SASIGNAL ATLASCross-industry intelligence / Research desk
SIGNAL ATLAS / RESEARCH DESK

Memory supercycle: HBM-led DRAM/NAND pricing at unprecedented margins

Scenario set · Semiconductors, advanced hardware & quantum · to 2030

Six alternative branches. These are subjective probabilities from the seed, not observed frequencies or investment recommendations.

Why this trend

The dossier states it plainly: the single most important question in this sector is whether 2026 is the fifth memory cycle or a genuine structural break. Three Tier A disclosures — Micron at 84.9% gross margin, SK hynix at 76% operating margin, Samsung DS at KRW 89.2tn operating profit on KRW 127.5tn revenue — describe margins with no precedent. Everything else in the sector, including consumer device pricing and the AI accelerator cost curve, is downstream of the answer.

Load-bearing assumption

Long-term agreements and prepayments cover enough of general-purpose DRAM output to change the price-discovery mechanism, rather than merely delaying the price signal by two quarters.

Preconditions

Capability

clear — HBM4 is in qualification and HBM absorbs roughly 30% of DRAM wafer capacity

Economics

clear at the top of the cycle — 76-85% margins are disclosed and audited

Supply

constrained by design — capacity additions land 2028-29, which is the whole question

Demand

clear but concentrated — a handful of hyperscalers and two frontier labs drive the marginal order

Permission

partial — Section 232 Phase 2 on equipment remains reserved and unimposed

Capital

clear but expensive — SK hynix raised $26.5bn in US markets, Micron capex ~$27bn, into a hawkish-leaning FOMC

Peak in 2027, cushioned normalisation

base · 40% · band 30-50%

Memory pricing peaks during 2027 and normalises rather than collapsing, cushioned by long-term agreements covering 20-30% of general-purpose DRAM. Margins fall from the high 70s and 80s into the 40s and 50s over 2028-29 rather than to zero. The decelerating element is bit demand growth: hyperscaler capex digestion slows it before capacity additions arrive, so the two do not collide at full force. TSMC holds leading-edge share, CoWoS roughly doubles by 2028, and the binding constraint passes from packaging to power.

Mechanism

Capacity committed in 2026-27 lands in 2028-29 into demand that has decelerated but not reversed. Long-term agreements absorb enough volume that the spot market — historically the price-discovery mechanism that overshoots — is a smaller share of the total, so the descent is shallower than in 2011-12 or 2018-19. Consumer demand destruction, already visible in Samsung's acknowledgement of partial moderation in mobile and PC, removes some of the demand that would otherwise have to be cleared by price.

Preconditions, early indicators and assumptions
Calibration Basis

base_rate

Preconditions
  • Micron quarterly DRAM bit-growth guidance staying in the low-to-mid 20s percent through 2027
  • Contract prices flattening rather than falling through 2027
  • No hyperscaler capex guidance cut
Early Indicators
  • Indicator

    Micron DRAM bit-growth guidance in the low-to-mid 20s percent with contract prices flattening rather than falling

    Source

    S-03-05 Micron investor relations, quarterly

    Would Be Visible By

    2027-12

  • Indicator

    A quarter-on-quarter decline in conventional DRAM contract prices in at least one of the first three quarters of 2027 (forecast F-03-03)

    Source

    S-03-14 TrendForce, quarterly

    Would Be Visible By

    2027-11

  • Indicator

    Korean monthly semiconductor export values still positive year on year through 2027

    Source

    S-03-24 MOTIE, monthly

    Would Be Visible By

    2027-12

Affected Industries
  • 03
  • 01
  • 12
  • 15
  • 10
What Businesses Should Do

Buy memory forward where you can and accept a price above spot for supply certainty through 2027; renegotiate the 2029 tail. If you sell devices, plan the 2028-29 bill of materials on falling memory, not on the 2026 run rate.

Precedence Note

Distinguished from downside by the depth of the margin decline, not by its occurrence: base is normalisation toward 40-50% operating margins; downside is single digits.

Would Change Our Mind

Two consecutive quarters of DRAM contract prices rising more than 10% with bit demand growth above 30%.

Assumptions
  • Text

    Capacity additions and hyperscaler capex digestion do not arrive in the same four quarters.

