State of Memory: what the note is saying and where it fits

The note's core argument is that the memory cycle has entered a new phase: price and volume have decoupled, and contract structure — not just the spot price — now determines how earnings react at cycle turns. Longer-term agreements with take-or-pay commitments, price floors/ceilings, and customer prepayments are moving end-of-cycle volatility from suppliers onto customers. But the protection is uneven supplier-by-supplier, so "protected earnings" cannot be assumed as an industry-wide feature. The practical discipline the note keeps returning to: separate the price of available memory from the amount suppliers can actually deliver 114.

The context: why contract terms matter specifically now

This note is the late-cycle update to a thesis Creative Strategies has been building since early 2026. The underlying condition is a physical shortage, not just an HBM premium story. The mechanics:

• HBM is a "capacity tax" on everything else. HBM3e consumes roughly 3× the wafer capacity of standard DDR5 per bit, and HBM4 widens that to ~4:1 (about 1.5× HBM3e per bit) due to more I/O and a larger logic base die. With total DRAM capacity growing only ~8% in 2026, nearly all incremental wafers go to HBM, leaving non-HBM commodity DRAM to grow ~2% 7.
• Tightness spread well beyond HBM, into conventional DRAM, NAND, and older nodes — which is why the earlier framing was "duration, not demand." Micron's April update described customers still receiving only ~50% to two-thirds of medium-term demand 4.
• The market was simply under-modeled. 2024 actual memory revenue of $165B came in 27% above the $130B consensus forecast from late 2023, and Creative Strategies now models $550–570B for 2026 and $800–850B for 2027 — 2.6× and ~4× the old 2024-vintage projections (explicitly a forecast revision, not a realized outcome) 15.

Within that, the contract discussion is the newest variable. Earlier in the cycle the debate was whether demand was real; now the reusable question is how much of that revenue is committed and at what terms 4. The SanDisk work from May flagged the direction this was heading: the old model let customers keep optionality while suppliers absorbed volatility; the new one "moves part of that volatility back to customers through committed demand, financial guarantees, and purchase obligations" 6.

Key implications

1. Revenue growth has stopped being a reliable measure of supply reaching the market. The cleanest illustration is Sandisk's fiscal Q4: revenue up 51% sequentially, with management attributing ~two-thirds of that to pricing rather than volume. That is not a 34% ASP increase — but it shows why strong supplier revenue can coexist with constrained system deployments 114. Price and deliverable bits must be tracked separately.

2. LTAs change how a downturn actually reaches earnings — asymmetrically. On covered business, a falling market price need not produce the same revenue decline (floor + take-or-pay). But in an upcycle, ceilings cap the upside. The note's summary: "how much business has protection and what the terms allow" matters far more than "how many years appear in the agreement" 114.

3. Protection is very uneven — this is the note's most important table:

Source: Creative Strategies Atlas note, Aug 30, 2026; Samsung terms await verification. Evidence gaps ≠ no protection 14.

Three cautions the note attaches to this map: Kioxia's number is a target, not signed coverage; SK hynix's disclosure doesn't quantify pricing protection; and Sandisk itself warns guarantees "may cover only part of the loss if a customer fails to purchase." Financial guarantees are not necessarily cash received, and uncovered volumes remain fully exposed to the cycle 114.

4. There are two distinct constraints, and they solve differently. If HBM capacity reductions bind, memory pooling and efficient model-state placement become more valuable (provided transfer overhead fits the workload). A bandwidth reduction is a separate constraint that adding capacity alone can't fix. The note's nuance for buyers: judge a lower-memory design by "completed work at the required latency," not by its aggregate memory count 1.

5. The buyer side feels it, but not uniformly. Apple is a useful contrast: it accepted a reported 100% price hike from Samsung on LPDDR5X for the iPhone 17 even as DRAM contract prices rose 134% QoQ and NAND 256%, with memory at 10–15% of flagship BOM — but long-term agreements and procurement scale kept it insulated relative to the rest of the market while memory inflation still pressures margins through 2027 109.

