06-reference/research

hbm fixed price offtake primary filings

2026-11-01·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
memory-cyclehbmprimary-sourcesinvesting-thesiscontract-structure

Five-year take-or-pay memory agreements exist, pricing is dominantly a floor-and-ceiling band, a fixed-price subset exists, and no filing ties a fixed-price agreement to HBM or a hyperscaler

The question

Verbatim: "Does a 3+ year fixed-price hyperscaler HBM offtake agreement actually exist as of September 2026? Resolve from primary sources (10-Q/10-K purchase-commitment disclosures, earnings-call transcripts), not trade press."

Context: this is open follow-up #2 of [[2026-09-09-hbm-plateau-hypothesis-backtest-variant]], which named it as the single fact that most moves the plateau prior. The vault carried an unreconciled three-way split: [[2026-05-18-hbm3e-hbm4-capacity-timeline-phase-b-end]] said a 3+ year fixed-price hyperscaler offtake "has not happened," [[2026-05-18-memory-cycle-v1.1-v2-honest-rerun]] cited "multi-year offtakes" as plateau evidence, and trade press claimed five-year long-term agreements (LTAs). This brief resolves what filings and company-published management remarks actually say, and marks precisely where they say nothing.

Date-framing note. The question was filed 2026-09-09 and framed "as of September 2026." The disclosure that resolves the structure and term conjuncts predates it: Micron's FQ3 FY2026 10-Q (filed 2026-06-25) and FQ3 prepared remarks (2026-06-24). Three sources postdate the framing and are used as confirmation and extension: Micron's FQ4 prepared remarks (2026-09-30), Micron's FY2026 10-K (filed 2026-10-09), and SK hynix's Q3 2026 results release (2026-10-28). So the vault's September contradiction was resolvable at the time on structure and term; nobody had read the 10-Q. The counterparty conjunct was not resolvable in September: the first category-level naming of cloud service providers anywhere in the primary record is SK hynix on 2026-10-28, and SK hynix's prior release (2026-07-29) said only "around 10 customers, including key strategic partners," with no five-year term.

What we already know (from the vault)

What the web says

Primary sources only. The Micron 10-K quotes, the Micron 60%-concentration figure and the Nvidia Note 10 quote were each re-verified directly against the raw filings by the synthesising agent rather than accepted on an extraction report.

Convergences and contradictions

Synthesis for RDCO

Verdict on each conjunct, separately. (a) Does a long-term agreement exist? Yes, confirmed, high confidence. Micron's FY2026 10-K discloses take-or-pay agreements with "binding commitments for specific contractually enforceable volumes," backed by $12.75B of customer cash deposits received in FY2026 and $7B of third-party letters of credit, with $134B of RPO where a year earlier RPO was "not material." That is an audited balance sheet with a new line item, not a press release. SK hynix and Samsung independently confirm multi-year memory agreements in their own primary disclosures. (b) Is the term 3+ years? Yes, confirmed, high confidence. Micron management: "typically ... a five-year term, from calendar 2026 through the end of calendar 2030," automotive at three years, some extending into 2031. SK hynix: "typically five years." The 10-K corroborates independently, since RPO is "expected to be fully recognized within the next five years" and deposits are "scheduled to be repaid between 2029 and 2031." (c) Is the price fixed? Split answer. (c1) The dominant structure is a band, not a fixed price: high confidence. The majority of Micron's agreements price within minimum and maximum bands, and SK hynix prices within an agreed range that tracks the market. (c2) A fixed-price subset is disclosed to exist, and is not sizeable from any primary source: existence moderate-to-high confidence, magnitude unknown. The confidence is not higher because every source sentence names fixed pricing only inside an "either fixed or banded/negotiated" disjunction rather than describing a fixed-price agreement on its own; the strongest single wording is the 10-K's RPO exclusion sentence, which treats "fixed pricing" as its own category. Micron's filing says pricing is "either fixed or periodically negotiated"; the FQ3 10-Q says "Pricing for most agreements is either fixed, or is subject to minimum and maximum pricing"; FQ4 says RPO covers agreements whose pricing "can be either a fixed price or subject to a pricing floor and ceiling." The only bound available is FQ3's "a modest portion of the SCA-related revenue," and even that merges fixed-price agreements with no-band agreements. So the full question, read as a conjunction, is not confirmable from primary sources and is also not refuted. No source ties a fixed-price agreement to HBM specifically, or to a named or category-identified hyperscaler counterparty, and two further conjuncts the question assumed are themselves unresolved: whether and how much HBM volume sits inside any strategic agreement is not disclosed, and the calendar-2027 HBM book's term and pricing form are not disclosed either. "The filings do not disclose this" is the honest answer to the fixed-price-HBM-hyperscaler conjunction, and it is not the same as "it does not exist."

