TL;DR
The memory cycle is now running two contradictory stories at once: conventional DRAM pricing is decelerating fast (Q4 2026 forecasts at +10-15% QoQ, down from +90-95% in Q1), while the 2027 HBM4 negotiation has hardened in suppliers' favor — Samsung asking 3x+, TrendForce projecting +121% HBM ASPs. Meanwhile the biggest HBM buyer is literally redesigning its flagship chip to need less memory, and Micron's take-or-pay contract book is being pitched as the thing that abolishes the cycle — right as its own capex plans join the everyone-builds-at-once pattern.
Pricing: the split market is now official
The pricing tape has bifurcated into a decelerating commodity side and an accelerating HBM side, and both are setting records. TrendForce's September 30 forecast marks the full arc: conventional DRAM contract prices went +90-95% (Q1) → +58-63% (Q2) → +13-18% (Q3) → +10-15% (Q4) — the rate of increase collapsed by roughly an order of magnitude even as absolute prices keep climbing (TrendForce, Sept 30, 2026). The consumer edge has already bent into outright demand destruction:
"PC brands are also reducing SSD capacities in mainstream models to lower BOM costs, resulting in declines in both procurement volumes and average capacity per device."
— DRAM Contract Pricing Deceleration
+1
Micron's own print confirms the convergence: DRAM ASPs up high-teens QoQ — the same band as TrendForce's 4Q26 conventional forecast — and FQ1 2027 gross margin guided to 86.25%, below FQ4's 87%, which management called the FY2027 floor with "a more moderate rate of price increases" beyond (Micron FQ4 Prepared Remarks). One anomaly worth staring at: DDR4 8GB now trades above DDR5 8GB ($148 vs $133) — legacy scarcity, not forward demand (Moomoo/Goldman Sachs).
What to watch: TrendForce's January forecast — a flat or single-digit Q1 2027 print extends the arc; an outright decline is the turn itself.
HBM4: the stall resolved in suppliers' favor
Last cycle flagged the Nvidia HBM4 negotiation as stalled; the evidence now says the stall was interim noise in a negotiation settling high. Samsung has put forward a 2027 HBM4 ask in "the mid-to-high $4 range per 1Gb" — more than three times the ~$1.50/Gb HBM3E mainstream price (Xenospectrum), and TrendForce projects HBM ASPs up 121% next year, with a 36GB HBM4 stack going from ~$600 to ~$1,300 (Seoul Economic Daily).
"We have completed agreements for the vast majority of our calendar 2027 HBM bit supply with significant price increases year over year, narrowing the gross margin gap with conventional DRAM."
— Micron FQ4 Prepared Remarks
Micron's lock-in strengthens the other suppliers' hands, and SK Hynix is reportedly going further — removing price caps from its HBM4 LTAs entirely (MoveSurge). The lone dissent: Bernstein cut its 2027 HBM hike expectations and trimmed SK Hynix's target from ₩3.3M to ₩2.7M (BigGo). Note the seed of the next overshoot inside the bull case: HBM's share of global DRAM wafers goes from ~20% to ~30% next year per Samsung's own memory EVP (Seoul Economic Daily) — the capacity valve re-opening at peak pricing.
What to watch: final 2027 settlements — whether they print near Samsung's 3x ask or Bernstein's marked-down hike.
Demand: the biggest buyer is rationing itself
The sharpest demand-side crack isn't a price decline — it's Nvidia redesigning Rubin Ultra around memory scarcity. After abandoning its 12-Hi HBM4E design, Nvidia is testing three fallbacks, and the front-runner is the memory-cut option: an 8-layer HBM4/192GB setup, ~33% below Rubin's 288GB (Shattered.io). Meanwhile spot HBM trades at a 4-5x premium to contract — a 36GB HBM3E module at ~$2,100 vs a $300-400 LTA range — the panic-price signature of buyers who missed allocation (Shattered.io).
