peak each price implies

Convention for this grid: trough fraction τ is applied to peak-year EPS, the through-cycle multiple is applied to that residue, and the resulting value is discounted back from the peak year at 10% with no interim cash. Implied peak EPS = price × (1.10)^{year − 2026} / (multiple × τ). The “now” column is the undiscounted case, which is the right comparison to a peak you think is imminent.

Where implied peak sits below the street number for that year, the price is justified even under that trough — the street peak is higher than the peak the price needs. Where implied peak sits above the street, the price needs either a higher peak, a shallower trough, or a higher multiple than the one in the header.

Micron — 10× through-cycle. Street reference: FY2027 $177, FY2028 $208.

Trough as % of peak	Implied peak, now	Peak in 2027	Peak in 2028	Peak in 2029	Peak in 2031
50%	$147	$161	$177	$195	$236
40%	$183	$202	$222	$244	$295
35% (base)	$209	$230	$253	$279	$337
25%	$293	$323	$355	$390	$472
10%	$733	$806	$887	$976	$1,181

The cell that matches the street is the point. FY2028 consensus of $208 is almost exactly the undiscounted 35% row ($209) and sits next to the 40% / peak-2027 cell ($202). So if your through-cycle multiple is 10×, today’s price is the market saying: the peak is the 2027–28 consensus, not a 2029 or 2031 blow-off, and roughly 35–40% of that peak is what is left afterward. A 25% retention — closer to the 2018-to-2020 experience — requires a peak of about $293 now, or $355 if the peak is 2028. That is 40–70% above the street. The price is not underwriting a 2023-style loss, and it is not underwriting a peak in 2031 unless peak EPS is several hundred dollars.

Move the multiple and the whole grid scales by 10/M. At 8× every implied peak is 25% higher; at 12×, 17% lower; at 15×, 33% lower. At 15× and a 35% trough, the undiscounted implied peak is about $140, which is below FY2027 street. That is the only memory assumption set in this note under which $733 is outright cheap versus consensus rather than a fair price for a normal fade. It requires you to pay 15× for cyclical earnings, the top of the range, on a business that has already printed an 87% gross margin.

Nvidia — 25×. Street reference: FY2027 $9.30, FY2028 $15.80.

Trough as % of peak	Now	2027	2028	2029	2031
70%	$10.2	$11.2	$12.4	$13.6	$16.4
50% (base)	$14.3	$15.7	$17.3	$19.0	$23.0
35%	$20.4	$22.5	$24.7	$27.2	$32.9

FY2028 street of $15.80 sits between the undiscounted 50% cell ($14.3) and the peak-in-2027 50% cell ($15.7). The market’s assumed peak is about the fiscal 2028 consensus, with half of it surviving, capitalized at 25× — or, equivalently, fiscal 2027 earnings surviving almost intact at 20×. A 2031 peak is in the price only if that peak is ~$23 at a 50% retention, or ~$16 if 70% survives. Nobody is publishing a clean 2031 number to test that against. Drop the multiple to 22× and every implied peak falls by 12%; the 50% undiscounted peak becomes $13.0, and FY2028 street then sits 22% above what the price needs. That is the bull case: not a higher multiple, a belief that more than half of $15.80 is durable.

Broadcom — 18×. Street reference: FY2027 $19.47, FY2028 $30.37.

Trough as % of peak	Now	2027	2028	2029	2031
70%	$25.4	$27.9	$30.7	$33.8	$40.8
55% (base)	$32.3	$35.5	$39.0	$43.0	$52.0
40%	$44.4	$48.8	$53.7	$59.1	$71.5

The match is tight and specific. FY2028 street of $30.37 is the 70% retention, peak-in-2028 cell ($30.7). The undiscounted retention that equates $319 to 18× FY2028 EPS is 58%. FY2027 street of $19.47 does not appear anywhere in the grid except as “almost no decline”: you need about 91% of FY2027 to survive to justify $319 at 18× with no wait. The price implies the market’s peak is fiscal 2027 earnings treated as durable, or fiscal 2028 earnings with roughly 40% given back. It does not imply a 2029 peak near $40-plus, and it does not imply the second AI-revenue doubling sticking at a mid-teens multiple. That second doubling, held and capitalized at 18×, would be a stock near $550, not $319.

TSMC — 18×. Street reference: 2027 $22.2 (range to $23.4), 2028 $28.2 (range to ~$30).

Trough as % of peak	Now	2027	2028	2029	2031
85%	$25.4	$27.9	$30.7	$33.8	$40.9
70% (base)	$30.8	$33.9	$37.3	$41.1	$49.7
50%	$43.2	$47.5	$52.2	$57.5	$69.5

2027 street is below every cell. You do not need a trough model to justify $389 if 2027 EPS is $22–23 and you will pay 18× for it; fair value on that identity is about $400–$420 before any further growth. 2028 street of $28 is a 23% fade at 18×, which lands between the 85% row and the 70% row. A peak in 2029 or 2031 is not in the price unless EPS gets to the mid-$30s or high-$40s. One local broker has talked about 2028 EPS above NT$170, which is about $26 per ADR at the November 9 rate — inside the street range, not above the grid. [26]

The TSMC caveat that the P/E grid does not capture: Deutsche Bank has lifted 2027 and 2028 capex forecasts to $84 billion and $100 billion after management said it would leave no AI demand unmet. [8] EPS can keep rising while free cash flow does not. A through-cycle P/E on earnings overstates the value if the capital intensity of the “trough” year is still a build-out year.

What each price implies about the market’s assumed peak

Put the four implied peaks in one sentence each.

• Micron at $733 assumes the peak is the 2027–28 consensus, on the order of $180–$210, and that something like $60–$90 of EPS is what remains. That residue is still about 8–12× the fiscal 2022 peak of ~$8. The market has accepted that this cycle’s level is different and is arguing about duration. It is not paying for a peak in 2029 or 2031. Supply is the tell on duration: the next Idaho fab’s first output has been described as mid-2027, with a second fab later in 2028, which is why a peak year of 2028 or 2029 is the fundamental debate, not 2031. [27] Customer agreements change the left tail — management has described non-cancellable deals with defined price floors running toward 2030 — but they do not change the identity. They change τ. [19] [28]
• Nvidia at $179 assumes a peak near fiscal 2028 street earnings, with about half given back, or a peak that is basically fiscal 2027. It does not assume the $15.80 becomes the new baseline at a 25× multiple. That would be a stock closer to $400.
• Broadcom at $319 assumes the same shape one year earlier in the narrative and one year later in the product cycle. Fiscal 2027 is the durable number. Fiscal 2028 is a peak only if a large minority of it does not last. The announced path to $230 billion of AI revenue is a fundamental forecast the price has not underwritten.
• TSMC at $389 assumes 2027 is not a peak in any memory sense. The market’s “assumed peak” is next year’s earnings, held. A later peak is upside the price does not require and does not pay for. The risk is capex and geography, not a 60% earnings air-pocket.

The early-November drawdown did part of this work already. From the October 6 area, Nvidia was down on the order of 20% and Micron on the order of 30% by the first week of November, and the November 9 closes sit 26% and about 42% under the 52-week highs of $243 and $1,255. [29] [11] An October 27 snapshot had Nvidia at $218, Micron at $929, TSMC at $437, and Broadcom at $342. [30] At those prices the same 10× memory identity required durable Micron EPS of $93 rather than $73, and Nvidia at 22× required durable EPS of $9.9 rather than $8.1. The selloff did not create the cyclical discount in memory. It widened it. It did, for Nvidia and Broadcom, pull the price down from “paying for a slice of the 2028 ramp” toward “paying for 2027 only.”

How to use this if you are underwriting the stocks

Treat the forward P/E on Micron, SK hynix, and any other memory name as a statement about τ, not as a bargain signal. The question to underwrite is whether post-cycle Micron can earn $50, $75, or $90, and in which year you are forced to find out. If the answer is $75 in 2029, the stock at $733 and 10× is roughly the undiscounted fair value of a 35% retention of the 2028 street peak, and the upside is duration — contracts and a late fab pushing the peak past 2028 — not a rerating to 15× on peak earnings. If the answer is a loss year, every cell in the memory grid is wrong in the bullish direction, and the historical precedent is fiscal 2023, not fiscal 2019.

For Nvidia and Broadcom the argument is not the multiple. Both already trade at a through-cycle multiple on next year’s consensus. The argument is whether fiscal 2028 is a peak or a base. The price says peak. You only need a reverse-valuation model if you disagree with that and want to know how much of the 2028 number you must believe is permanent. For Broadcom that hurdle is high: believing the $30 sticks at 18× is a very different stock from the one at $319.

For TSMC the cycle framework is the wrong tool at this price, unless you think 2027 EPS is itself a peak. The tool that matters is free cash flow after $80–100 billion of capex. An 18× P/E on rising EPS can coexist with a much fuller multiple on the cash that is left.

Three things would move the grid without any change in philosophy. A Micron print that makes fiscal 2027 look like $210 rather than $177 — the screen is already hinting at that — lowers the decline embedded in a flat price, which is how memory stocks get more expensive while the forward P/E falls. A Nvidia or Broadcom guide that pulls fiscal 2028 EPS back toward the durable residual ($7–9 and ~$18) would mean the price had been right and the street had been capitalizing a peak. A TSMC capex number that keeps climbing after EPS flattens would break the “18× 2027 is fair” conclusion without any change in the earnings multiple at all.

---

Recent Findings Supplement (November 2026)

Recent semiconductor earnings guidance, market forecasts, and valuation resets (post-May 2026) provide fresh inputs for a reverse-valuation framework on AI-exposed names like Nvidia (NVDA), Broadcom (AVGO), TSMC (TSM), and memory players (Micron/MU, SK Hynix). These updates point to stronger near-term peaks (pushed into 2028 or later), sustained high margins from HBM and advanced nodes, but also compressed multiples reflecting competition and potential later-cycle normalization.[1][2]

Nvidia’s August 2026 results and outlook signal that the market is pricing in durability well beyond 2027 while applying lower through-cycle multiples than historical AI-era averages. The company guided ~70% revenue growth for fiscal 2028 (ending ~Jan 2028), with Q3 FY2027 revenue forecast at $108 billion; it highlighted persistent memory supply constraints and customer forecasts implying even faster growth if supply allowed. Forward P/E compressed to ~16.5x (lowest since 2015 per LSEG), with trailing multiples around 22–23x in early November amid broader sector rotation. A record $150 billion buyback increase (total capacity $235 billion through FY2028) underscores management confidence in sustained cash generation.[1][2][3]

• AMD’s November 2026 guidance (overall revenue +>80% in 2027, data-center +97% in Q3) reinforces broad AI accelerator demand and eases fears of an abrupt slowdown, indirectly supporting higher peak assumptions for Nvidia-like players.[4]
• These figures imply that current NVDA prices embed assumptions of peak earnings in 2028–2029 (or later) with only moderate troughs (e.g., 60–80% of peak) or elevated through-cycle P/Es (20x+ rather than 15–18x), given the explicit multi-year roadmap including Vera Rubin ramp and hyperscaler deployments.

Memory leaders reported record results and locked in multi-year visibility via long-term agreements (LTAs), shifting trough assumptions higher and supporting elevated cycle P/Es for the segment. Micron’s Q4 FY2026 (ended ~Sep 2026) delivered $54.2 billion revenue (record, + strong sequential ASP growth in DRAM/NAND), 87% gross margin, and $33+ billion FCF; consensus now sees FY2027 revenue roughly doubling to ~$277 billion. SK Hynix Q3 2026 posted >100 trillion KRW revenue (~first time) and 80% operating margin, driven by HBM4 ramp (priced ~50% above HBM3E) and LTAs covering >50% of 2027–2028 capacity; analysts note shares trading at crisis-era ~3x 2027 EPS despite these commitments.[5][6][7][8]

• Memory revenue is now expected to dominate 2026 industry totals (~54% per Gartner), with DRAM/NAND ASPs surging on AI/HBM focus and supply tightness persisting into 2027.[9]
• For reverse valuation, this data supports using higher trough percentages (70–90% of peak) or later peak years (2028–2029) for memory names, as LTAs and HBM premiums reduce cyclical downside versus historical 8–15x through-cycle P/E assumptions.

TSMC raised 2026 growth targets and highlighted structural under-supply in advanced nodes, implying its current price assumes sustained high utilization and pricing power into the late 2020s. Q3 2026 revenue reached $46.7 billion (+41% YoY), with full-year 2026 growth outlook lifted to ~44%; HPC/AI accounted for 68% of revenue. 2nm contributed 8% of wafer revenue (ramping), with 2026 capex guided at $60–64 billion and further increases planned. Forward P/E sits near 16x, below its 5-year average.[10][11][12]

• Analysts project EPS more than doubling from 2025 to 2027 levels, with 2nm capacity doubling in 2027.
• This supports sensitivity grids showing that TSM’s price justifies peak earnings around 2028–2031 with modest normalization (troughs at 65–80% of peak) under 18–22x through-cycle P/Es, rather than sharper cuts.

Industry forecasts were revised sharply higher in mid-to-late 2026, extending the implied earnings peak and reducing the severity of any trough needed to justify current valuations. Gartner (Aug 2026) sees global semi revenue at $1.555 trillion in 2026 (+92% YoY) and $1.94 trillion in 2027, with memory at $837 billion (2026) and >$1 trillion (2027); AI data-center share rising to 53% by 2030. Omdia (Nov 2026 updates) raised 2026 growth forecasts to 63–94% range on memory strength. McKinsey (Sep 2026) lifted its 2030 projection to $2.3 trillion (servers/data centers to $1.2 trillion). Hyperscaler capex plans total ~$800 billion (2026) and $1 trillion (2027).[13][9][14]

• Other outlooks (Yole, TechInsights) cluster around $1.8T+ for 2026 and $2.2–2.6T+ for 2027, confirming memory and advanced logic as the primary drivers.[15][16]
• These revisions mean sensitivity grids for peak years 2027–2031 would show current prices embedding later peaks (2028–2031) or higher trough ratios, especially if through-cycle P/Es for leaders are held at 15–20x+ versus lower memory benchmarks.

