Skip to main content
Invest AI
From market noise to cited evidence

Hot-Topic InvestigationsIssue No. 3

Public research

Has the Memory Cycle Topped? I Checked Five Chains of Evidence, One by One

Independent investigation · Published October 26, 2026 · Evidence as of October 26, 2026

On the morning of October 26, a widely watched channel asked whether the memory cycle is about to top out, and argued that the market is already pricing 2028. I hold Micron (MU) and other memory names, so I did not borrow the question's answer — I checked the evidence myself, chain by chain: HBM contracts, contract and spot prices, inventories, the capex timetable, and hyperscaler demand. My verdict differs from the headline in one important way: the evidence does not show a cycle top. It shows something subtler — the slope of the price climb cresting while price levels keep rising. Those are not the same statement, and they should not be traded as one.

Cover — Hot-Topic Investigations No. 3: has the memory cycle topped? 2027 HBM ASP revised up 121%; Q4 contract prices still rising; over 75% of Micron's FY2027 shipments committed under long-term agreements.
Hot-Topic Investigations No. 3 — the “memory cycle top” claim, checked against five chains of evidence. Data as of October 26, 2026.

The conclusion first

After checking all five chains, the evidence available on October 26 does not support “the cycle has topped”: 2027 HBM contract prices were revised up, not down; fourth-quarter contract prices are still rising; more than 75% of Micron's fiscal-2027 shipments are locked into long-term agreements; and the new capacity everyone fears contributes no meaningful bit supply until after 2028. The one genuine plateau signal is in the slope, not the level — Micron guided gross margin slightly down, and management itself says the pace of price increases is moderating. The mid-October selloff in memory stocks priced a 2028-2029 supply scenario, not a 2026 fundamental turn. Those two things need to be kept apart.

  • HBM: TrendForce raised its 2027 blended HBM ASP forecast to +121% year on year (Sep 29); Micron confirmed most of its calendar-2027 HBM bit supply is already agreed, at prices significantly higher than 2026.
  • Contract prices: TrendForce sees Q4 2026 conventional DRAM +10-15% quarter on quarter, +15-20% including HBM, NAND +15-20%, and enterprise SSD +23-28%.
  • Spot prices have split: DDR5 stays tight and sellers hold back, while NAND wafers soften and DDR4 trading thins — worth watching, not yet a reversal pattern.
  • Capex: all three makers are spending heavily, but their own timetables put the supply contribution after 2028. The top-call is early by its own evidence.

The five chains of evidence, checked one by one

Before checking anything, I fixed what a real top would have to look like, so the standard cannot move mid-argument: contract-price gains turning negative or negotiations breaking down; spot prices leading down for weeks; inventories building at makers and customers at the same time; big customers cutting capex or delaying pull-ins; and new capacity arriving ahead of schedule. Here is how each chain scores against that standard today.

#Evidence chainVerdictBasis
1HBM contracts — 2027 prices are being negotiated up, not downNo top patternTrendForce (Sep 29) raised 2027 blended HBM ASP to +121% YoY; Micron's Sep 30 call confirms most CY2027 HBM bits already agreed at significantly higher prices, and over 75% of FY2027 shipments committed under LTAs
2Contract prices are still rising — but the spot market has splitRising, with a split underneathTrendForce Sep 30 outlook for Q4 2026 (DRAM +10-15%, incl. HBM +15-20%, NAND +15-20%, enterprise SSD +23-28%); Oct 21 spot update shows DDR5 tight, DDR4 trading thin, NAND wafer -1.28% on the week
3Inventories — no evidence of a build at the makersNo build shown; precise weeks unverifiedSK hynix and Samsung 2Q26 official releases show record results and LTA signings, no inventory-turn deterioration; the “high inventory” lines in spot reports describe module makers in the NAND chain, a different world from maker server contracts
4Capex is jumping at all three makers — but the new supply lands after 2028A real risk, in the wrong yearMicron FY26 capex about $27B, FY27 guidance above the mid-$40B range with increments mostly in greenfield construction and first meaningful output after 2028; SK hynix 2026 capex in the high-₩40T range, Yongin phase-1 to Feb 2027; Samsung Q2 capex ₩16.8T
5Demand — hyperscaler capex is still being revised up, partly because memory costs moreNo demand breakPost-July-2026 guidance: Amazon about $220B, Alphabet $195-205B, Meta $130-145B (later fine-tuned to about $135-145B), Microsoft about $175B after an accounting re-presentation; combined about $720-745B, per TMT Finance and Motley Fool summaries in agreement

1. HBM contracts — 2027 prices are being negotiated up, not down

No top pattern

TrendForce's September 29 revision puts 2027 blended HBM average selling prices up 121% year on year, on persistent supply constraints, a rising share of pricier HBM4, and HBM4E joining in the second half of 2027. EE Times Asia, Aju Press and The Investor (citing KB Securities) all carry the same figure.