    Confidence

    medium

    Load Bearing

    true

    Basis

    Dossier 03 §12; SK hynix P&T7, M17 NAND and a new cluster are phased to customer demand

    If Wrong

    They collide, and probability mass moves from base to downside.

  • Text

    HBM continues to absorb roughly 30% of DRAM wafer capacity, keeping the commodity pool short.

    Confidence

    medium

    Load Bearing

    false

    Basis

    T-03-01 evidence, 2026-07-12 estimate

    If Wrong

    The commodity pool loosens earlier and the peak arrives sooner.

The cycle is re-engineered

upside · 12% · band 5-19%

Long-term agreements genuinely change the mechanism. AI-driven bit demand grows faster than capacity through 2028, memory makers sustain operating margins above 50% for multiple years, and capacity is funded without a bust. Equipment billings hold above $170bn. The industry's defining volatility, four boom-bust cycles since 2007, is dampened rather than deferred.

Mechanism

A majority of output is contracted years forward at prices set in advance, so the spot market stops setting the marginal price. Makers phase capacity to contracted demand rather than to price signals, which removes the overshoot mechanism. Customers accept above-spot pricing because supply certainty is worth more to them than price — the same logic that produced take-or-pay contracts in LNG and long-term wafer agreements at TSMC.

Preconditions, early indicators and assumptions
Calibration Basis

analogue

Preconditions
  • A calendar year with no quarter-on-quarter decline in conventional DRAM contract prices
  • Disclosed long-term agreement coverage above half of general-purpose DRAM output
  • Capacity additions phased to contracted rather than forecast demand
Early Indicators
  • Indicator

    No quarter-on-quarter decline in conventional DRAM contract prices across a full calendar year — something the corpus records as not having happened since the 1990s

    Source

    S-03-14 TrendForce, quarterly

    Would Be Visible By

    2028-02

  • Indicator

    SEMI worldwide equipment billings for calendar 2026 above $170bn (forecast F-03-07)

    Source

    S-03-10 SEMI WWSEMS, quarterly

    Would Be Visible By

    2027-04

  • Indicator

    SK hynix or Micron disclosing long-term agreement coverage as a percentage of output

    Source

    S-03-03 SK hynix newsroom and S-03-05 Micron IR, quarterly

    Would Be Visible By

    2027-12

Affected Industries
  • 03
  • 01
  • 12
  • 15
What Businesses Should Do

If you are a large buyer, sign the long-term agreement — this branch is the one in which the contract is worth more than the option. If you are an investor, the tell is disclosed coverage percentages, not margin levels.

Precedence Note

Requires disclosed contract coverage, not merely high margins. High margins alone are the base case at the top of a cycle.

Would Change Our Mind

A single quarter of DRAM contract price decline exceeding 10%, which would show the spot market still sets the marginal price.

Assumptions
  • Text

    Long-term agreement prices are enforceable in practice when spot falls below them.

    Confidence

    low

    Load Bearing

    true

    Basis

    T-03-09 risk register — 'contracted prices becoming above-market and unenforceable in practice'; terms are commercially confidential

    If Wrong

    Contracts are renegotiated as they were in prior cycles, the mechanism is unchanged, and this branch collapses into base or downside.

  • Text

    AI bit demand growth exceeds capacity additions through 2028.

    Confidence

    low

    Load Bearing

    false

    Basis

    Depends on AI model economics that do not yet exist — dossier 03 §11 'nobody can know'

    If Wrong

    The branch fails on demand rather than on contract structure.

The fifth memory cycle

downside · 26% · band 17-35%

2026-27 capacity commitments land into a hyperscaler capex digestion in 2028-29. DRAM prices fall by more than half; Korean operating margins go from 76% to single digits; the firms that raised US equity at the peak carry the impairment. Consumer devices get cheaper, which is the one genuinely good outcome in this branch and the reversal of the entry-tier device problem visible in 2026.

Mechanism

Capacity ordered at the top arrives after the demand that justified it has decelerated. Double-ordering and pull-forward, invisible at the time, unwind — the corpus flags this explicitly as a risk on T-03-01. Long-term agreements are renegotiated rather than enforced, because a supplier that enforces an above-market contract loses the customer's next generation. Korean monthly export values turn negative year on year first, then contract prices, then margins, in the sequence observed in all four prior cycles.