Valuation compression across the group (NVDA ~16–23x, TSM ~16x, SK Hynix low single-digit forward in some broker views) alongside persistent supply tightness and China-related guidance creates a narrower band for implied earnings downside. Recent notes highlight NVDA instructing cloud clients to limit China exposure and broader concerns over competition/custom silicon, yet buybacks and LTAs signal resilience.[1][17][18]

For a reverse-valuation framework, these developments collectively suggest updating assumptions toward later peak years (favoring 2028–2029 grids), higher trough percentages (especially memory via LTAs), and modestly elevated through-cycle P/Es for leaders given structural AI demand—while stress-testing 2031 scenarios for normalization risks. New data on specific growth rates, capex, and market sizes directly inform the “market’s assumed peak” implied by current prices.

Memory stocks have repeatedly looked cheapest on earnings multiples at the cycle top, because the share price turns months before profits do. Telecom and fibre names in 2000, and Nvidia in late 2021, peaked the other way: on very high multiples, with subsequent price damage driven more by multiple compression (and, for fibre, a real revenue collapse) than by the “low P/E trap” that defines memory.

Memory: stocks peak before earnings, so trailing P/E is a warning at the top

Morgan Stanley’s historical count, reported in secondary coverage of a client note, is that memory share prices peaked about four months before earnings in the 2008 and 2012 cycles and about eight months before earnings in 2018. Korean market commentary on the 2017–18 supercycle matches that pattern: Samsung Electronics’ operating profit peaked in the third quarter of 2018, but the stock had already made its high in November 2017; SK hynix still posted a then-record 2018 operating profit while the shares fell more than 40% from that year’s peak.

Micron is the cleanest public case.

• 2018. The stock’s intra-year high was about $62.62 on 29 May 2018; the trough was about $29.02 on 24 December 2018, a peak-to-trough drop of roughly 54% in about seven months. By early August 2018, with the memory-price peak still close, trailing P/E was about 5.2x — the bottom of its prior five-year range — and the shares then fell about another third into mid-2019. Fiscal 2018 (year ended August 2018) net income was $14.1 billion. That fell to $6.3 billion in fiscal 2019 (about −55%) and $2.7 billion in fiscal 2020 (about −81% from the peak) over roughly 24 months. The stock bottomed in late 2018, well before the earnings trough in 2020. Price-to-book near the May 2018 price high was only about 2.2x; by November 2018, after the price collapse and with book value still rising, P/B was about 1.2x. Year-end 2018 EV/EBITDA sat near the bottom of Micron’s historical range (sources put the 2018 reading around 1.7–4.5x depending on the exact window), because the price had already fallen while EBITDA was still near a peak.
• 2021–22. Micron’s calendar-2022 high was about $97 in mid-January 2022; the low was about $49 in late September 2022, a drop of roughly 50% over about eight months. Fiscal 2022 net income was $8.7 billion. Fiscal 2023 was a loss of $5.8 billion — a swing from a large profit to a large loss in one fiscal year — and revenue nearly halved to $15.5 billion, with gross margin collapsing from the mid-40s to about −9%. Industry DRAM pricing had already started to crack from late 2021; by early 2023 DRAM prices were down on the order of 50%+ from the prior peak. The shares bottomed in late 2022 while losses were still widening into 2023, then recovered as the cycle turned — the mirror image of the “cheap P/E” signal at the top. At the August 2022 window, trailing P/E was still only about 7x even as the downturn was underway.
• 2000. Public point-in-time multiples are thinner. Micron’s year-end trailing P/E was extremely high in 1999 (on the order of 150x in one long-run series) and had already compressed into the low-to-mid teens by the end of 2000 as earnings surged into the boom and the price rolled over. The subsequent DRAM bust produced losses, not just slower growth. The same lead-lag — price turning before the earnings peak — is the pattern analysts cite for that era as well, but the exact month gap is less cleanly documented in free sources than for 2008, 2012, and 2018.

SK hynix and Samsung show the same valuation signature. Morningstar-style calendar data put SK hynix’s 2018 price/earnings near 2.7x and EV/EBITDA near 1.4x — classic peak-earnings, already-falling-price readings — versus roughly 9x P/E and about 4x EV/EBITDA in 2021, before the 2022–23 collapse. Samsung’s group operating profit set a record in 2018 (about KRW 59 trillion) and then fell sharply in the fourth quarter of 2018 as memory demand rolled over; the stock had peaked the prior year.

What this means for anyone comparing today’s quotes to “cheap memory.” A 3–6x forward P/E on Micron, SK hynix, or Samsung is not evidence the cycle has been de-risked. It is the historical signature of a market that already expects earnings to fall, and that has often been right about the direction even when it has been wrong about the depth.

Current memory multiples sit in that historical “cheap at the top” band

As of early-to-mid November 2026 (prices moved sharply through the autumn, so forward multiples are date-sensitive):

• Micron. Trailing P/E about 9.3x and EV/EBITDA about 6.8x around 6–9 November 2026, with price-to-book about 5.6x. Mid-October commentary put the forward P/E near 5.6x after a pullback from the June 2026 high (about $1,213–$1,255). Price-to-book is not cheap versus Micron’s own history: the 2018 stock peak was closer to 2x book, and cycle troughs have traded near or below 1x.
• SK hynix. One U.S.-listed quotation series in late October / early November 2026 showed trailing P/E about 6.9x, forward P/E about 3.6x, price-to-book about 2.7x, and EV/EBITDA about 3.4x. Korean forward P/E readings in mid-October were cited around 3.2x.
• Samsung Electronics. 12-month forward P/E was cited near 3.6x in mid-October 2026, below the prior decade low around 6x in October 2018. Group multiples are diluted by non-memory businesses, so the memory franchise itself is even more peak-cycle in character than the headline P/E.

A October 2026 industry note framed the three suppliers at roughly 5.4–6.8x run-rate earnings and argued that, on a mid-cycle margin assumption, the market was already discounting a downturn larger than the last two cycles. That is the mechanism: low forward multiples are the market’s way of saying “these earnings will not last,” not a free option on the peak persisting.

Cisco and the 2000 telecom/fibre complex: high multiples, then a price collapse that earnings only partly justified

Cisco is the opposite of memory. At the 27 March 2000 close of $80.06, market value was about $555 billion and the stock was the world’s most valuable company for a moment. Contemporaneous and later reconstructions put the trailing P/E above 200 and enterprise value-to-sales around 31. Price-to-free-cash-flow was cited near 176. From that close to 11 October 2002 the shares fell about 87% (low near $10).

Earnings did not fall 87%. Fiscal 2000 revenue was $18.9 billion; fiscal 2001 revenue still rose 18% to $22.3 billion. Pro forma net income fell about 21% (from $3.91 billion to $3.09 billion). GAAP swung to a loss because of acquisition and restructuring charges. The fourth quarter of fiscal 2001 was the real air pocket: sales down 25% year on year and pro forma profit down about 86%. Over the following two decades revenue and profits grew substantially — later accounts describe revenues nearly quintupling and profits roughly quadrupling from the late-1990s base — but a buyer at the March 2000 close waited until December 2025 for a higher closing price. The stock peaked with the growth narrative, not months ahead of a known earnings collapse, and the subsequent return was almost entirely multiple compression. Today’s Cisco (early November 2026) trades nearer 30x trailing earnings and about 22x EV/EBITDA — a fraction of the 2000 multiple, even if the absolute price has only recently reclaimed the old high.

Fibre and telecom equipment was worse, because the earnings and revenue did collapse.

• Mid-2000 forward P/Es, from contemporaneous screens, were on the order of 60x for Corning, 74x for Nortel, 161x for JDS Uniphase, and well over 100x for several component peers. Trailing multiples on optical names were often higher still (press accounts cited JDS Uniphase and SDL in the 160–400x zone).
• Market-cap destruction from mid-2000 peaks into early 2001 was already enormous (Nortel from roughly $366 billion to about $82 billion in one published snapshot; JDS Uniphase from about $61 billion to about $11 billion). Peak-to-trough share-price declines of 80–95% were typical across the group. An index of fibre heavyweights was down about 78% from July 2000 into the following year in one contemporary tally, wiping out more than $1 trillion of market value.
• Revenue followed. JDS Uniphase quarterly sales fell from about $920 million in the year-ago quarter to about $262 million by early 2002 (roughly −70%), and the company posted multi-billion-dollar losses dominated by goodwill write-downs. Carrier capex, not a gentle inventory correction, was the driver. Many of these stocks never regained their 2000 highs. Lucent and Nortel were effectively destroyed as independent investment stories.

Implication versus memory. In 2000 the dangerous signal was a triple-digit P/E on a picks-and-shovels story whose customers were overbuilding. In memory the dangerous signal has been a single-digit P/E while pricing and margins are still at a peak. Using Cisco’s 200x peak as the benchmark for “is memory expensive?” answers the wrong question.

Nvidia in 2021–22: a high-multiple growth peak, a deep drawdown, then earnings that more than grew into the old price

Nvidia’s late-2021 peak is closer to Cisco than to Micron, with one crucial difference: the earnings disappointment was a slowdown, not a multi-year bust, and the subsequent upturn dwarfed the prior peak.

• Around the November 2021 high (split-adjusted prices in the low-to-mid $30s in post-2024 split terms; unadjusted quotes near $330–$346), trailing P/E was about 90x at year-end 2021 in one quarterly series, and Morningstar’s calendar-2021 readings were about 91x price/earnings, about 57x forward earnings, about 31x price-to-book, and about 74x EV/EBITDA. Price-to-sales was about 31x.
• The shares then fell about 66% from the November 2021 high to the October 2022 low. Gaming revenue went negative year on year through 2022; total revenue growth decelerated from 50%+ to flat-to-down. Fiscal 2023 (ended January 2023) was the digestion year. The stock bottomed before the AI demand shock showed up in results.
• Earnings then reaccelerated so hard that today’s trailing multiple is far below the 2021 peak even though the share price is many times higher. As of early November 2026, trailing P/E was about 23x and EV/EBITDA about 21x, with price-to-book still elevated (high teens) but well under the 2021 reading near 30x. The 2021–22 episode was a multiple and growth-rate reset, not a memory-style profit collapse from peak to loss.

Cisco’s 2000 buyers were paid in earnings growth that never caught the entry multiple. Nvidia’s 2021 buyers who held through the 66% drawdown were eventually paid by an earnings explosion the 2021 price had not fully anticipated — but only after a drawdown that would have forced most of them out.

Comparison table

Figures are approximate, drawn from the sources above, and are not perfectly contemporaneous (stock-price peak vs fiscal-year-end vs “near the earnings peak”). Forward multiples at historical peaks are especially sparse outside memory, where the low forward P/E was the story. “n.a.” means a clean public figure was not found in the sources retrieved.

Name / episode	Approx. stock peak	Trailing P/E near peak window	Forward P/E	P/B	EV/EBITDA	Earnings after the peak	Price vs earnings turn
Micron, 2018	May 2018, ~$63	~5x by Aug 2018	Low-single-digit into the peak (street was already using ~6x FY2018)	~2.2x at the May high	Very low once price fell; ~2–5x range in 2H18	FY18 NI $14.1B → FY20 $2.7B, about −81% over ~24 months	Stock peaked ~months before earnings; −54% by Dec 2018, bottomed before the 2020 earnings trough
Micron, 2021–22	Jan 2022, ~$97	~7–10x into 2022	Compressed as estimates fell	~1.2–2.0x through 2022	~3x by FY2022 year-end as price had already fallen	FY22 NI $8.7B → FY23 loss $5.8B in ~12 months; revenue about −50%	−50% by Sep 2022; price bottomed while losses were still widening
SK hynix / Samsung, 2018	Samsung stock high Nov 2017; Hynix rolled over in 2018	Hynix ~2.7x in 2018; Samsung group similarly compressed	Forward P/Es in the mid-single digits at the late-2018 earnings peak	Hynix ~0.9x in 2018	Hynix ~1.4x in 2018	Samsung OP peaked 2018 then fell; memory OP down sharply from 3Q18	Stocks led earnings by roughly 4–8 months; 40%+ drawdowns while annual profits were still at records
SK hynix / Samsung, 2021–22	2021 highs, then 2022 decline	Hynix ~9x in 2021, then lower	Already discounting a downturn into 2022	Hynix ~1.5x in 2021, ~0.9x in 2022	Hynix ~4x in 2021, ~2.4x in 2022	Industry revenue down ~30–40% peak to 2023 trough; suppliers cut utilization	Same lead: prices fell in 2022 before the worst losses printed in 2023
Memory, 2000	Late-1990s / 2000 boom	Micron year-end 1999 very high (~150x in one series); end-2000 already ~14x	n.a. cleanly	Elevated into the boom, then compressed	n.a. cleanly	Subsequent DRAM bust produced losses	Price rolled over as earnings were still rising into the peak; less precisely dated than 2018
Cisco, Mar 2000	27 Mar 2000, $80.06, ~$555B	>200x	Still triple-digit on near-term estimates into early 2000 (one Jan 2000 quote ~109x FY2000)	n.a. at the peak in sources used	EV/Sales ~31 (EV/EBITDA not the metric used then)	FY01 revenue still +18%; pro forma NI −21%; GAAP loss on charges. Q4 FY01 sales −25%, pro forma profit −86%. Long-run earnings grew	−87% by Oct 2002. Price peak and narrative peak coincided. Took ~25 years to reclaim the high despite multi-fold earnings growth
Fibre / telecom equipment, 2000 (JDSU, Corning, Nortel, Lucent)	Spring–summer 2000	Often 100–400x where earnings existed	~60–170x on 2001 estimates in Aug 2000 screens	n.a. consistently	Extreme on sales; many names had little sustainable EBITDA	Revenue −50% to −70%+ within 12–24 months; multi-billion goodwill losses	Prices and fundamentals collapsed together. 80–95% drawdowns; many never recovered
Nvidia, Nov 2021	Nov 2021	~90x at end-2021	~50–60x forward	~31x	~70–75x	Growth slowed sharply in FY2023 (gaming down hard); not a peak-to-loss collapse	−66% to Oct 2022. Price bottomed before the AI earnings reacceleration. Subsequent earnings growth more than absorbed the old multiple
Micron, now (Nov 2026)	—	~9x	~5–6x (mid-Oct; price-sensitive)	~5.6x	~7x	Cycle still in the boom; FY2026 earnings far above any prior peak	Forward multiple is in the historical “peak-cycle” zone; P/B is above prior cycle peaks
SK hynix, now	—	~7x	~3.2–3.6x	~2.7x (U.S. line)	~3.4x	Same	Cheapest of the three on forward earnings; historically that has been the danger zone, not the all-clear
Samsung, now	—	Group trailing higher than pure memory	~3.5–4x forward (mid-Oct)	Group P/B well above 2018 troughs	Group EV/EBITDA mid-single digits in recent models	Memory is the swing factor inside a conglomerate	Forward P/E below the 2018 low that preceded the last bust
Nvidia, now	—	~23x	Mid-teens on published forward figures	High teens	~21x	Earnings have grown through the multiple	Far below the 2021 peak multiple because earnings outran the 2022–26 price rise
Cisco, now	—	~30x	~19x	~7–8x	~22x	Mature growth	Multiple is a small fraction of March 2000; absolute price only recently exceeded that peak

How prices behaved relative to earnings turning points

Three distinct mechanisms show up in the record.

