TL;DR
The memory cycle's fundamentals are still printing records — Samsung's first-ever ₩100T+ quarter, Micron's biggest quarter in memory history — but every marginal signal rolled over this month: contract-price deceleration is now the consensus, spot is flat, and the equity tape is falling on the records. The new wildcard is China: CXMT arrived as the swing supplier the cycle needed to end, undercutting incumbents by half in the commodity segments where pricing was already bending.
Pricing: deceleration is now the consensus, not a forecast
The pricing story has shifted from "how high" to "how fast is the rate of increase dying," and the answer is: on every margin at once. TrendForce's Avril Wu put the machinery on record:
"Long-term agreements represent an increasingly higher proportion of suppliers' total output. With ceiling-price mechanisms built in, the rate of price increases has slowed down."
— TrendForce DRAM Market Bulletin, Oct 7
The numbers confirm it: conventional DRAM contract prices are expected up 10–15% QoQ in 4Q26, "slowing from a second-quarter surge of roughly 60%" (Reuters via Asahi), while spot trading "remains quiet and flat" (TrendForce bulletin). Bernstein goes further — high single digits in Q4, prices normalizing in 2028 with industry gross margins easing from ~90% to the high-70s (Investing.com). This is the textbook pre-turn signature: suppliers are now deliberately throttling increases to avoid demand destruction, LTAs are capping price discovery, and spot has gone sideways — three of the classic turn markers visible simultaneously.
What to watch: whether the next Micron margin guide cites price rather than costs — a price-cited guide-down would be the top signal.
Equities: the market is trading the second derivative
Every memory maker is printing records and every memory stock is falling anyway — the tape has moved from the level to the rate of change. Samsung guided Q3 operating profit to ₩107.4T, up 782% YoY and the first tech company ever above ₩100T in a quarter, and shares "dropped 0.7% Thursday morning" anyway (CNBC). eToro's Josh Gilbert captured the dynamic:
"Samsung has delivered a record profit and still fallen short of expectations, which tells you just how demanding the AI trade has become… When expectations are so high for memory makers, the market doesn't hand out points for effort."
— CNBC, Oct 8
It's a pattern: SK Hynix beat Q2 EPS consensus by nearly 92% and the stock fell 14.71% (Investing.com); Samsung is ~25% off its June record despite four straight record quarters (Reuters via Asahi); MU is -15.2% off its high with RSI 48 (MU market view, Oct 12
). Underneath, estimates are rolling over — the LSEG SmartEstimate for Samsung was slashed 7.7% since end-August, and 90-day EPS revisions are negative for both Koreans. This is what the equity tape does at prior memory peaks: discounting the turn two to three quarters before the P&L shows it.
What to watch: SK Hynix's Oct 27 print — another beat-and-drop confirms the market has fully pre-priced the deceleration.
CXMT: the swing supplier just showed up
Every memory cycle ends with swing supply arriving into record prices, and this cycle's swing producer is Chinese. On Oct 12, Samsung fell 3.6% and SK Hynix 3.4% after CXMT flagged progress on a 4F² memory-cell structure — the same design both Koreans are pursuing — with President Cao Kanyu saying CXMT "planned to introduce a new generation of DDR5 registered dual in-line memory modules (RDIMMs) based on the technology by the end of 2026," aimed at enterprise servers (Investing.com via Yahoo).
The undercutting is already live: 32GB DDR4-3200 ECC modules at $138 — "less than half what Samsung and Micron charge for comparable server" memory (XDA) — and substitution is happening at the margin. Summit Insights' Kinngai Chan: "Our industry checks indicate that an increasing number of OEMs and ODMs are adopting Chinese DRAM and NAND" (Reuters via Asahi). CXMT holds roughly 9–10% of global DRAM bit share, ranked fourth, and the incumbents are responding like incumbents at a peak: Samsung offering up to $560,000 retention bonuses against CXMT poaching (SammyFans). CXMT won't touch HBM4 economics for years — Futurum's Rolf Bulk: still "two to three generations behind" (StockTwits) — but it caps the commodity tail and hands price-sensitive OEMs an exit ramp.
