DRAM 2028 Peak Forecast: −23% — Cycle-Timing Buyer Guide

The short version

For two years, every memory headline pointed one way. In the last two weeks, the market started pricing the other direction. Since CXMT's October 10 announcement of a denser 4F² DDR5 server module, memory stocks have sold off sharply: over the five days to October 15, SK Hynix fell 26%, Samsung 22% and Micron about 11% (CNBC). The trigger is not current business — current business is record-breaking — it is a forecast: that this supercycle ends, on a schedule analysts are now willing to date. The emerging forecast timetable: scarcity through 2027, a price turn sometime in 2028, then decline — Counterpoint expects DRAM prices to fall **about 23% in 2028 versus 2027**, with a Maeil Business column citing Korean academics expecting a 'flood of new supply' to turn the market into a 'red ocean' in 2029. Crucially, Counterpoint's −23% still leaves DRAM roughly **230% above 2025 levels**. For buyers, this changes the task. It is no longer only "secure supply in a shortage" — it is "secure supply in a shortage *without signing peak prices into the years after the peak*." Contract structure and expiry timing now matter as much as allocation.

What the market is suddenly pricing

The selloff's anatomy matters because it shows what changed — and what didn't. **What didn't change: the present.** Nanya's Q3 was a record quarter on its own release: revenue up 61.9% quarter-on-quarter, gross margin 85.7%, ASPs up over 60% in one quarter. TrendForce still projects Q4 2026 conventional DRAM contract prices up 10–15% and NAND up 15–20%. Micron has reportedly committed over 75% of its calendar-2027 output. Dell's Michael Dell warns 2027 looks *worse* than 2026. Nothing about the next twelve months got softer. **What changed: the far side of the peak became visible.** CXMT president Cao Kanyu said on October 10 that the company expects to launch a DDR5 RDIMM server module using its 4F² cell design — roughly a third smaller than today's 6F² at the same feature size — around year-end. CXMT has disclosed no specs, volumes, yields, or customers, and Rosenblatt's Kevin Cassidy is blunt: "We don't think this is going to swing the supply-demand equation at all." SemiAnalysis estimates CXMT shipments at about 12% of the DRAM market in 2027 on a data-capacity basis. But the announcement landed on a market already doing the 2028 math: new fab ramps scheduled across 2027–28, SEMCO's substrate expansion planned for September 2028 (#55), CXMT installed capacity forecast at about 500,000 wafers/month by end-2028 (SemiAnalysis estimate, via Seoul Economic Daily — installed capacity, not shipped output) — every supply line this series has tracked converges on the same window. The stocks appear to be pricing the *sum* of that 2028 supply, not CXMT alone.

The forecast timetable, source by source

- **Counterpoint Research:** DRAM prices fall ~23% in 2028. Even after the fall, prices sit ~230% above 2025. Senior analyst Jeongku Choi calls this a "New Normal": because buyers and suppliers now sign 3- to 5-year LTAs with upfront payments and **price collars (upper and lower limits)**, the down-cycle "is not expected to show the same sharp decline seen in the past." - **Korean academia/industry (Maeil, Oct 15):** A Maeil Business column citing SNU's Professor Hwang and KSIA's Ahn Ki-hyun expects a 'flood of new supply' to turn the market into a 'red ocean' in 2029. A new fab takes 3–5 years to build — today's construction starts are 2029's supply. - **The suppliers' own behavior:** Samsung targets 60-70% of its mid-to-long-term planned capacity under long-term agreements if current negotiations complete, and Micron's agreements cover more than 35% of projected revenue through 2030 — which both supports the "shallower decline" thesis and suggests, as our analysis rather than a reported fact, that the *spot* market in 2028 could be thinner and more volatile than the contract average.