1. Memory (2018, 2021–22, and the earlier cycles Morgan Stanley dated). The share price peaks first — on the order of four to eight months before earnings — because the market stops extrapolating peak margins as soon as pricing inflects or inventories build. Trailing P/E therefore prints its lowest readings near the earnings peak, which is the worst time to treat a low multiple as a buy signal. The stock often bottoms before, or around, the earnings trough, so the highest trailing P/Es (or “n/m” when earnings go negative) have coincided with the best subsequent returns. August 2018 Micron at 5x, followed by a further large decline, and June 2023 Micron at a very high P/E because earnings had collapsed, followed by a sharp rally, are the textbook pair.

2. Cisco 2000. The price peaked with the enthusiasm, not ahead of a visible earnings cliff. Near-term operating earnings dipped (pro forma profit −21% in the first down year; one quarter −86%) but the business did not go through a memory-style profit-to-loss cycle. Almost all of the 87% drawdown was the multiple. Long-run earnings growth of high-single to low-double digits compounded was not enough to bail out a 200x entry for a generation.

3. Fibre 2000 and Nvidia 2021, as the two tails. Fibre combined Cisco-like entry multiples with a genuine demand collapse (carrier capex), so both the multiple and the earnings base were destroyed. Nvidia combined a high entry multiple with a cyclical slowdown in gaming, then an earnings regime change in data-center AI. The 66% drawdown happened before that regime change was in the numbers; holders who needed the earnings to “prove it” before the bottom did not get that proof in time.

The direct answer on the question’s last point: yes. Memory stocks have historically peaked before earnings. The lag cited for the best-documented cycles is about four months (2008, 2012) to about eight months (2018), and the 2021–22 tape rhymed with that lead. Today’s 3–6x forward P/Es on Micron, SK hynix, and Samsung are in the same zone those stocks occupied when the last two peaks were already behind the share price and still ahead of the earnings peak. That is not a forecast that this cycle must end the same way — HBM mix, multi-year supply agreements, and the scale of AI demand are real differences — but it is the historical base rate the multiples are being compared against. Nvidia and Cisco at their famous peaks were expensive on a completely different metric: they were priced for many years of uninterrupted growth, and the subsequent path depended on whether that growth showed up (eventually, for Nvidia after a deep drawdown; not at the required rate, for Cisco) rather than on whether peak-cycle margins mean-reverted.

---

Recent Findings Supplement (November 2026)

Recent analyst reports (primarily Oct-Nov 2026) from firms like Bank of America, Morgan Stanley, Nomura, and others highlight a sharp de-rating in memory stock valuations amid a sell-off, even as fundamentals (earnings, margins, and supply tightness) remain strong or improving. These notes reference historical cycle patterns but do not provide exhaustive new datasets on exact 2000/2018/2021 peak multiples across all requested metrics (P/E, P/B, EV/EBITDA) for memory makers or comparables like Nvidia, Cisco, and telecom/fiber names.[1][2][3]

No major new regulatory/policy updates, comprehensive historical research publications, or revised peak-trough earnings/price datasets appear in post-11/10/2025 sources. Focus is on current (2026) multiples versus implied historical norms, with qualitative nods to past behavior.

Current memory stock multiples (as of Oct-Nov 2026) sit at or below prior upcycle ranges, driven by a 20-58% sell-off from June 2026 highs despite rising profits and multi-year sold-out supply.[4][5][6]

• SK hynix and Samsung trade at ~2.5-3.7x 2027E or next-year forward P/E (some reports cite 2.7-3.0x 2027E / 2.3-2.8x 2028E); Micron at ~4.8-5.2x forward P/E. These are 30-50%+ below the 4-7x observed in prior upcycles (e.g., pre-2001, 2008, 2018, 2022). P/B ratios near 1.0x 2027-28E book value for the Koreans, levels historically seen at cycle bottoms.[1][3][7]
• Run-rate or current earnings multiples for the three (Micron, SK hynix, Samsung) cluster in a ~5.4-6.8x peak-cycle range, implying investor expectations of significant earnings declines.[4]
• This follows a period of explosive earnings growth (e.g., Micron revenue +256% FY2025 to FY2026 in one account; Samsung Q3 op. profit surge). Stocks fell sharply in Oct 2026 (e.g., SK hynix -24% in one week, broader memory names down >50% from peaks for some), pushing multiples lower even as DRAM prices rose and supply remained constrained.[5][6]

Recent reports confirm the historical pattern that memory stock prices have often peaked before earnings (e.g., ~4 months ahead in 2008 and 2012; ~8 months in 2018 per Morgan Stanley).[2]

• Earnings declines post-prior peaks included Samsung op. profit halving (~2018-2019) or falling 85% (~2022-2023); Micron EPS dropping ~80% (FY2018 to two years later). Current expected EPS are multiples higher than at the 2022 peak.[5]
• Low forward multiples today are framed as pricing in downturns larger than recent cycles (hypotheticals in one note suggest 1.5-3.5x the size of prior ones under certain margin assumptions).[4]
• SK hynix disclosed production effectively sold out through 2029 (with shortages potentially persisting to 2030), supporting views that the cycle differs due to AI/HBM demand and long-term agreements, though capacity ramps are still flagged for 2027+.[8]

Nvidia and Cisco comparisons in recent commentary emphasize Nvidia's materially lower multiples versus Cisco's 2000 peak, with no new granular 2021-22 Nvidia peak data.[9][10]

• Cisco 2000 peak: ~200-234x trailing P/E; ~130x forward P/E.[9][11]
• Nvidia (as of late 2026): Trailing P/E ~22-35x range across reports; forward P/E ~11-14x (or ~13x in one snapshot); EV/EBITDA ~19x TTM. Some notes place it at ~40x forward in earlier 2026 context or note de-rating from higher prior levels.[12][13]
• Broader context: Nvidia's profitability (high margins, cash flow) and customer base (hyperscalers) are contrasted favorably with 2000-era dynamics, though overbuild risks are noted. No specific new multiples for telecom/fiber names at 2000 peaks beyond legacy references to extreme valuations (e.g., hundreds of x P/E for some fiber components).

Overall, recent sources position current memory multiples as unusually compressed relative to both historical upcycle norms (4-7x) and ongoing/expected earnings strength, with stocks appearing "cheapest at peak earnings"—consistent with past cycles where prices turned before fundamentals.[2][3] Direct apples-to-apples historical tables for all requested peaks/metrics are not newly published in these sources. For precise peak-trough quantification or 2000-era telecom/fiber details, older data or custom modeling would be needed. Current Nvidia multiples remain elevated versus broad market but far below 2000 Cisco extremes.

The post-Q3 2026 demand picture is real, but it is not one number. The hard commitments are multi-year volume contracts with deposits and price bands at the memory makers. The “sold out through 2029” and “$1 trillion visibility” lines mix those contracts with negotiations and customer forecasts. Hyperscaler capex levels are still rising into 2027; what is peaking, on Goldman’s reading, is the growth rate of that spending as a share of GDP.

Memory contracts are binding; the 2029 sellout is only partly so

Micron’s book is the cleanest contractual evidence in the industry. As of its September 30, 2026 fiscal Q4 call, it had 26 strategic customer agreements, up from 16 a quarter earlier, covering more than 35% of estimated revenue through 2030, with some deals and extensions running into 2031. All 26 include take-or-pay volume commitments. Remaining performance obligations tied to those agreements were about $150 billion, calculated at minimum prices, so recognized revenue over the life of the deals is expected to be higher. Customer financial commitments reached $32 billion, mostly cash deposits, and contract liabilities on the balance sheet were $12.9 billion, up from $568 million in May. Three-quarters of expected SCA revenue sits inside a defined pricing framework, mostly floor-and-ceiling bands; the rest is repriced periodically at market. Management said that even at floor prices, gross margins on the banded agreements would be meaningfully above prior-cycle quarterly peaks. Combining SCAs with other customer commitments, more than 75% of fiscal 2027 output was already secured. That is prepayment, volume obligation, and a price floor—not a verbal claim. [1] [2] [3] [4]

SK hynix’s October 28 call is stronger on volume coverage and weaker on enforceability disclosure. The company has signed long-term agreements with 14 key customers, including global cloud providers, up from 10 in Q2. Those LTAs are typically five years, cover more than 50% of planned 2027 and 2028 production capacity, and use a price range that reflects market conditions rather than a single fixed price. The company said it has included safeguards such as deposits so that purchase plans become orders and shipments. The headline—“DRAM and HBM volumes through 2029 are effectively sold out”—explicitly combines already-signed LTA volumes with demand still under detailed discussion. Management said the volumes under discussion are based on specific purchasing plans and have a high likelihood of materializing. That second half is not yet a contract. Separately, CFO Kim Woo-hyun projected that supply will lag demand through 2030 because of process complexity and fab lead times, and the company said customer volume requirements are more than double what it can supply today. Those are outlook statements, not take-or-pay. SK hynix has not disclosed a Micron-style remaining-performance-obligation dollar figure or a cash-deposit total. [5] [6] [7] [8] [9]

For anyone trying to underwrite the cycle, the distinction matters. Micron has put a floor under a minority of revenue (a bit over a quarter through 2030 once the market-priced quarter of SCA revenue is stripped out) and collected cash against it. SK hynix has locked more than half of 2027–28 capacity into volume agreements with deposit language and a price band, then described 2029 as sold out only after adding negotiations. Neither has contracted the entire industry’s output, and neither’s price mechanism fully freezes today’s spot price for the whole book.

Nvidia’s “backlog” is mostly other people’s plans, plus its own supply bets

Nvidia has not yet reported the quarter that covers calendar Q3. Fiscal Q3 FY2027 is scheduled for November 17, 2026. The latest official visibility statements are from the August 26 fiscal Q2 call. [10] [11]

Three different things are often collapsed into “Nvidia backlog”:

• Customer demand visibility. Jensen Huang has described “$500 billion of very high confidence demand and purchase orders” for Blackwell and Rubin through 2026, and “at least $1 trillion” of cumulative Blackwell and Rubin revenue from calendar 2025 through calendar 2027. CFO Colette Kress has said the company has “full confidence” in that $1 trillion figure. Investor materials treat this as a forward-looking estimate, not GAAP backlog, not recognized revenue, and not a fully non-cancellable order book. How much is shipped, firm, or still a customer plan is not broken out. [12] [13]
• Cloud industry backlog, which is not Nvidia’s order book. On the August call Kress said cloud industry backlog was greater than $2 trillion, and that capex by the top five hyperscalers was expected to reach nearly $800 billion in 2026 and $1.3 trillion in 2027. That is a statement about customers’ contracted cloud demand and their spending plans. [14] [15]
• Nvidia’s own upstream commitments. As of July 26, advance purchase orders and pledges to suppliers—dominated by high-bandwidth memory—stood at $279 billion, up from $119 billion a quarter earlier. Some of those supply commitments can be adjusted or canceled. This is Nvidia locking memory and components, not customers locking Nvidia. [16] [17] [18]

What is closer to a purchase order: production shipments of Vera Rubin began in August 2026, management said it already had purchase orders from every major hyperscaler, AI cloud, and system maker, and it expected Rubin to be about 20% of data-center revenue in fiscal Q3. AWS is deploying an additional 2 million Nvidia GPUs from fiscal Q3 2027 through fiscal Q2 2029. Management guided roughly 70% revenue growth in fiscal 2028 and said customers’ own forecasts imply faster growth, with supply—especially memory—the constraint. Huang has said demand is higher than the 70% figure. [19] [20] [18]

The implication for a competitor or a supplier: Nvidia’s revenue path is still gated by memory and power, not by a lack of stated customer intent. The $1 trillion and $2 trillion figures should not be modeled as firm, non-cancellable chip orders.