This is what a shortage peak looks like from the demand side: not falling prices, but the largest customers engineering the scarce input out of their products. Micron fielded a direct question about "one large customer de-speccing HBM" on its FQ4 call and deflected to "latent demand," while conceding server content growth is running at "a modestly lower rate... than prior expectations" (Micron FQ4 Prepared Remarks). Suppliers asking 3x+ for 2027 HBM4 while their biggest customer cuts 33% of memory from its next flagship are both true in a shortage — but the second one is what ends it.
What to watch: the HBM spot-to-contract premium — a collapse toward 1-2x would mark the turn; a further blowout means the shortage is still deepening.
Micron: the contract book that claims to abolish the cycle
Micron's FQ4 is the fullest expression yet of the take-or-pay model: SCAs grew from 16 to 26 in one quarter, ~$150B in remaining performance obligations, $32B in customer financial commitments (mostly cash deposits), agreements extending into 2031, and more than 75% of 2027 output already committed (Micron FQ4 Prepared Remarks). CFO Mark Murphy's claim: "even at floor prices, we expect margins meaningfully above any prior cycle peak margins" (Micron FQ4 Prepared Remarks).
But the capex tell is now official on Micron's own paper: 1H FY27 capex of ~$25B with 2H higher, skewed to construction accelerating cleanroom availability "in late calendar 2028 and beyond" — while management simultaneously claims "no line of sight to when supply and demand will return to balance" (Micron FQ4 Prepared Remarks). Building greenfield cleanrooms into 2028-29 on a permanent-shortage thesis is the third leg of the everyone-adds-at-once pattern. The market isn't buying the structural story either: at $1,074.89, Micron trades at 14.45x P/E with RSI at 75 — a peak multiple on peak margins (MU market view, Oct 5, 2026). And labor remains a live claimant: Taiwan's strike-rights vote (concludes Oct 6) seeks 15% of operating profit quarterly — ~$6.6B on FQ4 alone (Hundred East).
What to watch: whether Samsung's Oct 8 print (first >₩100T quarterly operating profit expected) and SK Hynix's Oct 27 results beat the raised bar — Q2's miss cost Hynix 10% of its stock.
What surprised us
- The HBM4 "stall" was the buyers' high-water mark. One cycle ago the near-100% hike request looked deadlocked; now Samsung is asking 3x+, SK Hynix is stripping price caps from LTAs, and Micron has already locked most of CY2027 at "significant increases." Buyer pushback peaked early (Xenospectrum, MoveSurge).
- Nvidia de-speccing is the cycle tell hiding in plain sight. The largest HBM buyer cutting 33% of memory from its flagship — while suppliers celebrate 3x pricing — is the demand-elasticity channel that ends shortages, and almost nobody is pricing it (Shattered.io).
- DDR4 over DDR5 is a scarcity tell, not a demand tell. A 10% DDR5 premium for the older generation says the wafer allocation fight, not end demand, is setting prices.
- The take-or-pay question is now the cycle's biggest open bet. Micron's $150B RPO and $12.7B of returnable customer deposits may genuinely dampen the bust — or just relocate default risk to customers, with a refund cliff waiting in the latter half of agreement terms (Micron FQ4 Prepared Remarks).
Open threads worth a vote
The Memory Supercycle — Change Edition
Since last time
Promoted
- Nvidia de-speccing memory. Last time this was a footnote — Vera Rubin BOM share (~29%) and a halved SoCAMM as context for HBM4 resistance. It's now a full core section: Rubin Ultra is being redesigned around an 8-layer/192GB setup, ~33% below Rubin's 288GB.
- Micron's take-or-pay contract book. Previously Micron appeared only via valuation and labor; its SCA/RPO model is now a core section and the cycle's "biggest open bet."
- The HBM spot premium. Previously a passing aside ("a thin squeeze on unallocated scraps"); the same $2,100-vs-$300-400 data point is now a formal cycle-turn tripwire (spot at 4-5x contract).