What to watch: hyperscaler/server qualification of CXMT's end-2026 DDR5 RDIMMs — that moves this from headline risk to allocation risk.
Micron: the contracted model meets its first margin step-down — and a strike vote
Micron's FQ4 was the biggest quarter in memory history ($54.23B revenue, 80.7% GAAP gross margin), but the first guided sequential margin decline of the cycle just landed: FQ1 2027 guided to ~86.3% from 87%, "partly due to employee compensation costs" (Reuters via Asahi). The canary is the attribution — costs, not pricing, for now (Micron SWOT analysis
+1).
The labor story escalated sharply. On Oct 7, the Taoyuan union at Micron's Taiwan operations — its crucial DRAM/HBM production hub, where unions cover >80% of ~15,000 workers — secured strike authorization with 99% of ballots cast, demanding a permanent 15% profit-sharing mechanism (≈$14.9B on recent-quarter profit scale) and rejecting a package that included a T$1M cash bonus (Reuters via The Star; Epoch Times). A snap strike wasn't ruled out, with an Oct 19 Taipei rally planned (Taipei Times). Samsung's precedent — averting a strike by distributing 10.5% of chip-division operating profit — shows where this ends in a record-profit year. The take-or-pay SCA architecture converts price risk into volume and credit risk, but it does nothing about labor and cost inflation eating the marginal margin at peak profitability.
What to watch: the Taiwan standoff — an actual walkout is a near-term supply/pricing tailwind; a settlement at the Samsung precedent sets the industry's cost baseline.
What surprised us
- The HBM margin inversion. In FQ4, the Cloud Memory unit carrying HBM earned an 83% gross margin versus 90% in both other business units — the "structural" product is currently the lowest-margin line because it was priced under year-old contracts while commodity DRAM/NAND repriced quarterly (Micron SWOT analysis+1). The record quarter came mostly from the cyclical part of the business.
- Even the bulls changed the basis of the bull case. Gilbert again: buyers committing to multiyear supply agreements give Samsung "greater visibility over demand in a business where investors have always worried about the next downturn" (CNBC) — the market no longer pays for the shortage, it pays for cycle-dampening contracts.
- CXMT's timing is the cruelest part. Its bits land in commodity DDR4/DDR5 exactly where three other brakes are already applied — LTA ceilings, deliberate supplier throttling, and flat spot (TrendForce bulletin).
- The overbuild signature is still absent at Micron itself. Capex runs ~31% of operating cash flow with ~$68B net cash — the classic late-cycle tell (>45% of OCF while margins fall) isn't in the numbers yet (Micron SWOT analysis+1).
Open threads worth a vote
The Memory Supercycle — Delta Briefing
Since last time
Promoted
- The equity tape. Last cycle it was one paragraph of valuation skepticism inside the Micron section (MU at 14.45x P/E, RSI 75). Now it's a full section: every maker printing records, every stock falling, estimates rolling over.
- Taiwan labor. Was a passing mention (strike-rights vote concluding Oct 6, ~$6.6B claim). Now it's half the Micron section plus an open thread: strike authorization secured with 99% of ballots, a $14.9B profit-sharing demand, and an Oct 19 rally date.
Escalated
- Pricing deceleration. Last time the +10-15% Q4 print was the tail of TrendForce's arc — a forecast to watch. Now it's consensus, spot has gone flat, and Bernstein is calling high single digits and 2028 normalization. The "what to watch" has moved from whether deceleration prints to which line item Micron blames.
Demoted
- The take-or-pay contract book. Was a full section ($150B RPO, SCAs 16→26, $32B customer commitments, the capex tell). Now a single sentence: the SCA architecture converts price risk into volume and credit risk "but does nothing about labor."
- Micron's capex tell. The greenfield-cleanroom-into-2028 "everyone-builds-at-once" argument is gone from the body — replaced by a surprise bullet arguing the overbuild signature isn't in Micron's numbers yet (capex ~31% of OCF, ~$68B net cash). That's a framing reversal, not a repetition.