Micron said the same thing from the other side of the table on its September 30 earnings call: the vast majority of its calendar-2027 HBM bit supply is already under agreement, at prices significantly higher than those negotiated for 2026; more than 75% of its total fiscal-2027 shipment plan is committed under long-term agreements, and customer allocation discussions have already been pushed into 2028.

Samsung's 2027 HBM4 talks point the same way: Korean media report asking prices in the range of a bit over $4 per gigabit, roughly three times its HBM3E level (about $1.50/Gb). That is an asking price, not a transacted price, and Samsung has not confirmed individual terms — I use it as direction evidence only, not as a settled number. Long-term contracts, prepayments and upward price revisions are the standard moves of a seller's market. This chain shows no top.

2. Contract prices are still rising — but the spot market has split

Rising, with a split underneath

TrendForce's September 30 outlook for the fourth quarter of 2026: conventional DRAM contract prices +10-15% quarter on quarter, +15-20% including HBM, NAND +15-20%, and enterprise SSDs +23-28% — the only category flagged as accelerating, on hyperscalers expanding purchases for AI inference. Multiple financial outlets carried the same figures.

Spot is a different picture. In TrendForce's October 21 spot update, DDR5 supply is tight and sellers are holding back — buyers cannot get volume even when they bid up — but branded DDR4 trading is thin, module makers are tightly controlling cash and inventory, mainstream DDR4 1Gx8 spot averaged $46.357 (+0.84% on the week), and NAND 512Gb TLC wafer spot fell 1.28% in a week to $18.592, with module-maker inventories high and trading quiet. The early-October spot report showed the same pattern: DRAM sideways, NAND wafers down 2.9% in a week.

Micron's fiscal Q4 (ended September 3) actuals confirm price levels still running high: revenue of $54.23B (up 379% year on year), non-GAAP gross margin of 87.0%, DRAM prices up high-teens sequentially and NAND prices up about 30%. My read: contract prices (big customers, servers) and spot prices (module makers, consumer) are diverging. Cycle turns historically tend to start in spot — but today's spot weakness is concentrated in NAND wafers and DDR4 volumes, which looks like buyers stepping back after being squeezed by price, not sellers cutting prices to move stock. The real turn signal would be DDR5 spot and server contract negotiations. Neither has cracked yet.

3. Inventories — no evidence of a build at the makers

No build shown; precise weeks unverified

At the makers, none of the big three's latest results releases an inventory-build signal. SK hynix's second quarter (released July 29) printed revenue of ₩79.3187 trillion (+257% YoY), operating profit of ₩60.5426 trillion (+557%) and a 76% operating margin, with multi-year LTAs signed with about ten key customers. Samsung's second quarter set records as well — revenue ₩171.5 trillion, operating profit ₩89.5 trillion, memory records on both lines — with official language pointing to continued strong second-half server demand and a market still short of supply.

On the customer and channel side, TrendForce's spot reports repeatedly flag module makers' inventories as high. That is the NAND spot chain's condition. It is not the makers' server contract chain, and merging the two into one inventory story is a category error.

An honest gap, stated plainly: the specific inventory-week figures circulating in the market — “makers hold about four weeks”, “suppliers under ten days” — reach me only through single second-hand summaries. They do not meet my two-source standard, so this report does not use them (see the not-used list below). On this chain I can support “no evidence of a build”. I cannot honestly give a precise week count.

4. Capex is jumping at all three makers — but the new supply lands after 2028

A real risk, in the wrong year

This is the top-callers' strongest argument, and it deserves a timetable rather than a shrug. Micron's fiscal-2026 capex came to about $27B after several in-year raises, and fiscal-2027 guidance sits above the mid-$40B range — but the increment is mostly greenfield fab construction. The company itself says the new cleanrooms supply the period after late 2028: first wafers from the first Idaho fab around mid-2027, the second at the end of 2028, and the first New York fab's output not until 2030.