Capex levels are still going up; Microsoft’s lower guide is a timing slip

The four companies do not report capex the same way. Amazon guides cash purchases of property and equipment. Alphabet reports purchases of property and equipment. Microsoft and Meta include finance-lease principal in the figure they discuss as capex. None publishes a clean AI-only line. Adding the latest 2026 guides produces about $720 billion to $740 billion: Amazon about $220 billion, Alphabet $195–205 billion, Microsoft about $170 billion for the calendar year, Meta $135–145 billion. That is roughly 78% above combined 2025 spending of about $410–414 billion, depending on the compilation. It is a level comparison, not a claim that the growth rate is still accelerating. [21] [22] [23]

Company by company after the October prints:

• Amazon reaffirmed about $220 billion of 2026 cash capex. Q3 purchases of property and equipment were $77.3 billion, versus $35.1 billion a year earlier. On the October 29 call, CEO Andy Jassy said the company expects to spend “meaningfully more in 2027,” with a line of sight to returns, and that AWS can grow north of 40% in both 2027 and 2028. There is no 2027 dollar guide. A separate, earlier verbal claim—on an October 23 podcast—was that 2027 AWS capacity is “largely sold out” for the large blocks, with a good chunk of 2028 also spoken for, while some capacity is held back for on-demand EC2 and Bedrock. That is a utilization statement, not a disclosed take-or-pay schedule. [24] [25] [26]
• Alphabet held 2026 capex at $195–205 billion. Q3 purchases of property and equipment were $51.4 billion, versus $24.0 billion a year earlier. CFO Anat Ashkenazi said capex will “increase significantly in 2027,” with detail deferred to the Q4 call. Alphabet has also disclosed $924 billion of future purchase commitments; those are maximum contractual amounts, and the company says the final bill can be lower because of variable pricing and volume flexibility. [27] [21] [28]
• Meta narrowed 2026 capex to $135–145 billion from $130–145 billion, raising the floor. Q3 capex including finance-lease principal was $34.91 billion. CFO Susan Li said the 2027 budget is still being built and that the company expects to keep investing aggressively in the core business and AI. No 2027 number. [29] [30] [28]
• Microsoft is the only one that cut the calendar-2026 figure, and the cut is not a demand cut. The path was about $190 billion in April, about $175 billion on the July 29 call after more leases were classified as operating leases (Hood said underlying investment plans were unchanged), then about $170 billion on October 28 “due to the 3P capacity delays.” Fiscal Q1 capex including finance leases was $45.2 billion, below the more-than-$50 billion the company had expected; cash additions to property and equipment were $39.4 billion. Hood said some GPU-as-a-service deliveries slipped into the second half of Microsoft’s fiscal year (January–June 2027), that December-quarter capex should exceed $50 billion, and that spending should accelerate in that second half on CPUs, GPUs, higher memory and storage costs, and the shifted deliveries. She also said customer demand continues to exceed available supply. [31] [32] [33] [34] [35]

Quarterly cash spend is already outrunning operating cash flow at several of these companies. In calendar Q3, purchases of property and equipment (Microsoft: additions) were about $202 billion against operating cash flow of about $176 billion across the four, on one compilation. Alphabet’s quarterly free cash flow was negative $9.1 billion; Amazon’s trailing-twelve-month free cash flow was an outflow of $42.4 billion. The 2027 story from the companies themselves is “higher level, no number yet,” constrained by third-party data-center delivery and components, not by a stated intention to stop. [27] [36] [37]

Outside estimates of the 2027 level are higher still, and they are forecasts, not guides. Nvidia management’s “nearly $800 billion / $1.3 trillion” for the top five in 2026 and 2027 is one such statement. A Goldman research note reported in late September had the largest U.S. hyperscalers at $800 billion in 2026, $1.2 trillion in 2027, and $1.4 trillion in 2028, with the growth rate slowing even as the level rises. Moody’s has said capex at the five largest U.S. hyperscalers plus CoreWeave could approach $1 trillion in 2027. These ranges should not be stacked as if they measured the same set of companies. [14] [38] [39]

Hatzius is calling a peak in the growth impulse, not in the dollars spent

On October 26, before Amazon and several peers had reported, Goldman chief economist Jan Hatzius wrote that the growth impulse from the AI boom was set to peak in late 2026 or early 2027, “even before concerns around the sustainability of the AI investment boom intensified.” AI investment is still rising rapidly in dollars. What he expects to slow over the next year is the increase in that investment as a share of GDP. A secondary summary of the same note said the year-over-year change in hyperscaler investment as a share of U.S. GDP should peak in the fourth quarter of 2026 at about 1.5%. He also pointed to a fading boost from restored bonus depreciation and a cooling equity wealth effect (Goldman had put that wealth effect at about 0.25% of GDP over the past year). Goldman has separately put recent U.S. AI-related capex at roughly $600 billion, and 2026 AI investment at 1.9% of GDP in one related brief. [40] [41] [42] [43]

That is compatible with Amazon spending “meaningfully more” and Alphabet spending “significantly” more in 2027. A rise from, illustratively, $730 billion to something above $900 billion or $1 trillion is still a higher level. It is a slower growth rate, and a smaller incremental contribution to GDP growth, than the jump from roughly $410 billion in 2025 to roughly $730 billion in 2026. Hatzius has been consistent on the mechanism since at least September: even if the buildout is productive, an investment phase in which the volume rises sharply gives way to an exploitation phase in which investment growth slows or the level eventually falls. He has not said the level has peaked. [44] [44]

Apollo’s Torsten Slok, writing around the same weekend, made a different claim: “zero financial payoff to corporate America outside of tech,” with margins and revenue per worker up at the largest tech firms and flat for the rest of the S&P 500. That is a return argument, not a capex-level argument. It does not cancel the memory contracts. It does say the handoff from capex to broad productivity, which would be needed to justify another leg of spending growth, is not yet in the data. [42] [45]

Robot and physical-AI memory is a content story, not a contracted book

Two estimates are in the market, and they do not describe the same machine.

Micron, on the September 30 call, said Level 4 and higher autonomous vehicles typically carry more than 200 GB of DRAM and several terabytes of storage, and that humanoid robots are expected to have comparable content. It said physical AI can become a significant driver of DRAM and NAND demand by the end of the decade, and that several physical-AI customers are already sampling next-generation products. That is an analogy plus early sampling, not a volume contract and not a shipment forecast. [46] [47] [48]

Counterpoint Research, on October 11, projected DRAM demand for humanoid computing systems rising from about 8,300 terabits in 2026 to about 172,000 terabits in 2030, roughly 20.7 times, while average DRAM per humanoid only rises from 19 GB to 39 GB. Most of the growth in that forecast is more robots, not more memory per robot. One reading of those two Counterpoint figures implies on the order of 550,000 humanoids in 2030 and total robot DRAM equal to a few hundredths of one percent of 2026 industry DRAM output. Counterpoint also argued that battery limits favor LPDDR and NAND over HBM. [49] [50] [51] [52]

The gap is the trade. Nvidia’s own robot modules sit in between: Jetson Thor at 128 GB, a July T3000 module at 32 GB that Nvidia pitched in part as a way to reduce cost amid high memory prices, and an entry module at 16 GB. Hold Counterpoint’s implied unit count and substitute Micron’s 200 GB content, and 2030 demand is several times the Counterpoint total—and still small next to AI-server DRAM unless unit forecasts move from hundreds of thousands toward the multi-million or, on Morgan Stanley’s 2050 scenario, hundreds of millions. None of this is in the LTA or SCA books that cover 2027–2030 server memory. [51] [53]

What the evidence does and does not support

Contractual demand that can be pointed to in filings and calls: Micron’s 26 take-or-pay agreements, $150 billion of minimum-priced remaining obligations, $32 billion of customer financial commitments, and price floors on most of that contracted revenue through 2030; SK hynix’s 14 LTAs covering more than half of 2027–28 capacity, with deposit safeguards and a market-linked price band; Nvidia’s purchase orders for the Rubin ramp and its $279 billion of supplier commitments, heavily memory. Hyperscaler 2026 guides of roughly $720–740 billion are company plans, not customer contracts with the chipmakers, though Alphabet’s $924 billion purchase-commitment ceiling shows how large the contractual maximums have become.

Verbal or forecast demand that should not be treated as firm: SK hynix’s “sold out through 2029” once negotiations are included; the shortage-through-2030 projection; Nvidia’s $1 trillion cumulative visibility and the $2 trillion cloud-industry backlog; Jassy’s “largely sold out through 2027”; every 2027 hyperscaler dollar figure that is not yet a company guide; and both robot memory estimates.

The growth-versus-level point is the one that reconciles Hatzius with the earnings calls. A peak in the GDP impulse in late 2026 or early 2027 can coexist with a higher capex level in 2027, with Microsoft’s $5 billion calendar-2026 trim (a delivery delay, with spend expected to reaccelerate into the first half of 2027), and with memory suppliers that have already sold, under contract, a large minority to a majority of near-term bits. The risk to that setup is not that 2027 capex has been guided down. It is that the contracts cover only part of the book, that price bands still move with the market inside a range, that third-party data-center delays are already slipping spend across the year boundary, and that the productivity payoff outside the builders themselves is not yet visible in the data Hatzius and Slok are using.

---

Recent Findings Supplement (November 2026)

SK Hynix reported record Q3 2026 results on October 28, 2026, with DRAM and HBM volumes described as “effectively sold out” through 2029. This combines secured long-term agreement (LTA) volumes with customer demand under active negotiation, supported by contracts with 14 major customers (including cloud service providers), up from 10 in Q2. These LTAs cover more than 50% of planned 2027–2028 production capacity, typically span up to five years (with some extensions under discussion), include deposits, and feature pricing mechanisms that reflect market conditions within agreed ranges rather than fixed prices.[1][2][3]

• HBM demand is stated as roughly double available supply; a structural supply-demand imbalance is expected to persist through 2030.[1]
• The company plans significantly higher capex in 2027 to expand advanced DRAM/NAND capacity.[1]
• Record quarterly operating profit reached KRW 83.33 trillion (~$62B at then-prevailing rates) on revenue above KRW 100 trillion for the first time, with an ~80% operating margin.[1]

These LTAs provide more durable visibility than prior verbal outlooks, though pricing flexibility within bands means revenue remains sensitive to market conditions. For competitors or entrants, this underscores the challenge of securing multi-year supply commitments from hyperscalers amid tight HBM/DRAM availability.[4]

Micron’s Q4 FY2026 results (reported around November 9, 2026) showed 26 Strategic Customer Agreements (SCAs) covering more than 35% of expected revenue through 2030 (with some extending into 2031). These are take-or-pay contracts with binding volume commitments, backed by ~$32 billion in customer financial commitments (mostly cash deposits) and ~$150 billion in remaining performance obligations (RPO) calculated at minimum pricing. Three-quarters of the SCA revenue has defined pricing frameworks, often with floor/ceiling bands; management expects gross margins at floor prices to exceed prior cycle peaks.[5][6]

• SCAs span data center, consumer, and automotive segments; terms are generally five years (calendar 2026–2030) for most, shorter for automotive.
• This builds on 16 SCAs disclosed in Q3 FY2026 (~$100 billion minimum revenue at the time).[7]
• Physical AI applications (e.g., self-driving cars) are flagged as a substantial demand driver starting around 2030.[8]

The contractual structure (take-or-pay + deposits + pricing bands) differentiates these from purely verbal claims and aims to stabilize earnings in a cyclical industry. New entrants or rivals face high barriers to replicating this locked-in visibility and customer pre-commitments.

Nvidia’s recent Q3 results (period ended ~October 26, 2026) highlighted continued strong AI demand visibility, with data center revenue at $51.2 billion (up 66% YoY) and cloud GPUs described as sold out. Blackwell Ultra is now the leading architecture, with Rubin ramping and expected to contribute meaningfully; hyperscalers and partners (including Google, Microsoft, Oracle) have multi-year infrastructure commitments.[9]

• Management noted accelerating and compounding demand across training and inference, with an expanding ecosystem of model makers, startups, and sovereign customers.
• Production and supply chain planning reflect multi-year visibility, though specific backlog dollar figures were not newly quantified in the latest reports.

Verbal statements on sold-out status and backlog are supported by ongoing hyperscaler capex and partnerships but lack the detailed take-or-pay or deposit disclosures seen from memory suppliers. This reinforces Nvidia’s position while highlighting supply constraints that benefit memory makers with contracted volumes.

Big Tech capex guidance updated in late October 2026 shows sustained high levels with some 2027 increases signaled, though Microsoft’s calendar-2026 figure was trimmed due to timing delays rather than reduced plans. Growth rates in spending may be peaking even as absolute levels remain elevated or rise further.

• Microsoft: Calendar 2026 capex guidance lowered to ~$170 billion (from ~$175 billion previously) due to third-party capacity delivery delays now expected in H1 FY2027 (Jan–Jun 2027); Q1 FY2027 spend was $45.2 billion. Plans call for acceleration in H2 FY2027 driven by CPUs/GPUs, higher memory costs, and shifted deliveries. Earlier reductions partly reflected accounting shifts to operating leases.[10][11]
• Alphabet: 2026 capex guidance held at $195–205 billion (Q3 spend $51.4 billion, contributing to negative FCF). Expects a significant increase in 2027 (analyst estimates often $285 billion+, with some projections exceeding $300 billion).[12][13]
• Amazon: 2026 cash capex guidance held at ~$220 billion (majority for AWS); expects to spend “meaningfully more” in 2027. Strong visibility into 2027–2028 (and some of 2029) demand, with AWS projected to grow >40% in both years; capacity described as largely sold out through 2027 in some commentary.[14][15]
• Meta: 2026 capex guidance $135–145 billion (low end raised); non-cancelable commitments and leases total ~$825 billion in future obligations. Equity raises planned to fund AI buildout; S&P revised outlook to negative citing FCF deficits (forecast ~$22 billion in 2026, peaking higher in 2027).[16][17]

These figures represent capex levels that are high and, in most cases, still rising into 2027; they differ from growth rates, which Goldman Sachs economist Jan Hatzius flagged as peaking in late 2026/early 2027. Hatzius noted AI-related capex (estimated ~$600 billion annualized recently) will continue rising in absolute terms but its marginal contribution to GDP growth will slow as the share of GDP stabilizes, alongside fading wealth effects and other factors.[18][19]

Physical-AI/robotics memory demand remains a longer-term qualitative driver rather than a near-term quantified surge in recent reports. SK Hynix highlighted enterprise/personal AI in the short term and physical AI (autonomous driving, robotics) over the mid-to-long term as sustaining growth. Micron echoed substantial demand from physical AI applications starting around 2030. No new specific market size estimates or revised forecasts for robot/physical-AI memory appeared in the latest earnings or analyses reviewed.[20][8]

Implications for competitors/entrants: Contractual commitments (take-or-pay LTAs/SCAs with deposits and pricing bands) from SK Hynix and Micron provide stronger demand evidence than verbal “sold out” statements and create stickiness that is hard to displace. Big Tech capex remains demand-led and elevated into 2027 despite some timing shifts and growth-rate peaking signals, supporting memory visibility through the decade but pressuring FCF for spenders. Hatzius’s view suggests the investment boom’s incremental economic boost is maturing, which could eventually moderate growth rates even if absolute spending stays high. New players would need differentiated technology, cost advantages, or alternative supply models to compete in this contracted, capacity-constrained environment.