The model quantities the price/volume split

The Aug 30 Core Memory V2 working model makes the central point numeric. In the 2027 "High" case, HBM revenue is ~13% higher than Base despite shipping 34% less memory, because realized price is modeled near $54/GB rather than $31:

Working scenarios; calibration and approval pending — not an approved house forecast, and it deliberately does not model individual LTA protections 1214.

Related Creative Strategies / The Diligence Stack research (the memory thread)

In publication order, these build the arc the note sits on:

• WFE Equipment — Why This Time Is Different (Jan 27, 2026) — the HBM 3:1→4:1 trade ratio and the "capacity tax" on commodity DRAM 7.
• Memory's $200B Inflection (Feb 19, 2026) — original framing: HBM share (SK hynix ~50–53%, Micron scaling to ~20–22%), conventional DRAM ASP toward $1.20–1.30/Gb, memory rising to 65–70% of server BOM 3.
• Foundry Economics in the AI Age (Apr 21, 2026) — the $17/GB HBM4 pricing trigger (below which the scarcity premium compresses) 8.
• The Next Debate in Memory Is Duration, Not Demand (early Apr 2026) — the tightness is broad, not an HBM-only story, and "multi-year agreements, tighter fulfillment, assured supply" signal changing buyer behavior 4.
• SanDisk's NBM Moment (May 1, 2026) — the commercial-model shift: volatility moves back to customers via commitments and guarantees 6.
• Memory in the Age of Inference (May 28, 2026) — the forward question: inference turns memory into a live-state capacity problem (KV cache, context, CXL) once pricing cools 11.
• Where AI Constraints Become Pricing Leverage (Jul 2, 2026) — memory framed as the "cleanest pricing-power case," with a ladder grading price-move evidence 5.
• In 2023-2024 We Weren't Bullish Enough (Jul 24, 2026) — the model-revision retrospective 15.
• Custom HBM / zHBM — Co-Design packet (Aug 23, 2026) — related and important: co-design is pulling purchasing decisions earlier in the accelerator cycle, tying commit terms to tape-out and qualification rather than quarterly pricing 12.
• Micron Core Client Intelligence Packet (Aug 29, 2026) — the house's structural-vs-cyclical separation: HBM qualification/yield is differentiated value; conventional DRAM/NAND still cycle; "peak pricing or margin should not be treated as a permanent HBM moat" 13.

Questions worth asking next

These fall out of the note's "What We're Watching" and the surrounding research — separated into coverage, realization, and trigger checks:

Coverage (is the protection real, not just disclosed?)

1. What share of each supplier's next two years of volume is actually signed under enforceable commitments — not targets? The gap between Micron (~20% DRAM signed), Sandisk (~half to two-thirds of bits), and Kioxia's ~50% "target" is the whole game 14.
2. Where do the floors/ceilings actually bind — on which contracts and volumes? Micron's are only in its "largest agreements"; Sandisk's are variable 14.
3. Are the $16.5B/$22B "guarantees" cash received or conditional instruments — and exactly what is covered if a customer defaults? Sandisk itself caveats partial coverage 14.

Realization (is it price, or is it bits?) 4. How much of the latest supplier beat is price vs. volume? Sandisk's ~⅔-pricing attribution is the template — ask it of every memory read 1. 5. Are customers over-buying or actually consuming? Rising customer inventories plus order cuts would signal purchases exceeded consumption and stress-test every LTA 1.

Triggers (what would change the thesis) 6. Is HBM4 realized pricing holding above ~$17/GB, and are packaging yields improving faster than expected? That's the single most important near-term data point for the scarcity premium 8. 7. Is qualified shipment growth accelerating? Faster growth in sellable output would directly weaken scarcity pricing 1. 8. Are lower-memory configurations forcing more remote-memory traffic or more accelerators per workload? That's what would strengthen the pooling/CXL case — and make its economics hinge on the interconnect and software 111.

The meta-question the note wants you to keep in front of all of these: am I looking at the price of available memory, or at the amount suppliers can actually deliver? They're now two different numbers, and the cycle's next turn will be decided by which one you were watching 1.