SK hynix raised its 2026 capex guidance into the high-₩40 trillion range (call commentary) and pulled its Yongin phase-1 cleanroom opening forward to February 2027; its official 2Q26 release simultaneously stresses “CapEx Discipline” and phased investment paced to customer demand. Samsung spent ₩16.8 trillion on capex in the second quarter alone (about ₩15.4 trillion of it in DS) and is negotiating five-year rolling LTAs with its top-ten tech customers, planning to lock a substantial share of capacity into contracts (the 60-70% figure is Korean-media reporting of management commentary — direction evidence).

So the three makers genuinely are expanding at once, and that is the seed of a 2028-2029 supply risk — the top-callers are not anxious about nothing. But on each maker's own timetable, this capacity's real contribution to bit supply lands after 2028. Taking a 2028 risk and using it to declare the cycle topped in Q4 2026 does not line up on the calendar.

5. Demand — hyperscaler capex is still being revised up, partly because memory costs more

No demand break

The four hyperscalers' 2026 capex guidance after the July earnings season: Amazon raised to about $220B, Alphabet to $195-205B, Meta to $130-145B (fine-tuned again after Q3 to about $135-145B), and Microsoft to about $175B — its step down from about $190B was purely a lease-accounting re-presentation, not a smaller build plan. Combined, roughly $720-745B, a clear upward revision from the start-of-year figures, with TMT Finance, Motley Fool and other summaries in agreement.

One detail explains the chain better than any total: Amazon's CFO named higher memory costs as one of the reasons capex was raised from about $200B to about $220B. The demand side is not only not flinching — it is paying up for the more expensive memory.

As long as hyperscaler capex guidance keeps being revised up and long-term prepayments keep growing, the “demand suddenly goes out” precondition for a cycle top is not met. This is also the most fragile link in the whole structure: the quarter capex guidance gets cut, this narrative can reverse very fast — which is why it is item one on my watch list.

Flagged, not settled

  • Makers' precise inventory weeks — the figures in circulation (“about four weeks”, “under ten days”) are single-sourced; I report direction only: no build is evidenced, no week count is given.
  • The DDR4 contract monthly-gain sequence (14.3% in July narrowing to about 4.0% in September) comes from a single analyst summary — kept as a reference signal, not a conclusion input.
  • CXMT's actual mass-production capability on its 4F² DDR5 — no first-hand specification, yield or customer evidence exists yet; today it is a sentiment variable, not a supply variable.

The other side's real cards

A check that only lays out one side is not a check. Here is what the top-callers actually hold:

The slope may have crested — and for the stock, that can be enough

Micron guided fiscal Q1 2027 non-GAAP gross margin to about 86.25%, below fiscal Q4's 87.0% — the first downward guide of this cycle. Management calls Q1 the fiscal-2027 margin low, with recovery after, but “the first step down” is itself a signal, and management says the pace of price increases is moderating. Contract monthly gains are narrowing as well. A crest in the growth rate is not a crest in price levels — but for a share price, the first can be enough. I record the distinction rather than blur it.

2028 already has an institution's price on the decline

Counterpoint Research forecasts DRAM prices down about 23% in 2028 versus 2027 (reported by CNBC on October 15). The same work stresses that even after that fall, prices would remain about 230% above 2025, and that three-to-five-year LTAs with price floors and ceilings make the down-leg look nothing like the old crashes. I keep it labelled as what it is: a single firm's forecast, not a multi-house consensus.

Consumers are being priced out — and the stocks already fell

Reports of PC and phone shipments hurt by memory costs keep coming; consumer weakness is real, currently outweighed by server demand. CXMT's October announcement of a 4F²-architecture DDR5 server module by year-end triggered a round of selling, though analysts such as Rosenblatt publicly question its real supply impact — no specifications, yield or customer evidence — and its HBM scale and access to advanced equipment remain constrained. In mid-October the stocks already fell hard in a week. The market is pricing 2028-2029. That half of the originating title is right; the “cycle top” half is, on today's evidence, called too early.

The calendar problem

Every strong card on the other side is dated 2028 or later — the supply wave, the priced decline, the capacity math. Every current-quarter card — contract prices, HBM pricing, committed volumes, capex guidance — still points up. A 2028 risk reprices a stock. It does not date a cycle top in the fourth quarter of 2026.

Browse the Why I Own It research series →

Where I differ from the loudest version

  1. With the “cycle top” framing: the standard top pattern — contract prices rolling over, inventories building, demand guidance falling — is absent from every piece of first-hand evidence I checked. What is present is a slope change. I record it as a slope change, no more and no less.
  2. With the “market is already pricing 2028” framing: I agree — and that is precisely why the stock move and the industry verdict must be separated. The October selloff priced a 2028-2029 supply scenario, not a 2026 fundamental turn.
  3. With the “pays back in about three years” valuation shorthand for the Korean pair: the multiples behind it are single-site computations from consensus estimates on inconsistent bases. I dropped the precise multiples (see below) and keep only the qualitative point — the market is already discounting peak earnings.