The memory industry is still in a shortage that producers have pre-sold through 2028–29. The first year a glut is even mechanically possible is 2029, and only if several greenfield waves land together while AI bit growth slows. The nearer risk is NAND, not DRAM/HBM.

The binding constraint is wafers, not announcements

HBM does not add bits in proportion to the cleanroom space it consumes. TrendForce’s published path has HBM taking about 18% of DRAM wafer starts at end-2025, about 22% in 2026, and about 30% by late 2027, while delivering only about 8%, 9%, and 13% of DRAM bits. The mechanism is a roughly 3× wafer-area penalty versus DDR of similar density, heading toward 4× as HBM4/HBM4E die sizes grow. Every incremental HBM wafer is taken out of the conventional DRAM pool until new cleanrooms exist. Samsung’s own comment matches that: HBM near 30% of industry DRAM wafer capacity in 2027, up from about 20%, with its own HBM starts rising from roughly 180,000 wafers a month in 2026 to about 250,000 in 2027. [1] [2] [3]

That is why sold-out language and new-fab press releases can both be true. On late-October earnings calls, Samsung said 2027 HBM is fully booked and tightness runs into 2028, with a target of locking 60–70% of capacity under long-term agreements through 2030. SK hynix said LTAs with 14 customers already cover more than half of 2027–28 capacity, customer HBM requests are more than double supply, and combining signed volume with negotiations still in process leaves DRAM and HBM effectively sold out through 2029, with imbalance expected through 2030. Micron has most of 2027 HBM contracted, more than 75% of calendar-2027 output committed, and told investors calendar 2027 and 2028 will be tighter than 2026, with no line of sight to balance. Jensen Huang, in an October 22, 2026 CNBC interview, said he sees no memory glut through 2029. [4] [5] [6] [7] [8]

For anyone trying to underwrite supply, the useful unit is qualified commercial wafers after yield and customer qualification, not groundbreaking dates. Tool lead times have stretched toward six to eight months, so a 2027 cleanroom opening is a 2028 bit story.

Where the new fabs actually land

The Big Three’s greenfield calendar clusters in two waves: a thin 2027 slice, then a much larger 2H28–2030 stack. CSIS’s September 2026 assessment frames the same shape: limited additions in 2026, more meaningful expansion in 2027, and a post-2028 wave large enough to either end the shortage or recreate a glut. [9]

Micron is the most geographically spread. First output is packaging and a mid-2027 wafer start, not a 2026 flood.

• Singapore HBM packaging: initial production early 2027, about $7 billion through the end of the decade. A second Taichung HBM packaging line opened in October 2026.
• Idaho ID1 DRAM wafers around mid-2027; Tongluo (acquired from Powerchip) around the same window, with more meaningful HBM/DRAM shipments into 2028.
• Gap, then 2H28: Idaho ID2, Hiroshima expansion, and a Singapore NAND wafer fab.
• Hiroshima: groundbreaking July 2026, about $9.3 billion, equipment from 2H28, commercial shipments around summer 2028, aimed at HBM and 1γ DRAM.
• Clay, New York: first production around 2030 inside a multi-fab complex planned in phases through the 2040s. CSIS cites an eventual ~200,000 wafers per month for the first Boise fab at full run-rate late in the decade, with a second Boise fab of similar scale planned. [10] [9] [11] [12] [13]

SK hynix has the earliest cleanroom date and the slowest path from cleanroom to full output.

• M15X in Cheongju is the near-term HBM4 lever: operating in 2H26, industry estimates around 40,000 wafers per month initially and about 80,000 in 2027, with management pushing it to maximum by end-2026.
• Yongin Y1 first cleanroom pulled forward to February 2027 from May. Ramp estimates diverge and should not be averaged. Supplier briefings describe six cleanrooms at 60,000 wafers per month every six months, implying a full Y1 near 360,000 by 1H30. CSIS is more conservative: 150,000 wafers per month by end-2028. LS Securities sits in between (about 60,000 in 2027, 180,000 in 2028, 300,000 in 2029).
• Y2: groundbreaking July 2027, first cleanroom June 2029, investment cited at 35.2 trillion won ($26.5 billion). M17 NAND cleanroom targeted for December 2028. Indiana HBM packaging is a 2H29 story, not a 2027 one. [14] [15] [16] [9]

Samsung is converting existing Pyeongtaek lines now and starting the big shell later.

• P5 tool move-in is being discussed for 2Q27 rather than 3Q27, with operations targeted in 2028 and an eventual 100,000–120,000 wafers per month on CSIS’s reading, biased to HBM and 1c DRAM. The shell is being redesigned as a triple-deck fab, which raises ultimate capacity but does not pull first wafers forward.
• First Yongin National Industrial Complex fab is now targeted for 2029 operation, one to two years earlier than the old plan. Six Yongin fabs and two Honam fabs are on the long roadmap; they are not 2027–28 supply. [10] [17] [18]

SEMI’s October 2026 symposium forecast is the clean industry-level check: global DRAM capacity, in 8-inch equivalents, from 2.08 million wafers per month last year to 2.94 million in 2030, up 41%, a 7.1% CAGR, concentrated in ≤15 nm. That is steady compounding, not a 2027 step-function. SEMI also warned that HBM’s extra wafer and back-end intensity can leave both advanced and legacy DRAM short even while headline capacity rises. [19]

What this means for a competitor or a buyer: 2027 is a conversion-and-packaging year. The first year multiple new DRAM shells contribute at once is 2028, and full run-rate is 2029–31. Anyone modeling a 2027 glut is modeling announcements, not wafers.

China adds commodity bits, not a global HBM reset

CXMT is now a real fourth DRAM supplier and still a non-factor in merchant HBM outside China.

Counterpoint put CXMT at about 9.5% of global DRAM revenue in 2Q26 and about 9–10% around 3Q26, fourth behind Samsung, SK hynix, and Micron. Capacity estimates cluster around 320,000 wafers per month now, rising toward roughly 420,000 by 2027; more aggressive commentary talks about 500,000 by 2028. TrendForce’s read is that the extra wafers mostly backfill China’s standard DRAM imports (DDR5, LPDDR5) as the Big Three starve those lines for HBM, and that China’s own AI and cloud build can absorb the increment through 2027. [20] [21] [22]

HBM is where the gap is still a generation. Reuters reported on October 28, 2026 that CXMT has sampled HBM3E to domestic AI chip designers and plans limited production within about two months, at yields of roughly 20–30% versus about 70% at SK hynix. Samsung and SK hynix are already in HBM4 volume, with HBM4E mass production targeted for 1H27 at Samsung. Even a successful CXMT ramp leaves China on HBM3E while the export market is qualifying HBM4. In late October CXMT also claimed engineering-sample validation of a 4F² DRAM platform and said it would sample server DDR5 RDIMMs by year-end; mass-production yield is unproven, and Micron has reportedly declined to commercialize 4F² on yield grounds. [20] [23] [24]

YMTC is the volume threat, and it is in NAND. Counterpoint put YMTC at 16.5% of global NAND shipments in 2Q26, third worldwide; Omdia put its China NAND share at 38.5%, ahead of Samsung. Current output is about 200,000 wafers per month across two Wuhan fabs. Fab 3 starts late 2026 and ramps in 2027. Omdia’s wafer forecast is 2.01 million wafers in 2026 to about 2.50 million in 2027, an increase several times larger than Samsung’s or SK hynix’s planned NAND adds, because the Koreans are spending on DRAM/HBM. Fabs 4 and 5 are each talked about at roughly 100,000 wafers per month when full; Fab 4 construction is expected in 2027, aimed at >300-layer NAND with 60–70% domestic tools. The long-term company target cited in market commentary is about 500,000 wafers per month including a DRAM side-bet. That DRAM side-bet is still a pilot: LPDDR5 at about 5,000 wafers per month in test since late 2025, with trial production discussed as early as end-2026, and a faster 4F²/VCT version not yet in mass production. An IPO sized around $4.9 billion has slipped into 1H27. [21] [25] [26] [27] [28]

Export controls shape this, they do not freeze it. The December 2024 BIS rule (ECCN 3A090.c) controls HBM above a memory-bandwidth-density threshold and reaches foreign-produced HBM via the Foreign Direct Product rule, so Korean and US HBM cannot be sold into Chinese accelerators in the normal channel. CXMT is not on the BIS Entity List; it is on the Defense Department’s 1260H list, which is a procurement and reputational mark, not an Entity List licensing ban. The practical result is a captive Chinese customer that will buy 20–30% yield HBM because the alternative is the gray market. Smuggled HBM3E has been reported around $4,000 per stack, and a co-packaged loophole—packages whose primary function is compute are judged on the processor, not the memory—has been used to move HBM inside server modules. Huawei and other domestic accelerator vendors have raised chip prices on memory cost. The September 2026 US-China summit left advanced-chip and tool controls in place. Controls therefore slow CXMT’s yield curve (no EUV, restricted bonders and metrology) while guaranteeing it a domestic order book. They do not create a global HBM surplus. [29] [30] [31] [32]

For a non-Chinese buyer, CXMT is a 2027–28 pressure valve on commodity DDR/LPDDR inside China, not a qualified second source for HBM4. YMTC is the name that can soften NAND contract prices a year or two before DRAM turns.

CoWoS expands fast and still does not free memory

TSMC’s advanced packaging is a parallel bottleneck, not a substitute for DRAM wafers. Published capacity figures disagree on the exact level and agree on the slope:

• End-2026: TSMC roughly 120,000–140,000 CoWoS wafers per month, with industry totals near 160,000–200,000 once OSATs are included.
• 2027: TSMC targets cited around 180,000–200,000; KGI has raised its end-2027 TSMC forecast to 185,000–190,000 and still sees a 15–20% gap versus revised customer demand. UBS has talked about industry CoWoS near 270,000 by end-2027, with ASE’s increment more aggressive than TSMC’s.
• 2028: street figures around 260,000 for TSMC, Yole nearer 270,000. One supply-chain account allocates about 60% of that to Nvidia. [33] [34] [35] [36] [37]

TSMC has said CoWoS capacity compounds above 80% a year from 2022 to 2027. That still only keeps pace with AI accelerator wafer demand, and package size is eating the gain. CoWoS-L is moving toward roughly 9 to 9.5 reticle equivalents in 2027; only a handful of those interposers fit on a 300 mm wafer, versus a dozen or more of today’s smaller packages. NVIDIA Rubin Ultra-class designs stack on the order of 16 HBM cubes in one socket. Wafer-per-month headlines therefore overstate chip throughput. TrendForce-linked reporting has TSMC and OSATs weighing another 10–20% CoWoS price increase, with equipment lead times out to about eight months. Meaningful non-Taiwan relief—Amkor in Arizona, GlobalFoundries interposers for TSMC—is an early-to-mid-2028 event. TSMC’s panel-level CoPoS is a 2028 start with meaningful volume more like 2029. Intel’s EMIB-T is the credible second source being shopped to hyperscalers, not a 2026 release valve. [38] [39] [40] [41]

Implication: even if HBM bits loosen in 2028, accelerator shipments can stay packaging-constrained. A memory glut and a CoWoS glut are different events. CoWoS expansion supports HBM demand; it does not create spare DDR5.

The 2029 oversupply case, and what it assumes

There is no single published house forecast that treats a 2029 DRAM glut as the base case. The date shows up as the first year the announced greenfield stack can reach balance, and as the year equity markets are already discounting.

The most explicit supply-side dating is an October 2026 Aephel sector note (an AI-authored research desk piece, not a bank): announced fabs add on the order of 1.9 million wafers per month of DRAM capacity across 2026–30, including about 475,000 at CXMT, 360,000 at SK hynix Yongin, and 420,000 across Samsung P4/P5. On those schedules the market reaches balance in 2029 and a modest surplus in 2030. Its demand assumption is DRAM bit demand at a 21% CAGR to 2030, with servers at 43% and HBM bits at 39%. It has demand/supply at 108–112% through 2026–28. A stress case in which China reaches 25% of DRAM by 2030 pulls the surplus forward to 2029 and cuts 2030 incumbent DRAM profit by about 30%; the authors treat 25% as plausible in NAND and not in DRAM without a 4F² or 3D-DRAM breakthrough. [42]

That sits against a tighter near-term tape:

• TrendForce, via an October 2026 Goldman memory expert conference: DRAM stays tight through 2027, the price upcycle extends into 1H28, and meaningful new capacity is hard before 2028 because of cleanroom space and tool lead times. HBM blended ASP was projected up about 122% year over year in 2027, to about $3.7 per gigabit, with HBM industry revenue above $200 billion. Each Korean supplier was seen adding only 80,000–100,000 wafers per month by 2027, much of Samsung’s add going to HBM, so conventional bit growth depends on node migration. [43]
• Citi in early November 2026 still has 2027 undersupplied, but narrowed the DRAM gap to −3.8% from an earlier −8.7%, with supply bit growth of 24% against demand of 28%. A September Citi note had been more aggressive, arguing continual-learning workloads could widen the gap in 2028 and stretch the shortage to 2031, with 2027 DRAM wafer capacity up only about 8% in one Korean press account of that work. The November revision is the more current number; the 2031 call is a demand scenario, not a capacity fact. [44] [45] [46]
• CSIS (September 2026): the capex required to end today’s shortage can recreate the next glut if AI demand slows while Samsung, SK hynix, Micron, and China all ramp together—a replay of past memory “chicken” cycles. If AI demand keeps growing at recent rates, the same wafers are absorbed. A new memory fab still costs roughly $15–20 billion and two to three-plus years. [9] [47]
• Company guidance is the high end of tightness: SK hynix through 2030, Samsung into 2028, Micron explicitly tighter in 2027–28 than in 2026. Deloitte’s July 2026 view was that new supply does not really arrive until 2029 or 2030. Nvidia’s CFO has framed memory constraints as weighing on growth at least through the end of fiscal 2028. [6] [48] [8]

The assumptions that have to hold for a 2029 balance, rather than a 2028 collapse or a 2031 shortage, are therefore:

1. Bit growth decelerates but does not break. Near-term demand in the mid-20s to about 30% (Citi’s 28% for 2027 DRAM) has to slow toward something closer to Aephel’s 21% CAGR by the end of the decade. Node migration no longer delivers “free” bits the way older shrinks did, so wafer adds have to do the work.
2. Capex is large but lagged. Producers are not showing classic discipline—they are accelerating shells—but cleanroom and tool lead times insert 18–30 months between the decision and sellable bits. LTAs covering half to 70% of 2027–28 output also reduce the incentive to flood the spot market. The glut risk is simultaneous 2028–30 ramps, not 2026 capex prints.
3. HBM yield stays a tax on wafers. At ~70% mature yield and a 3–4× area penalty, 30% of wafers still produce only low-teens bit share. A faster yield climb on 12-high and 16-high HBM4 would release conventional bits earlier than the 2029 models assume. CXMT at 20–30% does the opposite for global supply: it consumes Chinese wafers without exporting competitive HBM.
4. China stays mostly domestic through 2028. TrendForce’s absorption thesis has to fail—CXMT DDR5 or 4F² qualifying at non-Chinese hyperscalers in volume—before China alone pulls a DRAM surplus forward.

