06-reference/research

hbm dram phase 2 3 transition q3 2026

2026-09-25·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
memory-cyclephase-markersdramhbminvesting-thesismarkov-pipeline

Q3 2026 indicator review: four of five tripwires are colder than they were in July, and the one that mattered most is unreadable until Sept 30

The question

"As of Q3 2026, how have the Phase 2→3 transition indicators (HBM spot pricing, Samsung/Micron capex guidance, DRAM contract prices, fab utilization) moved, and does current data support beginning to size into Phase 3 rotation positions?"

Context: the question was written in July 2026. This brief answers it as of 2026-09-25, five days before Micron's FQ4 FY2026 print. Every reading below carries its own as-of date, because the honest headline of this update is that several indicators have no Q3 reading at all and the July values are the newest verified numbers in the vault.

What we already know (from the vault)

What the web says

Indicator scorecard — each against its own previously-stated threshold

# Indicator Threshold (as set 2026-06-28 / 2026-07-07) Newest verified reading + as-of date Move since 2026-06-28 Verdict
1 ASP rate-of-change (2nd derivative) QoQ contract-price gain decelerates 2+ consecutive quarters 2Q26 +58-63% → 3Q26 +13-18% (server DRAM), as of 2026-07-09 No new reading in 11 weeks. 4Q26 figure paywalled STILL FLICKER (1 of 2). Cannot advance or clear. Qualitative reports of a lifted 4Q26 outlook would, if true, reset the flicker
2 Supplier inventory days Supplier DRAM inventory >5-6 weeks for 2+ consecutive quarters (off a ~3-4wk floor) Supplier ~3-5 weeks, as of Jul 2026 (third-party). Micron total DIO 120.5 days, FG days 8.7, as of 2026-05-28 (primary) DIO still falling (123.1→121.9→120.5); FG days at series low NOT TRIGGERED. Moving the wrong way for Phase 3. Next read Sept 30
3 Capex composition → greenfield bits online 2+ vendors bring greenfield wafer capacity ONLINE (not node/packaging), OR Samsung hits ~250k HBM wpm on/ahead of schedule Zero greenfield DRAM fabs online, as of 2026-09-15 register. First step: Yongin Y1 tools Q2 2027, Micron ID1 mid-CY2027. Samsung 250k status unverified Announcements up (Korea $575B, Jul 2026); online dates unchanged or later NOT TRIGGERED. The clock did not move this quarter
4 ASP / contract-price rollover (the confirm) DRAM contract/spot declining 4-6 consecutive months, OR any maker reports HBM ASP -10% QoQ Prices still rising through 3Q26, as of 2026-07-09. Zero months of decline recorded No progress toward the threshold NOT TRIGGERED, and furthest from firing. Even a decline starting today puts confirmation in Q1-Q2 2027
5 Exogenous entrant (CXMT / YMTC) 5a DRAM: CXMT bit share ≥~15-20% while DRAM bit-demand growth ≤~18% (re-derived 2026-09-08; supersedes the old 12-15% level). 5b NAND: YMTC share + supply contribution 5a: CXMT ~12% forecast for 2027 — a 2028-29 event. 5b: YMTC 14% of global NAND bits, Q2 2026 5b already crossed on NAND; 5a demonstrably unreachable against strong demand 5a NOT TRIGGERED. 5b TRIGGERED — but on NAND, which the DRAM/HBM thesis does not rest on
6a Fab utilization (confirming) Aggregate memory fab utilization <90% with new capacity online Last known sustained 95%+, as of Jun 2026. No Q3 2026 reading verified this run Unverified UNVERIFIED — not scoreable tonight
6b Gross-margin peak (confirming) Margin peaks, then compresses 2-3 quarters into the downturn Micron GAAP GM 84.6% (period ended 2026-05-28, primary); FQ4 guided ~86% Still climbing NOT TRIGGERED. No peak has printed
6c Lead times / allocation (confirming) Lead times compress <~26 weeks AND allocation frameworks relax >40 weeks, allocation-only, as of Jun 2026. No Q3 reading verified Unverified UNVERIFIED — not scoreable tonight

Composite rule status: P3 requires (#2 AND #4) or (#3 AND #4). #4 has zero progress on either leg. The composite is therefore arithmetically unable to fire today, and — because #4 needs 4-6 consecutive months of decline — it cannot fire before roughly Q1-Q2 2027 even if a rollover began this month.

Instrumentation note the founder should see: the anchors/dram-spot/ series' newest row is dated 2026-04-15. It has not been fed in over five months. The vault's structured price anchor is stale, which is why indicator #1 had to be scored off a July press release rather than a series.

Convergences and contradictions

Synthesis for RDCO

The direct answer: no. Current data does not support beginning to size into Phase 3 rotation positions. Scored mechanically against thresholds the vault set for itself in June and July, the Q3 2026 update is: zero of five primary triggers fired on DRAM/HBM; one fired on NAND (YMTC, 5b); the single flicker from July is stalled and possibly reset; two confirming features are unverifiable tonight. The composite rule cannot fire without #4, and #4 requires four to six consecutive months of price decline that have not begun. That is not a close call, and the calibration discipline here runs one way: the evidence I could verify says later, and the evidence I could not verify is mostly the bullish-for-Phase-2 kind.