My judgment, in one paragraph

Strongly supported: as of October 26, 2026, the industry's fundamentals show no standard cycle-top pattern — contract prices are still rising, 2027 HBM long-term pricing was revised sharply up, makers' output is locked into long-term contracts, hyperscaler capex is still being raised, and the new supply that matters lands after 2028. Supported at medium strength: the steepest part of the climb may be behind us — a first down-tick in margin guidance, narrowing monthly gains, soft NAND spot. A long plateau in price levels can follow a crest in slope, and the difference matters. Not concluded, for lack of evidence: precise maker inventory weeks, an exact 2028 decline figure, and CXMT's real production capability. I hold MU and related memory names; that position is disclosed here, and the judgment above follows the evidence, not the position.

What I am watching in the next 30 days

  • SK hynix Q3 2026 results (Oct 28): its language on 2027 HBM and on inventories
  • Final Q4 contract settlements: do they land inside TrendForce's forecast ranges?
  • Hyperscaler earnings (Oct 28-29): capex guidance — still rising, or the first cut?
  • TrendForce's Q1 2027 contract outlook: the direction of the first print for 2027

How I checked this

I listed the claim actually circulating — that the memory cycle is topping — then fixed in advance what evidence a real top would require, and checked five chains against it: HBM contracts, contract and spot prices, inventories, capacity timing, and end demand. Every figure was cross-checked against at least two independent sources, company releases and calls first. Where the two sources disagreed, I followed the primary one; where no primary figure could be obtained, I dropped the number instead of estimating it. Figures I checked but chose not to use are listed at the bottom of this page, with the reason for each.

Sources

Every figure on this page comes from our own investigation of October 26, 2026. Each number was cross-checked against at least two sources before publication; where sources disagreed, the primary source governed, and where a primary figure could not be obtained, the number was dropped rather than estimated. The originating video is a topic lead only — none of its figures or its narrative is cited or adopted anywhere on this page.

  1. TrendForce (Sep 29, 2026) — 2027 blended HBM ASP forecast raised to +121% year on year on persistent supply constraints, a rising HBM4 mix and HBM4E from 2H27; reported by EE Times Asia eetasia.com
  2. Aju Press (Oct 2, 2026) — “HBM heyday to continue — albeit at consumers' expense” (2027 HBM pricing and supply tightness) ajupress.com
  3. Aju Press (Oct 1, 2026) — “Micron signals longer AI bonanza for Samsung, SK hynix”: Micron fiscal Q4 revenue $54.23B; Q1 FY27 guidance $61.5B ±$1.5B at about 86.25% adjusted gross margin; more than 75% of FY27 shipments committed; SCA commitments about $32B, RPO about $150B; most CY2027 HBM bits agreed at significantly higher prices; 2027-28 supply-demand tighter than 2026 ajupress.com
  4. The Investor / Korea Herald (Oct 7, 2026) — “Samsung's HBM prices tipped to more than double in 2027” (citing KB Securities; HBM4 asking levels about 3x HBM3E per gigabit — asking, not transacted) theinvestor.co.kr
  5. TrendForce (Sep 30, 2026) — Q4 2026 contract-price outlook: conventional DRAM +10-15% quarter on quarter, DRAM including HBM +15-20%, NAND +15-20%, enterprise SSD +23-28%; as summarised by Guru3D and Evertiq (Oct 2, 2026) guru3d.com · evertiq.com
  6. TrendForce spot-price update (Oct 21, 2026) — DDR5 supply tight and sellers holding back; mainstream DDR4 1Gx8 spot +0.84% on the week to $46.357; NAND 512Gb TLC wafer spot -1.28% to $18.592, module-maker inventories high and trading thin trendforce.com
  7. TrendForce spot-price update (Oct 7, 2026) — DRAM spot largely sideways; NAND wafer -2.90% that week (same series as the Oct 21 update) trendforce.com
  8. Micron Technology — fiscal Q4 2026 results release (Sep 30, 2026): revenue $54.23B, non-GAAP EPS $33.42, non-GAAP gross margin 87.0%; fiscal Q1 2027 guidance revenue $61.5B ±$1.5B and non-GAAP gross margin about 86.25%; cross-checked against CNBC's Sep 30 report (actuals vs consensus) cnbc.com
  9. SK hynix — 2Q26 results release (Jul 28-29, 2026, via PR Newswire): revenue ₩79.3187 trillion (+257% YoY), operating profit ₩60.5426 trillion (+557%), 76% operating margin; multi-year LTAs with about ten key customers; 2026 capex guided to the high-₩40 trillion range, Yongin phase-1 cleanroom pulled forward to February 2027, “CapEx Discipline” language in the official release prnewswire.com
  10. Samsung Electronics — 2Q26 results release (Jul 29, 2026): revenue ₩171.5 trillion, operating profit ₩89.5 trillion, both records; Q2 capex ₩16.8 trillion (DS about ₩15.4 trillion); HBM4 to exceed 60% of HBM sales in 2H26 on the company's outlook; full text via LedInside, reported by TelecomLead ledinside.com · telecomlead.com
  11. TMT Finance — “2026 hyperscaler capex tops US$700bn” (analysis of the July 2026 earnings-season guidance: Amazon about $220B, Alphabet $195-205B, Meta $130-145B, Microsoft about $175B after a lease-accounting re-presentation) tmtfinance.com
  12. Motley Fool Research — “How Much Are AI Companies Spending on Data Centers?” (hyperscaler capex summary; Q2 2026 combined actual $171.2B, +78.6% YoY; Amazon CFO citing higher memory costs among the reasons for the $200B to about $220B raise) fool.com
  13. CNBC (Oct 15, 2026) — “Memory stocks are getting hammered. China and a 2029 oversupply are the big worries”: Counterpoint Research's forecast of DRAM prices about -23% in 2028 vs 2027 (still about +230% vs 2025), LTA price collars, the CXMT dispute and Rosenblatt's public scepticism (via BizToc repost noting the original) biztoc.com
  14. Hudson Labs — Micron earnings-call research summary (fiscal-2027 capex above about $45B with the increment in fab construction; management describing the pace of price increases as moderating) hudson-labs.com
  15. ad-hoc-news / Yahoo Finance (Jul 29, 2026) — SK hynix 2Q26 call summary (2026 capex in the high-₩40 trillion range, Yongin phase-1 pulled forward, about ten LTA customers) ad-hoc-news.de