Preconditions, early indicators and assumptions
Calibration Basis

base_rate

Preconditions
  • Capacity additions landing while hyperscaler capex decelerates
  • Long-term agreements renegotiated rather than enforced
  • No offsetting demand source replacing hyperscaler bit demand
Early Indicators
  • Indicator

    Korean monthly semiconductor export values negative year on year for two consecutive months (forecast F-03-01)

    Source

    S-03-24 MOTIE, monthly

    Would Be Visible By

    2028-06

  • Indicator

    Micron GAAP gross margin below 60% in any quarter (forecast F-03-02)

    Source

    S-03-05 Micron IR, quarterly

    Would Be Visible By

    2028-04

  • Indicator

    Micron quarterly DRAM bit-growth guidance falling below 15% for two consecutive quarters

    Source

    S-03-05 Micron IR, quarterly

    Would Be Visible By

    2028-03

Affected Industries
  • 03
  • 01
  • 12
  • 15
  • 10
What Businesses Should Do

Do not sign a five-year fixed-price memory agreement at 2026 prices without a market-reset clause. Device makers should plan for a 2028-29 bill-of-materials tailwind and resist locking it away in long contracts now.

Precedence Note

Distinguished from failure by where the cancellation starts: downside starts in memory pricing; failure starts at the equipment layer and works inward.

Would Change Our Mind

Korean export values staying positive year on year through 2028 while capacity additions land on schedule.

Assumptions
  • Text

    The mechanism that produced four boom-bust cycles since 2007 still operates, because contract coverage is 20-30% rather than a majority.

    Confidence

    medium

    Load Bearing

    true

    Basis

    T-03-09 estimate, 2026-07-12; the coverage figure is a Korean press estimate, not a disclosure

    If Wrong

    The mechanism is genuinely changed and mass moves to upside and base.

  • Text

    Current bit demand contains a pull-forward and double-ordering component.

    Confidence

    medium

    Load Bearing

    false

    Basis

    T-03-01 risk register

    If Wrong

    Demand is cleaner than assumed and the downcycle is shallower.

The package eats the chip

disruption · 7% · band 2-12%

Panel-level CoPoS and chiplet standardisation arrive faster than expected, commoditising interposers and letting second-tier designers assemble competitive parts from purchased chiplets. Leading-edge wafer share becomes less decisive than packaging and integration skill; optical I/O moves on-package and the fabric becomes the differentiator. Memory's position changes from scarce commodity to co-packaged component sold into a more competitive assembly layer.

Mechanism

CoPoS reaches mass production on schedule in late 2027 with a non-TSMC supplier qualified, which breaks the single-site concentration that currently prices advanced packaging. Chiplet interface standardisation lets designers buy compute, memory and I/O dies from different suppliers. HBM's pricing power depends on being one of two or three qualified suppliers into a constrained package; a wider packaging base weakens that.

Preconditions, early indicators and assumptions
Calibration Basis

analogue

Preconditions
  • CoPoS reaching mass production with a second qualified supplier
  • At least two independent parties shipping standardised chiplet-based accelerators
  • Advanced packaging capacity ceasing to be the binding constraint
Early Indicators
  • Indicator

    CoPoS panel-level packaging in mass production in late 2027 with a non-TSMC supplier qualified

    Source

    S-03-14 TrendForce and S-03-15 DigiTimes, daily

    Would Be Visible By

    2028-03

  • Indicator

    Broadcom AI semiconductor revenue at or above $30bn in a quarter, indicating the fabric and custom layer is capturing value (forecast F-03-05)

    Source

    S-03-08 Broadcom IR, quarterly

    Would Be Visible By

    2027-12

  • Indicator

    Amkor and SPIL monthly revenue growing faster than TSMC's for two consecutive quarters

    Source

    S-03-23 TWSE MOPS monthly filings, monthly

    Would Be Visible By

    2027-12

Affected Industries
  • 03
  • 01
  • 15
What Businesses Should Do

If you design accelerators, qualify a second packaging path now — the corpus's own rule is that one company is a project and two is a sector. If you buy them, ask which package your supply depends on, not which node.

Precedence Note

Requires a second qualified supplier. A single supplier reaching CoPoS production leaves the outcome in base.