Disappeared
- The entire HBM4 negotiation story. Samsung's 3x ask, SK Hynix stripping price caps from LTAs, TrendForce's +121% HBM ASP projection, Bernstein's dissent, Micron's 2027 lock-in quote — none of it appears. The only trace is a passing note that CXMT is "two to three generations behind" in HBM4.
- Nvidia de-speccing. Last cycle's "cycle tell hiding in plain sight" (Rubin Ultra cut to 8-layer/192GB, ~33% less memory) and the HBM spot-to-contract 4-5x premium get zero mention.
- Consumer demand destruction (PC brands cutting SSD capacity) and the DDR4-over-DDR5 anomaly ($148 vs $133) — both gone entirely.
Unchanged
- The headline conventional DRAM number: 4Q26 contract prices +10-15% QoQ. Same figure as last time — but it has shifted from forecast to consensus, so don't skim the framing.
Entirely new: CXMT as the cycle's swing supplier — its own section below.
Pricing: deceleration is now the consensus, not a forecast (Escalated)
The story has shifted from "how high" to "how fast the rate of increase is dying" — on every margin at once. TrendForce's Avril Wu put the machinery on record:
"Long-term agreements represent an increasingly higher proportion of suppliers' total output. With ceiling-price mechanisms built in, the rate of price increases has slowed down."
— TrendForce DRAM Market Bulletin, Oct 7
Conventional DRAM contract prices are expected up 10–15% QoQ in 4Q26, "slowing from a second-quarter surge of roughly 60%" (Reuters via Asahi), while spot trading "remains quiet and flat" (TrendForce bulletin). Bernstein goes further — high single digits in Q4, prices normalizing in 2028 with industry gross margins easing from ~90% to the high-70s (Investing.com). Three classic turn markers are now visible simultaneously: deliberate supplier throttling, LTA-capped price discovery, and sideways spot.
What to watch (changed): the previous watch — TrendForce's January print — is superseded by whether the next Micron margin guide cites price rather than costs. A price-cited guide-down would be the top signal.
Equities: the market is trading the second derivative (Promoted)
Every memory maker is printing records and every memory stock is falling anyway. Samsung guided Q3 operating profit to ₩107.4T — up 782% YoY, the first tech company ever above ₩100T in a quarter — and shares "dropped 0.7% Thursday morning" anyway (CNBC). eToro's Josh Gilbert:
"Samsung has delivered a record profit and still fallen short of expectations, which tells you just how demanding the AI trade has become… When expectations are so high for memory makers, the market doesn't hand out points for effort."
— CNBC, Oct 8
The pattern: SK Hynix beat Q2 EPS consensus by nearly 92% and fell 14.71% (Investing.com); Samsung is ~25% off its June record despite four straight record quarters (Reuters via Asahi); MU is -15.2% off its high with RSI 48 (MU market view, Oct 12) — note the swing from last cycle's RSI 75 at $1,074.89. Underneath, the LSEG SmartEstimate for Samsung was slashed 7.7% since end-August, and 90-day EPS revisions are negative for both Koreans. This is what the tape does at prior memory peaks: discounting the turn two to three quarters before the P&L shows it.
What to watch (carried over): SK Hynix's Oct 27 print — another beat-and-drop confirms full pre-pricing. (The Samsung Oct 8 watch from last cycle has resolved: it printed, and the beat-and-drop happened.)
CXMT: the swing supplier just showed up (New)
Every memory cycle ends with swing supply arriving into record prices, and this cycle's swing producer is Chinese. On Oct 12, Samsung fell 3.6% and SK Hynix 3.4% after CXMT flagged progress on a 4F² memory-cell structure — the same design both Koreans are pursuing — with President Cao Kanyu saying CXMT "planned to introduce a new generation of DDR5 registered dual in-line memory modules (RDIMMs) based on the technology by the end of 2026," aimed at enterprise servers (Investing.com via Yahoo).