Checked but not used

  • “Makers hold about four weeks / suppliers under ten days” of inventory, and a “2026 DRAM shortfall of 4.9%” — single second-hand summaries (an investment blog and forum reposts), no TrendForce or company source. Dropped; the text states only that no build is evidenced.
  • The three-way HBM per-gigabit price table (Samsung $4.50, Micron $4.00, SK hynix $3.60-3.80) — sourced to a social post citing an “expert interview”, no first-hand evidence. The table is dropped; Samsung's roughly-3x level appears only as Korean-media direction evidence, labelled asking price, not transacted.
  • Counterpoint's “2028 DRAM -23%” is a single firm's forecast. It is kept, but labelled as such in the text and never presented as a multi-house consensus.
  • Valuation multiples such as MU at about 4.2x FY2028 consensus EPS and the Korean pair at about 2.5-3x 2027 estimates — the arithmetic behind the “three-year payback” shorthand. Single analysis sites computing from consensus on inconsistent bases. Precise multiples dropped; only the qualitative point is kept.
  • Exact stock-drop percentages for the mid-October week (SK hynix -24% in a week, MU -14.7% from the October 9 close) — recorded by a single analysis site; CNBC confirms the heavy selloff only qualitatively. Used qualitatively, without the percentages.
  • The “over $500B SK hynix-NVIDIA partnership” and “memory is about 30% of hyperscaler capex” figures — mixes of multi-year framework agreements and estimates at inconsistent source levels. Dropped; not used anywhere in the text.
  • The DDR4 contract monthly-gain sequence (14.3% to 4.0%) — a single analyst summary; downgraded to a reference signal and labelled as such where it appears.
  • Every figure, and the “pricing 2028 early” narrative, from the originating TradingKey video — a topic radar only under this column's standing rule. Never cited, never adopted.

Data as of October 26, 2026. Q4 2026 contract-price figures are TrendForce forecasts published September 30, 2026; company figures are from the cited releases and calls. This page is a fixed research record and is not updated with live data.

For research and education only. This page is our own independent investigation; the views are ours, and named institutions' figures are summaries of cited reporting. This page is not investment advice, not a recommendation, and not a promise of returns. Read the full disclaimer before acting on anything here.

Full disclaimer →

Weekly digest: AI-map layer changes and new public 13F filings, each with its source. No score, no grade, no trade signal. Unsubscribe anytime.

Browse stock research →