Would Change Our Mind

CoPoS slipping more than two quarters with no second supplier qualified.

Assumptions
  • Text

    Packaging capacity, not wafer capacity, is what currently prices AI accelerators, so widening it changes the cost curve.

    Confidence

    medium

    Load Bearing

    true

    Basis

    T-03-02; CoWoS capacity is unpublished and every circulating figure is a channel estimate (see forecast F-03-08)

    If Wrong

    Widening packaging does not change pricing and the branch has no economic content.

  • Text

    Chiplet interface standardisation reaches the point where dies from different suppliers interoperate commercially.

    Confidence

    low

    Load Bearing

    false

    Basis

    T-03-11; the corpus's 450mm analogue is a decade-long industry format transition that was abandoned

    If Wrong

    Assembly stays vertically integrated and the disruption is deferred.

Section 232 Phase 2 lands on equipment

regulatory · 8% · band 2-14%

The January 2026 proclamation explicitly reserved expansion to additional semiconductors and to manufacturing equipment. Phase 2 is imposed. US fab construction costs rise materially, equipment makers face a tariff on their own inputs, and retaliation touches materials, where Japan and China both have leverage. The direction is restricting; a liberalising variant — Phase 2 abandoned and the existing bands narrowed — is the same branch run the other way.

Mechanism

New HTS subheadings under 9903.79 extend coverage to manufacturing equipment, or CBP issues CSMS guidance broadening the existing headings. Tool costs rise for every fab under construction in the US, which raises the capital cost of exactly the capacity the CHIPS programme was built to create. Japan and China respond in materials — photoresist, mask blanks, CMP slurry, wafer substrates — where a handful of firms hold near-monopolies that are invisible in English-language coverage until they bind.

Preconditions, early indicators and assumptions
Calibration Basis

base_rate

Preconditions
  • Phase 2 negotiation status resolving rather than staying open
  • New HTS subheadings or CBP guidance extending coverage
  • No court challenge that suspends the sectoral authority
Early Indicators
  • Indicator

    New HTS subheadings under 9903.79 covering semiconductor manufacturing equipment, or a CBP CSMS bulletin extending coverage

    Source

    S-03-13 USITC HTS and CBP CSMS, irregular

    Would Be Visible By

    2027-12

  • Indicator

    BIS Federal Register notices extending the advanced-computing control perimeter

    Source

    S-03-12 BIS, irregular

    Would Be Visible By

    2027-12

  • Indicator

    ASML or Applied Materials quantifying a tariff impact in quarterly guidance

    Source

    S-03-09 ASML IR, quarterly

    Would Be Visible By

    2027-10

Affected Industries
  • 03
  • 01
  • 09
  • 05
What Businesses Should Do

Map your tool and materials exposure to HTS headings now, not after the proclamation. The corpus's finding that the existing 25% tariff is functionally a toll on the export-control compliance pathway rather than a broad import tariff means the headline rate tells you nothing about your own exposure.

Precedence Note

Regulatory outranks base when a proclamation or CBP bulletin issues, even if volumes look base-like. Negotiation status alone stays in base.

Would Change Our Mind

The Phase 2 reservation formally lapsing, or a court vacating the sectoral authority as the Supreme Court did for the IEEPA tariffs.

Assumptions
  • Text

    The Section 232 sectoral authority survives judicial challenge, unlike the IEEPA tariffs the Supreme Court voided.

    Confidence

    medium

    Load Bearing

    true

    Basis

    Macro brief addendum — the tariff architecture was rebuilt on sectoral authorities precisely because they are more robust

    If Wrong

    The instrument disappears and the branch's mass returns to base.

  • Text

    Retaliation, if it comes, lands in Japanese and Chinese materials rather than in finished chips.

    Confidence

    low

    Load Bearing

    false

    Basis

    Dossier 03 §10 — the Japanese materials layer is a set of near-monopolies essentially uncovered in English

    If Wrong

    The transmission path differs and the cost lands elsewhere in the chain.

Cancellation from the tools inward

failure · 7% · band 2-12%

The FOMC's flagged financial-stability risk from high AI-firm valuations and leveraged infrastructure financing materialises. Orders are cancelled from the equipment layer inward, in the reverse order they were placed: tools first, then packaging, then wafers, then memory. Because equipment leads chip revenue by four to six quarters, this sector sees it before the AI sector does — and the firms that raised equity at the peak carry capacity that never runs.