What this means for component buyers

**1. Split your book at the peak: volume firm, price floating.** The defining mistake available in 2026 is signing a fixed-price deal that runs past the forecast 2028 peak. The defining skill is copying what the biggest buyers already do (per Counterpoint): multi-year volume commitments **with price collars** — a ceiling that protects you if the shortage worsens, a floor the supplier needs to commit capacity, and indexation or reopeners in between. Lock the allocation; band the price. A supplier who insists on flat peak pricing through 2029 is asking you to underwrite their cycle risk — price that accordingly or shorten the term. **2. Stagger expiries to land after the forecast turn.** Contracts expiring in 2027 reset into a still-rising market (TrendForce: increases moderate, but continue). Contracts expiring in late 2028 reset after the year in which Counterpoint expects prices to fall ~23%. Where you can choose, push renewal dates past the forecast 2028 turn and stagger them — a portfolio of expiries across 2028–29 reprices gradually into the decline instead of all at the 2027 peak or all at the 2029 trough. **3. Treat DRAM and NAND as different cycles.** The forecasts diverge: DRAM stays tight on server demand, while NAND spot has softened earlier — our reading is that the two cycles are diverging, not a reported certainty. On that reading, bundling them into one contract posture risks overpaying on NAND through 2027 and under-securing DRAM. Negotiate them separately, with NAND terms shorter and more spot-exposed than DRAM's. **4. Budget the plateau, not the return to 2025.** A ~23% fall across 2028 is not cheap memory — it is memory at ~3.3× 2025 prices (230% above 2025) under a new cost floor (LTAs with collars, higher wafer costs per #58, substrate costs per #55). Product plans and customer quotes that assume 2025 memory economics return in 2028 will be wrong in the expensive direction. Plan on a high plateau with a softer slope. **5. Track the four triggers that move the date.** The peak date is a forecast, and forecasts move. The variables that pull it forward or push it back: (a) CXMT 4F² yields and actual server-module volumes — today undisclosed; (b) fab ramp timing — Nanya's own release, for example, puts equipment move-in in Q1 2027 with a 30,000 wafer/month target only by end-2028; (c) hyperscaler capex — one scenario to test, not an observed result: flat capex could flip 2027 toward surplus, while strong capex growth keeps it short; (d) the LTA coverage ratio — the more output is collared, the gentler any decline. Review them quarterly; each is observable before it shows in prices. **6. Don't let the selloff talk you out of securing 2027.** Stocks are pricing 2029; your production line lives in 2027, when every source — including the bears — expects shortage, rising contract prices, and >75% of at least one major maker's output already committed. The buyers who lose this cycle will be the ones who read "glut coming" headlines in 2026 and left 2027 uncovered. Timing the cycle means covering the near shortage *and* refusing to pay peak prices beyond it — both, not either. The DRAM and flash lines in our catalog reset against exactly this calendar. Send us your memory requirements by year and we'll map which volumes to collar under LTA, which to leave to the post-peak market, and where your current contract expiries land against the forecast turn.

Frequently asked questions

**Is the memory shortage over?** No. Every near-term indicator points the other way: Q4 2026 contract prices +10–15% (TrendForce), record supplier margins, 2027 output largely committed. What has changed is that analysts now date the *end* of the shortage — tight through 2027 on current forecasts, a price turn expected sometime in 2028, easing thereafter. Those are forecasts, not scheduled facts. **How reliable is Counterpoint's −23% for 2028?** It is a forecast, not a fact — and Counterpoint itself frames the down-cycle as shallower than historical crashes because LTA collars damp it. Direction (a turn later this decade) is broadly consistent with supply models and with the Maeil-cited academic view that new supply floods in by 2029; magnitude and exact quarter are contestable. Plan around the window, not the decimal. **Did CXMT's 4F² announcement really cause the selloff?** It was the trigger, not the whole cause. CXMT's module has no disclosed volumes or yields, and analysts like Rosenblatt's Cassidy dismiss its near-term supply impact. The announcement landed when investors were already modeling 2028–29 capacity from Korea, substrates, and CXMT's expansion combined — the stocks are pricing the sum of 2028 supply, dated by the announcement. **If prices fall in 2028, should I just wait to buy?** Only for volumes you genuinely need in 2028 or later — and even then, "fall" means prices ~23% lower across 2028, at a level still ~230% above 2025. For 2027 requirements, waiting means buying into a rising, largely committed market. Match the purchase to the year of need; don't speculate with production supply. **What's a price collar, and should I ask for one?** A collar sets a maximum and minimum price for contract volumes over the term — you give up some downside if prices crash; the supplier gives up some upside if they spike. Per Counterpoint, 3- to 5-year LTAs with upfront payments and collars are now the standard structure among major buyers. For volumes crossing the 2028 turn, a collar is precisely the tool that converts cycle risk into a budgetable band. **Could the glut arrive earlier than 2028?** The identified early-trigger is demand-side: if hyperscaler capex flattens, surplus conditions could form in 2027. Supply-side surprises (faster CXMT yields, accelerated Korean ramps) are the other path. Both are watchable quarters before they'd hit contract prices — which is why trigger-tracking belongs in the procurement calendar. ---

Sources

Use the latest guidance for the exact part and manufacturer when deciding acceptance criteria.