A Korean press column and some market commentary have flagged 2029 as the year new supply turns competitive, and one secondary account has DRAM prices down about 23% in 2028 versus 2027 while still far above 2025. That is a peak-and-fade, not a glut. The 2029 date in serious work is a balance year under full execution of announced fabs, with a modest surplus more naturally in 2030.

Supply-demand timeline

Year	Supply that actually shows up	Demand / pricing signal	Balance
2026	HBM ~22% of DRAM wafers, ~9% of bits. M15X starting. CXMT HBM3E sampling at poor yield. TSMC CoWoS ~120–140k wpm.	Sold out on HBM; Korean inventories reportedly under 10 days. LTAs being written through 2028–29.	Deep shortage. Commodity DRAM starved by HBM mix.
2027	Partial only: Micron ID1 and Tongluo mid-year, Singapore HBM packaging early, SK Y1 cleanroom from February with limited wafers, Samsung HBM starts toward ~250k/month, CXMT toward ~420k mostly for China, YMTC Fab 3 ramp. HBM ~30% of wafers / ~13% of bits. CoWoS toward ~180–200k at TSMC, still short on large packages.	Citi: DRAM demand +28%, supply +24%, gap about −4%. Samsung and Micron: tighter than 2026. TrendForce: HBM ASP roughly +120% YoY.	Still short. First year of greenfield wafers, not a relief year.
2028	First broad wave: Samsung P5, Micron ID2 / Hiroshima / Singapore NAND, SK Y1 in real ramp, Micron Taiwan EUV fab. Tooling conditions improve from 2H27, so this is the first year new shells matter. Some CoWoS relief outside Taiwan.	TrendForce: price upcycle into 1H28, then the open question. Micron: still tighter than 2026. Aephel: DRAM peaks around 2Q28; NAND contract down 15–25% as the more likely soft spot.	Shortage easing at the margin in 2H, not a glut, unless AI capex stalls. NAND can loosen before DRAM.
2029	Overlap year: SK Y2 cleanroom (June), Samsung Yongin first fab targeted, Y1 further up the ramp curve, CXMT’s multi-year add largely in place. Micron New York still ~2030. CoPoS starts to matter.	Base case in the explicit wafer-add model: balance. Company guidance (SK hynix) still says shortage if AI demand holds. LTAs blunt a price collapse even if bits loosen.	Candidate balance year. Modest surplus is a 2030 story unless China or demand breaks.
2030–31	SEMI: DRAM capacity +41% vs 2025. Yongin and Pyeongtaek clusters plus Micron New York in ramp. Full run-rate of shells announced in 2025–26.	SK hynix and some demand scenarios (continual learning, physical AI) still short. CSIS: this is when a classic overbuild can show up if 2026–28 capex all lands and AI growth slows.	Decision year for the next downcycle. Capex can still be slipped; the shells will already exist.

Confidence is high on 2026–27 tightness, because it is in company disclosures and contract coverage. Confidence is medium on 2028 as an inflection rather than a collapse. Confidence is low on the exact quarter of balance in 2029–30, because it depends on HBM yield, China qualification, and hyperscaler capex three years out.

What would pull a glut forward of 2029

An earlier glut means surplus bits in 2027 or 2028, not a softer print in 2030. Six triggers would do it, in rough order of how fast they could hit.

1. An AI capex air pocket, not a slowdown in the narrative. Aephel’s sensitivity is the sharpest published one: DRAM stays short in 2027 only if top hyperscaler capex grows more than about 25%; flat capex implies roughly a 10% surplus that year. JPMorgan has flagged about 37 GW of US data-center capacity scheduled for 2028 as exposed to permitting risk. That matters only if it cancels memory orders already in LTAs, rather than sliding builds to other regions. Watch take-or-pay renegotiations, not headline capex guidance. [42] [49]
2. HBM yield jumping faster than the 3–4× wafer tax. If 12-high and 16-high HBM4 yields move well through the current ~70% mature level, the same wafer start produces more sellable stacks and frees DDR5. The opposite—yields stuck, customers already de-specing Rubin Ultra from 12-high toward 8-high—keeps the shortage. De-spec is rationing, not demand destruction, unless content per GPU falls faster than unit growth.
3. CXMT qualifying outside China. A 4F² or mainstream DDR5 RDIMM that hyperscalers will buy in volume, or HBM yields moving from the 20–30% range toward 50% with a product that can leave China, is the supply shock that pulls a commodity-DRAM surplus into 2028–29. Domestic absorption through 2027 is the reason this has not happened yet. Export controls make the HBM version of this shock slower, not impossible.
4. Tool lead times collapsing. If the six-to-eight-month equipment queue shortens and every 2028 shell (P5, ID2, Hiroshima, Y1 phases) ramps in the same two quarters without utilization cuts, the historical memory pattern returns: supply decisions made at the peak arrive together. Producers say they will pace this. They have said that before.
5. LTA breakage. Contracts covering 50–70% of 2027–28 output are the earnings floor. If cloud buyers force price resets or volume cuts when spot softens, the floor becomes the transmission mechanism for a glut rather than a cushion. Deposits and market-linked pricing at SK hynix are designed to prevent exactly that.
6. NAND first, then DRAM by contagion. YMTC’s 2027 wafer add, against Korean NAND capex deferred toward HBM, can put NAND into surplus in 2027–28 even while DRAM is tight. If enterprise SSD pricing breaks and hyperscalers reallocate budgets, DRAM contract negotiations in late 2027 feel it before any new DRAM fab is full. That is the earliest plausible “glut,” and it is a NAND glut.

The competitive implication is asymmetry. Incumbents have sold the shortage forward and are building the capacity that could end it in 2029–30. A new entrant or a buyer betting on cheap DRAM in 2027 is fighting cleanroom physics and a 3× HBM wafer penalty. A buyer who needs NAND, or a Chinese commodity DRAM alternative for non-HBM uses, has a real supply response arriving sooner. The trigger list above is the monitoring set: hyperscaler capex revisions, HBM4 yield commentary, CXMT qualification wins outside China, and whether 2028 tool move-ins are pulled in or quietly slipped.

---

Recent Findings Supplement (November 2026)

Micron has paused its key Taiwan HBM/DRAM expansion while accelerating Singapore and US sites, creating near-term supply uncertainty amid strong contracted demand.[1][2]

• In early November 2026 reports, Micron effectively halted construction on the $4.5 billion Fab 11E in Taoyuan (intended for EUV-based 1-gamma DRAM and HBM, with 2028 startup) due to a labor dispute with the Taoyuan union over profit-sharing bonuses and pay terms; minor safety work continues but major activity has stopped, with no resumption timeline disclosed. This follows an October agreement with the larger Taichung union.[3][4]
• Offsetting this, Micron confirmed progress on a $7 billion Singapore HBM advanced packaging facility (first outside Taiwan), with initial output targeted for early 2027, plus a $24 billion double-story wafer fab slated for 2H 2028 operations.[5][6]
• The company also announced a January 16, 2027, groundbreaking for its $100 billion New York megafab campus (up to four fabs, advanced memory for AI).[7]
• In Taiwan, the second advanced assembly/test facility (AATT2) in Taichung opened in mid-October 2026; HBM output through 2027 is fully contracted. Tongluo (acquired from PSMC, closed March 2026) targets mid-2027 wafer shipments.[8][9]
• A $9.3 billion Hiroshima expansion (announced earlier but with ongoing relevance) adds cleanroom space for 1-gamma/1-delta DRAM and HBM4E, with 2028 production.[10]

This mix of delays in Taiwan and acceleration elsewhere signals Micron prioritizing diversified, geopolitically resilient capacity while facing execution risks from labor issues; competitors may gain share if the pause extends.[11]

SK hynix and Samsung report near-total sellouts of DRAM/HBM capacity through 2028–2029 via long-term agreements (LTAs), with explicit expectations of shortages persisting into 2030.[12][13]

• On its October 28, 2026, Q3 earnings call, SK hynix stated that DRAM and HBM volumes are “effectively sold out” through 2029 when combining secured LTAs (with 14 customers, covering >50% of 2027–2028 capacity) and ongoing negotiations; customer HBM requests exceed twice current supply. It expects shortages to continue through 2030 due to fab lead times and process complexity. HBM4 is projected to surpass HBM3E in Q4 2026 shipments/revenue; M15X fab ramps to full capacity by year-end.[14][15][16]
• Samsung’s October 29, 2026, Q3 call confirmed 2027 HBM supply fully sold out (with ongoing additional requests); HBM4 revenue more than tripled QoQ. It signed LTAs with 10 major customers, is negotiating with 10 more, and targets 60–70% capacity coverage by 2030. HBM4E mass production begins H1 2027. Overall memory supply expected tight through 2028; record 2026 capex of 84.4 trillion won (mostly semiconductors).[17][18][19]
• Both companies note stable or improving HBM4 yields and pricing power from the tight market.

These sellouts and multi-year visibility via LTAs (with deposits and flexible pricing) lock in revenue and reduce cyclical risk, but they also highlight how capacity additions are lagging AI-driven bit demand; new entrants or Chinese players face high barriers to displacing contracted supply.[20]

Chinese players CXMT and YMTC are advancing despite technology gaps, with CXMT targeting limited HBM3E output by end-2026 at low yields and YMTC scaling NAND while eyeing DRAM.[21][22]

• Reuters reported October 28, 2026, that CXMT plans limited HBM3E production within two months (by ~end-2026), having shipped samples to domestic AI chip designers; yields are estimated at 20–30% (vs. SK hynix ~70%), with performance closer to prior-generation HBM3. The firm claims progress on high-aspect-ratio TSV processes. Q3 DRAM market share reached ~9–10%. It validated a 4F² DRAM platform and plans DDR5 RDIMM samples by end-2026.[23][24][25]
• YMTC (No. 3 globally in NAND by volume in recent data) filed for a Shanghai STAR IPO and is planning a fourth Wuhan fab (>300-layer NAND, AI/data-center focus) with construction targeted for 2027 and 60–70% domestic equipment (up from ~50%+ at Fab 3). Potential per-fab capacity ~100k wafers/month. It is also conducting R&D on LPDDR5/DRAM using Xtacking/VCT 4F² architectures.[26][27]

Low yields and process immaturity limit near-term impact on global supply, but domestic Chinese demand (and potential indirect channels) could absorb output; export controls add friction but have not fully halted technology leakage.[28]

TSMC is aggressively expanding CoWoS advanced packaging capacity (critical for HBM stacking with logic dies), targeting a doubling by 2028 amid price pressure and allocation battles.[29]

• Analysts (Economic Daily News, Goldman Sachs, etc.) project TSMC CoWoS at ~130k wafers/month by end-2026, rising to ~200k in 2027 and 260k by end-2028 (some estimates higher at ~270–290k). Expansions center on AP7 (Chiayi) and AP8 (Tainan), with US Arizona packaging also planned.[30][31]
• Reports from late October 2026 indicate potential 10–20% CoWoS price hikes due to equipment lead times stretching to 6–8 months and strong AI/ASIC demand. Allocation estimates include ~60% to Nvidia.[32][33]

Packaging remains a key bottleneck alongside HBM; faster CoWoS ramps or rival technologies (e.g., Intel EMIB) could ease constraints, but current trajectories support continued tightness.[34]

Published outlooks emphasize prolonged memory shortages rather than oversupply around 2029, driven by AI bit growth outpacing capacity additions.[35]

• SK hynix (Oct 2026) projects demand exceeding supply through 2030, citing long fab lead times and leading-edge process complexity; LTAs provide visibility but leave limited spot capacity. Micron and Samsung align on tightness into 2028.[36][37]
• Assumptions behind tightness (vs. potential glut): Strong AI-driven bit growth (HBM requiring 3–4x wafer capacity per bit vs. conventional DRAM); disciplined but lagged capex; improving but still-constrained HBM yields; and demand from hyperscalers/cloud/ASICs absorbing supply. Earlier Goldman notes projected undersupply persisting into 2028. No major 2029 glut forecasts appear in recent coverage; instead, extensions of the supercycle are highlighted.[38]

An earlier glut could be triggered by sharper-than-expected slowdown in AI capex, breakthrough yields or faster ramps from Chinese suppliers (CXMT/YMTC), or accelerated global capex leading to oversupply post-2028 once new fabs (e.g., Micron NY/Singapore, Samsung Texas, SK expansions) come online in volume.[39]

Export controls continue to constrain advanced HBM flows to China while prompting workarounds and license requests for maintenance at legacy fabs.[40]

• US rules (including December 2024 HBM controls) remain in effect; SK hynix sought Commerce licenses in October 2026 for Wuxi DRAM equipment maintenance (site produces ~48% of its DRAM). Indirect channels reportedly allow some Korean HBM to reach Huawei-linked supply chains despite bans.[41][42]

Controls raise costs and uncertainty for Korean/Japanese/Taiwanese supply chains but reinforce the premium on non-Chinese advanced capacity for Western customers.[28]

Overall supply-demand timeline (synthesized from recent company guidance and analyst estimates): Tight/shortage conditions through 2026–2028 (with HBM particularly constrained); gradual capacity additions ramping in 2027–2028 but unlikely to close the gap fully until after 2029–2030 absent major demand shifts. Chinese domestic supply grows but at lower yields/efficiency. No credible recent sources forecast a 2029 glut; the dominant view is extended tightness.