Escalated
- HBM4 negotiation. The stall you read about last time resolved — in suppliers' favor, hard. Samsung now asking 3x+, SK Hynix stripping price caps from LTAs, TrendForce at +121% HBM ASPs.
- Capex "everyone adds at once." Now on Micron's own paper (~$25B 1H FY27, cleanrooms into late 2028+), not just sell-side forecasts.
- Micron Taiwan labor. From a threatened strike to a strike-rights vote concluding Oct 6, now quantified (~$6.6B/quarter at a 15% claim).
- Consumer demand destruction. From "flexible pricing" anecdotes to PC brands physically cutting SSD capacities from mainstream models.
Demoted
- Nothing was merely reduced — several items dropped out entirely instead (below).
Disappeared
- Citi's 2027 industry capex forecast (+46.5% to $80.4B; Samsung $20.6B / SK Hynix $17.5B / Micron $15.8B DRAM splits).
- The physical supply constraints: ~8% 2027 wafer growth, EUV sold out through 2030, Yongin's June 2029 cleanroom, and KEPCO's grid prepayment demand.
- Citi's bull-case normalization path (DRAM ASP +242.4% in 2026 → +23.1% in 2027).
- Vera Rubin BOM-share/SoCAMM specifics (superseded by the bigger de-speccing story).
- Samsung's LTA price ceilings — ~$10B deposits, caps on ~70% of 2027 server DRAM. Notably ironic given SK Hynix is now removing caps.
- SK Hynix labor resolution (57.08% vote, ₩60.54T quarter, downside-sharing clause) and its 11.4x P/E valuation.
- The "late 2027–2029 collision window" framing.
Unchanged
- The spot HBM3E anchor data point (~$2,100 36GB module vs $300-400 LTA) — same numbers, new role as a tracked indicator.
Pricing: the split market is now official (Escalated)
The deceleration story you already know is now a fully quantified arc, and it has a mirror image. TrendForce's September 30 forecast lays out conventional DRAM: +90-95% (Q1) → +58-63% (Q2) → +13-18% (Q3) → +10-15% (Q4) — the rate of increase collapsed by roughly an order of magnitude even as absolute prices climb (TrendForce, Sept 30, 2026). The buyer pushback has moved from threats to design changes:
"PC brands are also reducing SSD capacities in mainstream models to lower BOM costs, resulting in declines in both procurement volumes and average capacity per device."
— DRAM Contract Pricing Deceleration+1
Micron's print confirms convergence: DRAM ASPs up high-teens QoQ, FQ1 2027 gross margin guided to 86.25% — below FQ4's 87% — which management called the FY2027 floor with "a more moderate rate of price increases" beyond (Micron FQ4 Prepared Remarks). New anomaly: DDR4 8GB now trades above DDR5 8GB ($148 vs $133) — legacy scarcity, not forward demand (Moomoo/Goldman Sachs).
What to watch: TrendForce's January forecast — flat or single-digit Q1 2027 extends the arc; an outright decline is the turn itself.
HBM4: the stall resolved in suppliers' favor (Escalated)
The near-100% hike request you read about as deadlocked was interim noise. Samsung's 2027 ask is now "the mid-to-high $4 range per 1Gb" — more than 3x the ~$1.50/Gb HBM3E mainstream price (Xenospectrum) — and TrendForce projects HBM ASPs up 121% next year, with a 36GB HBM4 stack going from ~$600 to ~$1,300 (Seoul Economic Daily).
"We have completed agreements for the vast majority of our calendar 2027 HBM bit supply with significant price increases year over year, narrowing the gross margin gap with conventional DRAM."
— Micron FQ4 Prepared Remarks
SK Hynix is reportedly going further — removing price caps from its HBM4 LTAs entirely (MoveSurge) — a direct inversion of the Samsung-ceiling story from last time. Lone dissent: Bernstein cut its 2027 HBM hike expectations and trimmed SK Hynix's target from ₩3.3M to ₩2.7M (BigGo). Seed of the next overshoot: HBM's share of global DRAM wafers goes from ~20% to ~30% next year per Samsung's own memory EVP (Seoul Economic Daily) — the capacity valve re-opening at peak pricing.