Confidence. High (call it 85%) on the narrow mechanical claim — no defined Phase-2→3 trigger has fired on DRAM/HBM as of 2026-09-25, because that is a threshold-matching exercise against readings I sourced this run. Moderate (call it 60%) on the broader claim that the turn is still three or more quarters out, because that claim leans on indicators with no Q3 reading — supplier weeks-of-supply, fab utilization, lead times, and the 4Q26 contract print — plus the 2018 announce→rollover lag, which the vault itself labels a floor. I am deliberately not converting "no trigger fired" into "the top is far away." Those are different statements and the second one is weaker.

The question contains a premise that does not hold, and this is the more useful finding. "Sizing into Phase 3 rotation positions" presupposes a defined Phase-3 basket. A full-vault search finds none — no Phase-3 rotation list, basket, or candidate set exists anywhere in ~/rdco-vault. Separately, memory-cycle-v1.1 was never deployed as a strategy; it remains pending a founder deploy gate. The memory names that actually sit in the paper portfolio — MU (via nvidia-supply-chain-v1), SNDK and INTC (via smart-money-mirror-v1) — were bought for other strategies' reasons. So "rotating the paper portfolio toward Phase 3" would today mean trimming positions held on non-memory logic, against a target basket that has never been specified. Even if a trigger had fired, the rotation is not executable as posed. The real gap is not the phase call; it is that the Phase-3 side of the trade was never built.

What to do instead, ordered by cost. First, wait five days. Micron's FQ4 FY2026 print on 2026-09-30 resolves more of this scoreboard than any research run can: DIO and finished-goods days (#2), FY2027 capex guidance and its composition (#3), whether ~86% gross margin actually printed or the peak arrived early (#6b), and — per the falsifiability table in [[2026-09-08-cxmt-ymtc-bit-share-phase3-threshold]] — whether a big-three capex cut appears. A brief written the week of Oct 1 would be worth several written tonight. Second, re-feed anchors/dram-spot/; a five-month-stale price anchor is why the highest-frequency indicator had to be scored off a press release. Third, if the founder wants the rotation to be a live option rather than a phrase, specify the Phase-3 basket now, while nothing is urgent — what the trade actually is (memory-consumer margin relief, equipment de-rating, pure shorts, or simply exiting memory to cash) is an unanswered design question, and designing it under a firing trigger is how overconfident sizing happens.

Illustrative only, not a recommendation, and not a buy list. If a Phase-3 basket were specified, the phase logic points at memory buyers whose input costs fall (server OEMs, module makers, handset and PC assemblers) and away from the producer and WFE complex that leads the cycle down — the names currently in the memory bucket (MU, SNDK, INTC, SMH) are the Phase-2 expression, which is the side being rotated out of. Nothing here is sized, screened, backtested, or gated. It is included to show what the rotation would have to be about, not which tickers to buy.

Falsification — what would change this call. (1) A verified 4Q26 DRAM contract-price forecast showing deceleration below the 3Q26 +13-18% band turns #1 from a flicker into a two-quarter trigger and moves the prior materially. (2) Micron FQ4 DIO rising with finished-goods days off the 8.7 floor starts #2's clock. (3) A guided big-three capex cut, or FY2027 capex composition shifting to greenfield wafers, flips #3's meaning. (4) Any single month of NAND contract prices negative while DRAM stays positive is the cheapest live test of the YMTC decoupling. (5) Conversely, a 4Q26 print at or above 3Q26's gain, plus Micron FQ4 gross margin printing ≥86%, would confirm the flicker has reset and push the earliest plausible turn window past H1 2027 — at which point the "HBM broke the cycle" plateau hypothesis deserves promotion from caveat to working assumption.

Why this is in the vault

This is the Q3 2026 scoreboard update for the memory-cycle-v1.1 thesis and for Layer 1 of the Markov pipeline's phase tracker — the structured review the backlog flagged as overdue since 2026-06-28. It directly answers whether the paper portfolio rotates this quarter (it does not) and records the two blockers that decide the next review: the paywalled 4Q26 contract print and the Sept 30 Micron FQ4 release.

Open follow-ups

Related

Sources

Vault

Web (all retrieved 2026-09-25)

Correction 2026-10-03. The "server DRAM" label on the 3Q26 +13-18% reading above is wrong. The figure originates in TrendForce's 2026-07-03 release, which reads "Conventional DRAM contract prices are forecast to rise 13-18% QoQ in 3Q26" - verified against the primary source on 2026-10-03, not taken on a secondary's word. A separate TrendForce release six days later (2026-07-09) does quote server DRAM at a coincidentally identical 13-18% range, and this doc cited that headline while disclosing in the same sentence that "the body was not fetched." That is how the two baskets got crossed. The 2Q26 +58-63% and 4Q26 +10-15% readings are also conventional DRAM at source, so all three tripwire readings are one consistent basket and there was no basket switch. Full chain: [[2026-10-03-dram-asp-series-basket-consistency-tripwire-1]].