The undercutting is already live: 32GB DDR4-3200 ECC modules at $138 — "less than half what Samsung and Micron charge for comparable server" memory (XDA) — and substitution is happening at the margin. Summit Insights' Kinngai Chan: "Our industry checks indicate that an increasing number of OEMs and ODMs are adopting Chinese DRAM and NAND" (Reuters via Asahi). CXMT holds roughly 9–10% of global DRAM bit share, ranked fourth, and the incumbents are responding like incumbents at a peak: Samsung offering up to $560,000 retention bonuses against CXMT poaching (SammyFans). CXMT won't touch HBM4 economics for years — Futurum's Rolf Bulk: still "two to three generations behind" (StockTwits) — but it caps the commodity tail and hands price-sensitive OEMs an exit ramp.
What to watch: hyperscaler/server qualification of CXMT's end-2026 DDR5 RDIMMs — that moves this from headline risk to allocation risk.
Micron: the contracted model meets its first margin step-down — and a strike vote (Escalated)
The margin guide itself isn't new — last cycle's briefing carried the FQ1 2027 ~86.3% guide vs FQ4's 87% and the "FY2027 floor" framing. What's new is the attribution and the context: FQ4 was the biggest quarter in memory history ($54.23B revenue, 80.7% GAAP gross margin), the decline is "partly due to employee compensation costs" (Reuters via Asahi), and the canary is that it's costs, not pricing, for now (Micron SWOT analysis+1).
The labor story escalated sharply. On Oct 7, the Taoyuan union at Micron's Taiwan operations — its crucial DRAM/HBM production hub, where unions cover >80% of ~15,000 workers — secured strike authorization with 99% of ballots cast, demanding a permanent 15% profit-sharing mechanism (≈$14.9B on recent-quarter profit scale) and rejecting a package that included a T$1M cash bonus (Reuters via The Star; Epoch Times). A snap strike wasn't ruled out, with an Oct 19 Taipei rally planned (Taipei Times). Samsung's precedent — averting a strike by distributing 10.5% of chip-division operating profit — shows where this ends in a record-profit year. The take-or-pay SCA architecture converts price risk into volume and credit risk, but it does nothing about labor and cost inflation eating the marginal margin at peak profitability.
What to watch: the Taiwan standoff — an actual walkout is a near-term supply/pricing tailwind; a settlement at the Samsung precedent sets the industry's cost baseline.
What surprised us
- [NEW] The HBM margin inversion. In FQ4, the Cloud Memory unit carrying HBM earned an 83% gross margin versus 90% in both other business units — the "structural" product is currently the lowest-margin line because it was priced under year-old contracts while commodity DRAM/NAND repriced quarterly (Micron SWOT analysis+1). The record quarter came mostly from the cyclical part of the business.
- [NEW] Even the bulls changed the basis of the bull case. Gilbert again: buyers committing to multiyear supply agreements give Samsung "greater visibility over demand in a business where investors have always worried about the next downturn" (CNBC) — the market no longer pays for the shortage, it pays for cycle-dampening contracts.
- [NEW] CXMT's timing is the cruelest part. Its bits land in commodity DDR4/DDR5 exactly where three other brakes are already applied — LTA ceilings, deliberate supplier throttling, and flat spot (TrendForce bulletin).
- [UPDATED] The overbuild signature is still absent at Micron itself. Capex runs ~31% of operating cash flow with ~$68B net cash — the classic late-cycle tell (>45% of OCF while margins fall) isn't in the numbers yet (Micron SWOT analysis+1). This reverses last cycle's capex-tell framing (the $25B 1H FY27 build with cleanrooms into 2028-29) — worth reconciling before you trade on either.
Open threads
New on the board:
Fate of last cycle's threads:
- HBM spot-to-contract premium collapse — dropped from the list with no update given; the underlying data (4-5x premium) also vanished from the body. Treat as unresolved, not closed.
- Consumer demand destruction — dropped, and the supporting evidence (PC SSD capacity cuts) is gone from the body entirely.
- Take-or-pay SCAs: dampen the bust or relocate default risk? — dropped from the vote list but partially absorbed into the Micron section's one-liner (SCAs convert price risk into volume and credit risk, do nothing about labor). Still unresolved.