What this does to the plateau prior: it moves it up, modestly, for a mechanism the plateau thesis never named, and it truncates the thesis's upside at the same time. The plateau case as carried through four vault documents rested on demand being contractually locked at a fixed price. It is not locked that way. What exists instead is better-targeted than the thesis asked for and smaller in scope: a price floor on a growing share of revenue, with real enforcement behind it. Micron's statement that "We expect gross margins from our strategic customer agreements with price bands, even at floor pricing levels, to yield gross margins meaningfully above our peak quarterly margins in any past cycle" (FY2026 10-K, Item 7 MD&A; the FQ4 remarks carry a similar sentence in Murphy's voice) is the load-bearing economic claim in this brief, because the historical memory-cycle crash is an ASP and margin event and a contractual floor is precisely a left-tail truncation on ASP. It is also a management expectation, not an audited outcome, and it has never been tested through a demand deceleration. Deposits and letters of credit make the floor collectible rather than aspirational. But the same instrument caps the right tail: ceilings anchored at calendar-Q2-2026 market prices mean the covered share cannot participate in further ASP escalation. The honest characterisation is therefore variance reduction, not direction, on a bounded and disclosed share: over 35% of Micron revenue through 2030, over 50% of SK hynix 2027-28 capacity. The next downturn looks shallower and later for the covered portion and fully spot-exposed for the rest, which is a materially different claim from "the cycle is broken." I would not attach a numeric probability shift to this. The defensible statement is that the plateau prior moves from "resting on an unsupported claim" to "resting on a partially supported, bounded and as-yet-untested mechanism," while the uncovered residual still carries the full historical cycle risk. The 2029-2031 deposit-repayment window is now a datable cash event, and it coincides with the subsidy-insulated capacity tail that [[2026-10-06-memory-fab-subsidy-share-2027-28-wave]] re-dated to 2029-31.

The tripwire consequence is a measurement problem and it needs an amendment. Tripwire #1 tracks DRAM contract-ASP deceleration off third-party series, and [[2026-10-03-dram-asp-series-basket-consistency-tripwire-1]] did the work of pinning it to one consistent conventional-DRAM basket. Band pricing breaks that series' relationship to realized revenue going forward: as agreement coverage grows, the published contract-price series increasingly measures the uncovered residual rather than blended realized ASP, and it will overstate both the rise and the eventual fall. The TrendForce headline the vault already had in hand, "Long-Term Agreements Cap Price Increases," was reporting exactly this and the vault read past it. The amendment is not to retire Tripwire #1 but to re-key it: a residual contract-price decline is now necessary but no longer sufficient for the ASP leg of a Phase-3 call, and the sufficiency condition should be a decline in blended realized ASP, which is computable from Micron's own revenue and bit-shipment disclosures. No other v1.1 tripwire changes, since inventory days, capex composition, greenfield online dates and bit share are all physical and indifferent to contract form. The 05-18 bull-capitulation test should be amended rather than deleted, and the candidate replacement is a three-part condition: banded-floor coverage above a named share of maker revenue, and collectible security behind it (deposits plus letters of credit, as Micron now discloses and SK hynix does not), and disclosure tying HBM volume specifically into those agreements. Noting honestly that on this brief's own reasoning a banded floor with a ceiling may not deserve to satisfy a bull capitulation test at all, since it caps upside as well as downside; that is a judgment for the quarterly review, not something this brief should settle.