What to watch: final 2027 settlements — near Samsung's 3x ask or Bernstein's marked-down hike.
Demand: the biggest buyer is rationing itself (Promoted)
The sharpest crack isn't a price decline — it's Nvidia engineering the scarce input out of its flagship. After abandoning 12-Hi HBM4E, the front-runner fallback is the memory-cut option: 8-layer HBM4/192GB, ~33% below Rubin's 288GB (Shattered.io). Meanwhile spot HBM trades at a 4-5x premium to contract — the panic-price signature of buyers who missed allocation.
Micron fielded a direct question about "one large customer de-speccing HBM" on its FQ4 call and deflected to "latent demand," while conceding server content growth is running at "a modestly lower rate... than prior expectations" (Micron FQ4 Prepared Remarks). Suppliers asking 3x+ while their biggest customer cuts 33% of memory from its next flagship are both true in a shortage — but the second one is what ends it.
What to watch: the HBM spot-to-contract premium — a collapse toward 1-2x marks the turn; a blowout means the shortage is deepening.
Micron: the contract book that claims to abolish the cycle (Promoted, with labor escalated inside it)
The take-or-pay model in full: SCAs grew from 16 to 26 in one quarter, ~$150B in remaining performance obligations, $32B in customer financial commitments (mostly cash deposits), agreements into 2031, and 75%+ of 2027 output committed (Micron FQ4 Prepared Remarks). CFO Mark Murphy: "even at floor prices, we expect margins meaningfully above any prior cycle peak margins."
But the capex tell is now on Micron's own paper: ~$25B 1H FY27 capex with 2H higher, skewed to cleanrooms arriving "in late calendar 2028 and beyond" — while management claims "no line of sight to when supply and demand will return to balance." Building greenfield into 2028-29 on a permanent-shortage thesis is the third leg of everyone-adds-at-once. The market isn't buying it: at $1,074.89, Micron trades at 14.45x P/E with RSI at 75 — a peak multiple on peak margins (MU market view, Oct 5, 2026).
Labor, escalated from last time: Taiwan's strike-rights vote (concludes Oct 6) seeks 15% of operating profit quarterly — ~$6.6B on FQ4 alone (Hundred East).
What to watch: Samsung's Oct 8 print (first >₩100T quarterly operating profit expected) and SK Hynix's Oct 27 results — Q2's miss cost Hynix 10% of its stock.
What surprised us
- [UPDATED] The HBM4 "stall" was the buyers' high-water mark. Last cycle's deadlock now reads as peak buyer leverage; Samsung is asking 3x+, SK Hynix is stripping caps, Micron has locked most of CY2027 at "significant increases" (Xenospectrum, MoveSurge).
- [NEW] Nvidia de-speccing is the cycle tell hiding in plain sight. The largest HBM buyer cutting 33% of memory from its flagship while suppliers celebrate 3x pricing — almost nobody is pricing it (Shattered.io).
- [NEW] DDR4 over DDR5 is a scarcity tell, not a demand tell. A 10% premium for the older generation says the wafer allocation fight, not end demand, is setting prices.
- [NEW] The take-or-pay question is now the cycle's biggest open bet. Micron's $150B RPO and $12.7B of returnable deposits may dampen the bust — or relocate default risk to customers, with a refund cliff in the latter half of agreement terms (Micron FQ4 Prepared Remarks).
Open threads
New this cycle:
From last cycle: the hyperscaler budget-ceiling tripwire was not carried forward as a tracked thread. Its substance has been absorbed into the Demand section — buyer pushback is now told through Nvidia's de-speccing and consumer capacity cuts rather than procurement-volume threats. If you were watching for explicit hyperscaler volume-cut announcements, that signal has been overtaken by design-level rationing.