Mechanism

A financing event in sector 01 removes the marginal buyer of compute. Equipment orders, which are placed furthest ahead, are the first thing cancellable without breaching a delivered-goods contract. SEMI billings fall sequentially while TSMC monthly revenue is still growing — the divergence is the tell, because it separates a demand break from a normal cycle. Memory makers with capacity mid-construction face the worst version of the classic error: capital raised at the top funding capacity that arrives into a trough.

Preconditions, early indicators and assumptions
Calibration Basis

base_rate

Preconditions
  • A visible AI capex retrenchment at the hyperscaler or neocloud layer
  • Equipment order cancellations rather than deferrals
  • No offsetting demand from non-AI end markets
Early Indicators
  • Indicator

    SEMI quarterly equipment billings declining more than 10% sequentially while TSMC monthly revenue is still growing — the divergence is the tell

    Source

    S-03-10 SEMI WWSEMS quarterly and S-03-01 TSMC monthly

    Would Be Visible By

    2028-06

  • Indicator

    ASML cutting full-year guidance by more than 10%

    Source

    S-03-09 ASML IR, quarterly

    Would Be Visible By

    2028-04

  • Indicator

    A memory maker announcing a capex cut of more than 20% or an impairment

    Source

    S-03-03, S-03-04, S-03-05 company IR, quarterly

    Would Be Visible By

    2028-06

Affected Industries
  • 03
  • 01
  • 05
  • 09
  • 11
  • 12
What Businesses Should Do

Watch the equipment layer, not the chip layer — it turns four to six quarters earlier and the data is quarterly and free. Anyone whose plan depends on 2029 memory or packaging capacity should know which supplier's balance sheet it sits on.

Precedence Note

Failure outranks all other branches on occurrence. Downside becomes failure at the point where equipment orders are cancelled rather than pushed out.

Would Change Our Mind

Equipment billings declining while hyperscaler capex guidance is still rising, which would indicate a normal digestion rather than a break.

Assumptions
  • Text

    Equipment billings still lead chip revenue by four to six quarters, as they have historically.

    Confidence

    medium

    Load Bearing

    true

    Basis

    Dossier 03 §7 'Cycles in equipment'

    If Wrong

    The lead indicator fails and the branch is undetectable in advance, which is the worst case for anyone relying on it.

  • Text

    SEMI's reported billings are comparable year to year despite the regional breakdown being subscription-only.

    Confidence

    medium

    Load Bearing

    false

    Basis

    Dossier 03 §13 item 3 — the regional table requires a paid subscription

    If Wrong

    The signal is noisier than assumed and the divergence test is harder to run.

Additional scenario notes

Probabilities Sum

1

What Must Be True To Grow
  • AI bit demand growth exceeds capacity additions through 2028
  • Long-term agreements are enforceable at contracted prices when spot falls below them
  • HBM continues to absorb enough DRAM wafer capacity to keep the commodity pool short
  • Advanced packaging expands fast enough not to cap accelerator volumes, and power expands fast enough not to cap deployment
What Could Stop It
  • Physical constraint — advanced packaging and, beyond it, the power and thermal envelope now capping deployable compute
  • Demand deflation — pull-forward and double-ordering unwinding, which the trend record names as a live risk
  • Capital cost — capacity committed at the cycle peak into a hawkish-leaning FOMC
  • Regulatory reversal — Section 232 Phase 2 on equipment, explicitly reserved and not yet imposed
  • Input chokepoint — Japanese photoresist, mask blanks, CMP slurry and wafer substrates are near-monopolies with no English-language coverage
Uncertain Assumptions
  • That long-term agreement coverage is 20-30% of general-purpose DRAM — this is a Korean press estimate, not a disclosure, and the terms are confidential
  • That TSMC's CoWoS capacity is near 130k wafers/month — every circulating figure is a channel estimate and they differ by roughly 8% (see forecast F-03-08)
  • That SEMI's own 2026 forecast of $145bn and its quarterly actuals are on the same basis
  • That CXMT's HBM yield and output are immaterial — CXMT is unlisted and discloses essentially nothing
Authored

2026-09-15

More from this sector