The sell-side “peak” is mostly a growth-rate peak, not a demand peak. Absolute AI capex in the latest bank models is still rising into 2028. What has peaked, on their own numbers, is the year-on-year rate of increase. The longer views from SemiAnalysis, Anthropic, Jensen Huang and Elon Musk do not put a ceiling on compute demand in this decade at all. They put the ceiling on power, debt and construction.

What the five banks actually dated

Wells Fargo’s strategist is the earliest level-risk call. On 7 October 2026, chief equity strategist Ohsung Kwon said consensus hyperscaler capex growth is already at its year-on-year peak, and that the risk not priced into the buildout trade is that spending flatlines after 2027 rather than rolls over. His arithmetic: AI demand has to reach about $1.4 trillion by 2028 to justify capex through 2027 (roughly $9,000 per worker per year). Another large capex year in 2028 pushes the required demand to $2.2 trillion. That is why he expects 2027 to be a year in which applications, software and the hyperscalers themselves outperform semis. This is a rate-of-change and free-cash-flow argument, not a claim that tokens stop growing. The semiconductor team under Aaron Rakers has been the opposite voice inside the same bank: in January 2026 he forecast industry sales above $1 trillion in 2026 (+29%) and another double-digit year in 2027, and in May 2026 Wells raised its Nvidia target to $315 from $265 on the view that 2027 earnings are sustainable, not peak-cycle. A February 2026 Wells note had top-five capex at only $716 billion in 2026 and $862 billion in 2027. Later street numbers are far above that. The internal split is the tell: coverage analysts keep lifting the level; the strategist is marking the second derivative. [1] [2] [3] [4]

Goldman is two desks, two peaks. Research (Ryan Hammond and team, 25 September 2026) has the five large U.S. hyperscalers at about $1.2 trillion of AI infrastructure spend in 2027, up roughly 54% from an ~$800 billion 2026 base, then only +12% to about $1.4 trillion in 2028. Cumulative AI infrastructure from 2026 through 2031 is modeled at about $7.6 trillion ($5.1 trillion compute, $2.1 trillion data centers, $358 billion power). Growth has already halved; the level has not peaked. The Delta One trading desk, in a note cited 3 November 2026, is the actual “2027 peak” call: SOX at about 12.2x forward earnings and Nvidia at about 10x (its lowest multiple since 2013) may be signaling “significant risk to estimates,” and 2027 could be an earnings peak because the cost of capital and debt-market capacity, not chip demand, bind first. The same desk still says use cases (robotics, enterprise workflow) are barely started. Goldman has also flagged hyperscalers possibly issuing a record ~$420 billion of bonds in 2027. Research is a deceleration call. The desk is a financing-constraint call. Neither is a published single-stock price target. [5] [6] [7] [8]

Barclays is the cleanest dated peak, and it has already slipped once. Tom O’Malley’s 11 March 2026 framework, built from OpenAI and Anthropic financials rather than GDP analogies, says hyperscale capex peaks in 2028 because that is when labs are expected to hit recursive self-improvement, which raises spend efficiency and cuts training cost. Training operating expense is assumed to peak in 2029, so capex must peak a year earlier (infrastructure has to be in place before the training ramp). The same note said Street capex was at least $225 billion too low across 2027 and 2028 combined, and that Nvidia was priced as if the peak were 2027 — about 17.5x calendar-2027 earnings and 14.5x 2028 even on conservative growth, “simply too great” a penalty. Head of U.S. equity strategy Venu Krishna repeated “peak in 2028” on CNBC on 13 May 2026. By late October, tactical strategist Alex Altmann was on the other side of the same numbers: capex of $1.2 trillion next year and $1.4 trillion the year after, against an AI revenue run-rate Barclays puts near $400 billion by end-2028, versus ~$1 trillion of revenue needed to pay for the spend. No fresh single-name target was attached to the March or May peak notes. Barclays’ own earlier macro call — AI’s GDP contribution peaking in 2025 and fading — has already been walked forward. That is the track record that matters. [9] [9] [10] [11] [12] [13]

UBS has not dated a level peak. On 2 November 2026 it raised global AI capex to $1.0 trillion in 2026, $1.4 trillion in 2027 and a new $1.5 trillion for 2028, and said investors should not abandon the theme on short-term volatility. A turning point is conditional, not calendared: lab revenue growth slows meaningfully, cloud utilization weakens, falling token prices fail to create enough new usage, algorithmic efficiency sharply cuts compute needs, or financial conditions stop the hyperscalers from funding the next round. Neoclouds are only ~8% of its capex estimate, so their credit stress is not, in UBS’s view, a system peak. A separate UBS monetization note says the six largest hyperscalers need roughly $1 trillion of incremental operating profit by 2031 (about $3.4 trillion of extra revenue at current cloud margins) to earn a 15% return — a 2031 payback problem, not a 2027 spend peak. Public summaries of these notes do not carry a single-stock price target. One caution: a Reuters item dated 10 November 2026 cites a narrower UBS hyperscaler series (growth slowing to 25% in 2027 and 6% in 2028 off a $673 billion 2026 base). That series does not match the 2 November global AI totals. Treat them as different scopes, not one number. [14] [15] [16] [17] [18]

Morgan Stanley’s level peak is after 2028; its growth peak is already behind it. An October 2026 forecast had AI infrastructure spending at about $980 billion in 2026, $1.58 trillion in 2027 (~+60%) and $1.81 trillion in 2028 (~+15%), with the cash-flow gap tightest in 2027 and a U.S. data-center power shortfall of 32 gigawatts by 2028 against 97 gigawatts of demand. The 8 November note, after the midterms, kept forecasts at about $1.6 trillion / $1.8 trillion for 2027 / 2028 (an eight-company exhibit printed $1.64 trillion and $1.76 trillion) and moved about 5 gigawatts of U.S. projects from 2028 into 2029 — an “air pocket” risk for the supply chain in 2028, not a cancellation. CEO Ted Pick, 3 November 2026, put hyperscalers at $800 billion to $1 trillion a year over a 7- to 12-year horizon and cumulative spend at $10–15 trillion, “second inning of a nine-inning game.” Preferred names in the November defense of the trade include Amazon, Meta, Alphabet, Microsoft, Nvidia, Broadcom, Equinix and several power names. No single price target is attached to the peak arithmetic in the public write-ups. Stephen Byrd’s data-center ROIC estimate of about 30% so far is the mechanism MS uses to argue the cycle elongates rather than breaks. [19] [20] [21] [22] [23] [24]

Table, sorted by implied peak year

Implied peak	Source and date	What actually peaks	Rationale	Price target attached	Incentive	Track record to weigh
Growth already; levels maybe after 2027	Wells Fargo, Ohsung Kwon, 7 Oct 2026	YoY capex growth now; risk that levels flatline after 2027	$1.4T AI demand by 2028 needed to justify spend through 2027; $2.2T if 2028 is another big year	None on this note. Semi team: NVDA $315 (May 2026, from $265)	Strategist is paid to rotate, not to cheer semis. Coverage team is paid to stay long the cycle	Feb 2026 top-five capex ($862B in 2027) was already low versus later street. Classic under-forecast, then a strategist fade
Possible 2027 earnings peak	Goldman Delta One, cited 3 Nov 2026	Semiconductor earnings, via financing, not demand	SOX ~12.2x, NVDA ~10x may be pricing “significant risk to estimates”; cost of capital binds; late-1990s “capacity before use” analogy. Desk still bullish on untapped use cases	None. Research team is not this call	Trading desk marks risk and client positioning; less tied to ECM/DCM than research	Separate from GS research, which still has spend rising in 2028. Desk is describing what the tape is implying, not a bottom-up model
Growth peak passed; level still up in 2028	Goldman research, 25 Sep 2026	Growth rate (54% in 2027, 12% in 2028)	Five hyperscalers ~$1.2T in 2027, ~$1.4T in 2028; ~$7.6T cumulative 2026–2031. Capex already above operating cash flow	None in the public summary	Research supports a financing super-cycle GS is underwriting (record IB fees; Solomon: robust demand for AI deals)	Street, including GS, has revised AI capex up repeatedly since 2023. Direction of error has been too low, too early
2028 level peak	Barclays, Tom O’Malley, 11 Mar 2026; Venu Krishna, 13 May 2026	Hyperscale capex level	RSI around 2028 raises efficiency and cuts training cost. Training opex peaks 2029, so capex peaks a year earlier. Street ≥$225B light in 2027–28 combined	No new PT. Valuation: NVDA ~17.5x CY27 / ~14.5x CY28 is too cheap if the peak is 2028	Semi analyst’s universe wins if the cycle is longer. Krishna is a strategist, less single-stock conflict	Barclays itself says the Street has underestimated capex since late 2022. Barclays’ own 2025 “GDP boost peaks this year” call has already been pushed out
No level peak through 2028; conditional later	UBS, 2 Nov 2026	None dated. Levels: $1.0T / $1.4T / $1.5T for 2026 / 2027 / 2028	Q3 cloud growth (Alphabet, Microsoft, Oracle) still confirms demand. Turning point only if lab revenue, utilization, or efficiency breaks	None. Message is selectivity, not a target	Wealth and IB franchise benefits from a continuing theme, with a volatility caveat that protects the franchise if it cracks	Same systematic underestimation. Sept 2026 memory revision (memory from 14% of AI capex in 2025 to 64% in 2027) shows the model is being rebuilt around price, not just volume
Growth peak passed; level peak after 2028, with a 2028 air pocket	Morgan Stanley research, Oct 2026 and 8 Nov 2026; Ted Pick, 3 Nov 2026	YoY growth (about +60% in 2027, then mid-single to mid-teens). Not the level	$1.58–1.64T in 2027, $1.76–1.81T in 2028. 5 GW of U.S. builds slip 2028→2029. CEO: $0.8–1T a year for 7–12 years, $10–15T cumulative, “second inning”	No single PT. Buys: AMZN, META, GOOGL, MSFT, NVDA, AVGO, EQIX, power names	MS is structuring the debt and equity (Meta, Anthropic and others). A long cycle is the advisory franchise	Forecasts have been revised up (June five-company 2027 figure was well below the November exhibit). Power-shortfall call (32 GW U.S. by 2028) has been more stable than the dollar capex call

The longer views do not have a peak year inside this window

SemiAnalysis (Dylan Patel, Runtime 2026, published early November) is a supply-rationing model, not a cycle model. Global compute went from about 60 GW at end-2024 to over 100 GW by end-2026. OpenAI and Anthropic together go from about 4 GW at end-2025 to about 10 GW this year and over 20 GW next year, taking roughly 12 GW of an estimated ~40 GW of 2027 additions (25–40% of new compute). He then has AI labs adding on the order of 30 GW in 2028 alone. Cumulative global AI capex by end-2029 is put at about $11 trillion, of which about $7 trillion is debt. H100 one-year rents are cited up from a trough near $1.70 per GPU-hour to a midpoint near $2.80. The mechanism is concentration plus training, not chatbot inference: he has said all of Claude Code is served by on the order of a gigawatt while Anthropic holds several. A peak in 2027 or 2028 is not in this framework. The risk he names is who finances $11 trillion, not whether the tokens are wanted. Incentive: SemiAnalysis also matches compute, so tightness is partly its product. Offset: its datacenter gigawatt path has been more stable across revisions than sell-side dollar capex, and it has been early on HBM and cluster bottlenecks. [25] [26] [27] [28]

Anthropic’s “view” is a contract, not a forecast. Amended S-1 figures put non-cancelable purchase and lease commitments at about $895 billion, mostly compute, with earlier reporting around $518 billion over the next decade and roughly 80% owed whether or not the capacity is used. Reported plans: more than 3.5 GW by end-2026, aiming to roughly double again in 2027 (press accounts differ on whether the 2027 target is ~7 GW or ~9.5–10 GW). Management’s stated constraint is compute, power and labor, not customers. Dario Amodei’s prospectus language is that AI abundance transforms every sector more than industrialization — a demand view with no peak inside the commitment window, which runs into the mid-2030s on some leases. Incentive is maximal: lock scarce supply while revenue is still compute-constrained (compute and infrastructure expense was 58% of revenue in the July–September 2026 quarter in the S-1/A). The risk to a 2027–28 peak thesis is that these obligations do not expire when a bank model says RSI arrives. [29] [30] [31] [32]

Jensen Huang’s number is an annual run-rate in 2030, not a peak. He has repeated, including at the Goldman Sachs Communacopia conference in September 2026 and on earnings calls since at least August 2025, that AI infrastructure spending reaches $3 trillion to $4 trillion a year by the end of the decade. He has also said reasoning and agentic systems may need on the order of 100x, then 1 million times, and in an October 2026 interview potentially up to a billion times more compute than earlier models, with land, power and shells constraining 2027 specifically. Needham’s Laura Martin has noted that hyperscaler consensus near $1 trillion by 2028 is roughly a quarter of Huang’s 2030 figure. Incentive is total: Nvidia captures a large share of each data-center dollar (he has said ~$35 billion of a $60 billion data center). Track record: the $3–4 trillion line has not been walked back as capex prints have risen toward it; the near-term constraint language (power, land, 2027) has been more accurate than the unconstrained demand language. [33] [34] [35] [36] [37]

Elon Musk’s peak is a power peak, and it arrives before demand does. At Davos in January 2026 he called electrical power the limiting factor. By late August 2026 he was citing analyst consensus of at least a 15 GW shortfall in 2027 for AI chips, with chip output up 40–50% a year and power outside China up 10–20%. SpaceX’s own plan, on its August 2026 call, was above 2 GW by end-2026 and closer to 10–15 GW than 20 GW by end-2027, using self-built generation so capacity does not wait on the interconnection queue. Morgan Stanley’s 32 GW U.S. gap by 2028 is the same mechanism with a larger number. Incentive: SpaceX and xAI sell powered compute into that gap. The directional call — chips will outrun the grid — is the one bank view that has not been revised away. [38] [39] [40] [41] [42]

Mapped onto 2027, 2028, 2029 and 2031

2027 — a financing and rate-of-change year, not a demand year. This is the Wells strategist scenario and the Goldman desk scenario. Cash-flow gaps are widest (MS says the mismatch peaks in 2027 even as dollars rise). Bond issuance is modeled at a record. Nvidia at ~10x forward, in the desk’s telling, is already discounting an earnings peak. What falsifies it: hyperscaler 2027 guides that are still “significantly” or “meaningfully” higher, which several have already signaled, and SemiAnalysis/Anthropic additions of ~12 GW for the two labs alone. A 2027 peak is a capital-markets event. It is not what the order book says.

2028 — the only year a major bank has explicitly labeled a capex peak. Barclays (O’Malley/Krishna) is the base case for a true level peak, via recursive self-improvement. UBS and Goldman research still have dollars up in 2028, but growth in the mid-single to low-teens, which trades like a peak even if the accountant does not. Morgan Stanley’s air pocket — 5 GW slipped from 2028 to 2029, U.S. additions down about 4% that year — can look like a peak in supplier revenue without any demand peak. Musk and MS power math say a chunk of 2028 chips may be built and dark. Huang’s constraint language also singles out 2027–28 physical limits. This is the year the four mechanisms (RSI efficiency, FCF math, grid queues, and election-driven permitting) can coincide even if tokens do not roll over.

2029 — where the slipped peak and the lab ramp meet. Barclays has training opex peaking in 2029, which is a spend-efficiency peak, not a smaller industry. Morgan Stanley has U.S. additions reaccelerating in 2029 after the slip. SemiAnalysis has labs still adding on the order of 30 GW in 2028 and does not show a rollover into 2029; the $11 trillion cumulative figure is an end-2029 stock of capital, which only works if 2029 is still a huge flow year. Anthropic’s doubling path, if it holds, is not finished in 2028. A 2029 “peak” in sell-side models is mostly the capex that could not be energized in 2028 showing up late. That is elongation, which is MS’s explicit investment conclusion: scarcity in 2027 raises ROIC on capacity that is already on, and delays relieve equipment and credit bottlenecks rather than kill the cycle.

2031 — monetization math, not a cycle top. UBS’s 15% ROIC hurdle ($1 trillion of incremental EBIT, ~$3.4 trillion of revenue) lands here. Goldman’s $7.6 trillion cumulative infrastructure total runs through 2031, with a “harvest” phase after 2027 in some write-ups but absolute dollars still enormous. Huang’s $3–4 trillion annual infrastructure number is a 2030 run-rate, which, if even half right, means 2031 is not a peak. Pick’s 7- to 12-year, $10–15 trillion frame puts the midpoint of the build past 2031. Anthropic’s non-cancelable leases run through the early-to-mid 2030s. The 2031 scenario is the one in which 2027–28 “peaks” were growth-rate peaks, the revenue gap Altmann flagged (~$400 billion run-rate versus $1 trillion needed by end-2028) either closes through inference scale or forces a capex cut that none of the lab contracts currently allow.

How to use this if you are positioned against the cycle

The bank disagreement is narrower than the headlines. All five have year-on-year growth collapsing after 2026–27. Only Barclays has a clear level peak (2028), and only Goldman’s trading desk and Wells’ strategist treat 2027 as a possible earnings or flatline year. UBS and Morgan Stanley still have higher dollars in 2028 than in 2027. The outside views (Patel, Amodei, Huang, Musk) agree with each other on one point the banks keep splitting: demand is not the binding constraint before 2030. Power, transformers, debt capacity and non-cancelable leases are.

Three practical distinctions follow.

First, a trade that is short “the peak” has to specify which peak. Short semis into a 2027 financing scare (Wells strategist, Goldman desk, MS cash-flow gap) can be right even if 2028–29 gigawatts are larger. Short the entire complex into 2028 on Barclays’ RSI date is a bet that efficiency gains dominate Anthropic’s contracted doubling and Patel’s 30 GW lab add. Those are different trades. The first has happened before in this cycle (multiple 20–40% semi drawdowns while capex guides rose). The second has not: every dated sell-side peak since 2024 has been pushed out, and Barclays has documented the Street’s systematic low bias.

Second, incentives are not subtle, and they cut both ways. Goldman and Morgan Stanley are earning record or near-record investment-banking fees underwriting this build (Solomon: an AI capex super-cycle across every instrument; Pick: 10–15% of the way through, later “second inning”). Their research having no level peak through 2028 is consistent with the franchise. It is also consistent with the order book. SemiAnalysis sells scarcity. Nvidia sells the pick-and-shovel share of Huang’s $3–4 trillion. Anthropic has already signed away the right to be small. The cleanest less-conflicted reads are the conditional ones: UBS’s turning-point checklist (lab revenue, utilization, algorithmic efficiency) and Musk/MS on megawatts that cannot be turned on. Those do not require anyone to be early on a demand peak.

Third, the 2028 air pocket is the non-obvious competitive opening. If MS is right that ~5 GW and about 20% of some U.S. project tranches slip from 2028 into 2029, equipment, memory and neocloud equity can have a down year in deliveries while the cumulative capex forecast is unchanged. That is a timing short, not a thesis that AI demand broke. Suppliers whose capacity can be repurposed (general-purpose components, power equipment with non-AI offtake) absorb a one-year slip. Suppliers whose lines were sized for a 2028 peak, and labs whose non-cancelable leases assume that power arrives, do not. The banks that look “cautious” for 2027–28 and the outsiders who look “uncapped” through 2031 can both be right if the thing that peaks is the growth rate and the energized-megawatt schedule, not the desire for compute.

---

Recent Findings Supplement (November 2026)

Recent (post-May 10, 2026) forecasts from major banks show a range of views on AI/hyperscaler capex trajectories, with several highlighting risks of growth deceleration or a shift away from pure infrastructure buildout around 2027–2028, while others maintain elevated spending projections into 2028.[1][2][3][4][5]

In contrast, longer-horizon perspectives from SemiAnalysis, Anthropic’s disclosures, Jensen Huang/NVIDIA, and Elon Musk/SpaceX emphasize sustained or accelerating demand and capacity additions well beyond 2028, driven by tight compute markets, rising per-GW monetization, and massive committed spend.[6][7][8][9]

These differences reflect varying incentives (sell-side banks balancing client relations and deal flow versus vendors/labs with direct stakes in expansion) and analytical lenses (near-term financial returns vs. structural demand). Track records are mixed: banks have revised hyperscaler capex estimates upward repeatedly amid strong demand signals, while specialized researchers and operators have consistently underestimated buildout scale in prior cycles.

Bank Peak/Growth Calls (Recent Updates)

Wells Fargo analysts, including equity strategist Ohsung Kwon (October 2026), flag 2027 as the year when AI applications could overtake pure buildout, with capex growth potentially flatlining afterward. They note that justifying continued heavy spending through 2027 requires AI-related demand reaching roughly $1.4 trillion by 2028 (or $2.2 trillion with another major 2028 round), and consensus already prices peak year-over-year growth.[1][10] Earlier 2026 notes from the firm highlighted rising per-GW costs but expected pass-through to customers, with hyperscaler AI capex at $1.1 trillion in 2027 (above consensus).[11]

Barclays (October 2026 commentary from Alex Altmann) stresses a revenue gap: AI needs ~$1 trillion in annual revenue by end-2028 to support capex levels of $1.2 trillion (2027) and $1.4 trillion (2028), but expects only ~$400 billion run-rate by then—a $600 billion shortfall that pressures free cash flow.[3][12]

UBS has repeatedly raised estimates (latest November 2026 notes), projecting $1 trillion AI capex in 2026, $1.4 trillion in 2027, and $1.5 trillion in 2028, with memory costs driving much of the increase. It sees a potential returns inflection in 2027 as capacity comes online and views the cycle as longer than markets price, though it flags volatility and 2027 debt-ceiling risks.[13][4]

Goldman Sachs (September 2026) projects hyperscaler capex at ~$800 billion in 2026, $1.2 trillion in 2027, and $1.4 trillion in 2028, with growth slowing from ~100% to 54% to 12%. It notes ~$300 billion in annual AI revenue needed to break even on average capex and highlights $3.2 trillion in off-balance-sheet commitments.[5][14][15]

Morgan Stanley (November 2026 note) keeps hyperscaler forecasts unchanged at $1.6 trillion for 2027 and $1.8 trillion for 2028 (with Google alone at $375 billion in 2027 for the tracked group), expecting firms to “spend and build as much as they can” despite potential 2028 downward bias from delays; it shifts some projects to 2029.[2]

Longer Views on Demand and Infrastructure

SemiAnalysis (Dylan Patel, October–November 2026 commentary) projects ~$11 trillion in global AI capex by end-2029, with AI labs (led by OpenAI and Anthropic) adding ~30 GW in 2028 alone and taking 25–40% of new compute in 2027. Compute remains extremely tight, with rising contract prices; US AI-related capex could approach $2 trillion annually, and revenue catch-up is feasible (only ~$3 trillion in AI revenue by 2031 would pose issues). The firm has raised 2027 datacenter forecasts despite skepticism.[6][16][17]

Anthropic’s October–November 2026 S-1 filings reveal enormous scaling commitments: hundreds of billions to nearly $900 billion in non-cancelable compute/infrastructure leases and purchases (many spanning to 2031–2045), targeting >3.5 GW by end-2026 and ~9.5–10 GW by end-2027 via deals with Google/Broadcom (TPUs), AWS (Trainium), AMD, and others. Compute costs represent a rising share of revenue (~58% in a recent quarter), underscoring a multi-year bet on demand growth.[7][18][19]

Jensen Huang/NVIDIA (August–November 2026 updates) emphasizes supply-constrained demand persisting into 2028, with FY2028 revenue growth guidance of ~70%. Per-GW revenue opportunity rises across generations (Hopper ~$18B, Blackwell ~$25B, Rubin ~$40B), and AI factories can generate substantial annual revenue relative to build costs. Power and memory bottlenecks are acknowledged but viewed as symptoms of robust demand outpacing supply.[8][20][21]

Elon Musk/SpaceX (October 2026 earnings commentary) outlines aggressive expansion: targeting ~10 GW incremental AI compute in 2027 (reaching at least 13 GW total by year-end, from ~2.5–3 GW end-2026), including a 440,000-GPU Vera Rubin cluster, alongside plans for excess power/cooling infrastructure. This aligns with broader xAI/SpaceX ambitions and includes expectations of very high ARR run-rates.[9][22]

Comparative Table (Sorted by Implied Peak or Deceleration Year)

Implied timelines are inferred from when sources see capex growth peaking, flatlining, or requiring major revenue justification (earlier = more cautious on sustained buildout). All data from post-May 2026 publications.

• ~2027 (early peak/flatline or growth deceleration): Wells Fargo (buildout risk after 2027; applications win in 2027); Goldman Sachs (growth slows sharply post-2027). Incentives: Sell-side research balancing stock ratings/PTs with client advisory; track record includes upward revisions amid demand strength but sensitivity to FCF/monetization. Maps to 2027 scenario (buildout peaks, shift to utilization/apps).
• ~2028 (revenue justification or continued high but slowing spend): Barclays (major revenue gap by end-2028); Goldman (continued rise but 12% growth); Morgan Stanley (high 2028 spend with potential downward bias). Incentives similar to above; track record shows repeated upward capex adjustments. Maps to 2028 scenario (sustained but maturing cycle with monetization pressure).
• 2028–2029+ (extended or accelerating cycle): UBS ($1.5T in 2028, longer cycle than priced, 2027 returns inflection); SemiAnalysis ($11T cumulative by 2029, massive 2028 additions); Anthropic (commitments through 2031+); NVIDIA/Huang (supply limits into 2028+, rising per-GW value); Musk/SpaceX (10+ GW additions in 2027, long runway). Incentives: Research/data firms (empirical modeling), vendors/labs (capital attraction and ecosystem growth). Track record: Consistent underestimation of prior buildout pace by markets/analysts. Maps to 2029 or 2031 scenarios (prolonged infrastructure boom, potential oversupply risks only later).

Implications for Market Participants and New Entrants

Banks’ nearer-term caution on 2027–2028 monetization and FCF could pressure valuations for pure infrastructure plays if revenue ramps disappoint, favoring selective exposure to applications, efficiency tools, or financially strong hyperscalers. Longer views from operators and specialists suggest durable demand tailwinds through at least 2029, driven by lab scaling and per-GW economics, supporting investment in power, memory, networking, and sovereign/neocloud capacity—but with execution risks around permitting, power, and financing.[4][6]

Competitors or new entrants should prioritize data on actual utilization, contract pricing trends, and regional power constraints over headline capex numbers; those aligned with extended-demand scenarios (e.g., specialized compute providers or efficiency solutions) may find more runway, while early-peak believers should stress-test against 2027–2028 revenue thresholds. Volatility around earnings, debt markets, and policy (e.g., 2027 debt limit) is likely to increase as the cycle matures.