The methodological finding is that contract architecture is now a disclosed, machine-readable quarterly series, and the vault should treat it as a new anchor rather than an anecdote. Four of these did not exist a year ago and are now recurring line items: Micron's RPO, its noncurrent customer contract liabilities, its agreement count and revenue-coverage percentage, and its letters-of-credit total. The natural cross-maker test is whether SK hynix's LTAs carry the enforceability Micron's do, but the timing of that test needs care: the RPO disclosure requirement in IFRS 15 is an annual financial-statement requirement, and quarterly reporting under IAS 34 and K-IFRS 1034 generally does not compel it. So absence of an RPO disclosure in SK hynix's 2026-11-14 quarterly business report would be close to uninformative, and the real tests are its FY2026 annual business report (approximately March 2027) and its first Form 20-F (approximately April 2027). What the November quarterly can show is the deposit and contract-liability balances, which speak to customer financial security rather than to price terms. One standing caution on the maker comparison: Micron is a US registrant under full SEC disclosure obligation while Samsung is not and SK hynix only became one in July 2026, so part of the disclosure gap is regulatory rather than commercial, and the 2027 annual filings are what separate the two.

Why this is in the vault

It closes open follow-up #2 of [[2026-09-09-hbm-plateau-hypothesis-backtest-variant]], amends the "no 3+ year fixed-price offtake" capitulation test in [[2026-05-18-hbm3e-hbm4-capacity-timeline-phase-b-end]] as mis-specified rather than merely stale, and supplies the Tripwire #1 re-keying that the memory-cycle v1.1 anchor set needs before the next quarterly review can read a contract-price series as a Phase-3 input.

Open follow-ups

  1. Compute Micron's blended realized DRAM ASP from its own disclosures and compare it to the TrendForce conventional-DRAM contract series over the same quarters. This is the test of how far the published series has already diverged from realized pricing, and the quantitative half of the Tripwire #1 re-keying. Micron reports revenue and bit-shipment growth by technology quarterly, so the series is constructible from primary sources.
  2. Read SK hynix's FY2026 annual business report (approximately March 2027) and first Form 20-F (approximately April 2027) for a remaining-performance-obligation disclosure and any non-current customer contract liability. Annual filings are where IFRS 15's RPO requirement actually binds, so this is the first clean test of whether the five-year LTAs have banded pricing in the accounting sense. The 2026-11-14 quarterly report is worth reading for deposit balances only.
  3. Establish whether the fixed-price-only share of Micron's agreements is ever separately disclosed. Management has only ever reported fixed and ceilinged together. If a future filing or remark splits them, conjunct (c2) becomes answerable with a magnitude instead of an existence statement.

Related

Sources

Research notes trail: every source consulted, with reachability

Micron Technology (CIK 0000723125), SEC filings via EDGAR, curl with declared user agent, all HTTP 200

Micron company-published management remarks (investor-relations CDN, curl, HTTP 200)

SK hynix (CIK 0002120882, Nasdaq: SKHY), SEC-registered since July 2026

Samsung Electronics (CIK 0000879316 exists but has filed no periodic report since 2015; a DART-only issuer)

Nvidia (CIK 0001045810), the intermediary test, all HTTP 200

Hyperscaler demand-side test, 14 filings, all HTTP 200, all zero-hit on memory commitments

Microsoft (CIK 0000789019): 10-K 2026-07-29, 10-Q 2026-10-28. Alphabet (CIK 0001652044): 10-K 2026-02-05; 10-Qs 2026-04-30, 2026-07-23, 2026-10-29 (the Q3 filing carries the Note 10 banded-pricing language quoted above: https://www.sec.gov/Archives/edgar/data/1652044/000165204426000084/goog-20260930.htm). Amazon (CIK 0001018724): 10-K 2026-02-06; 10-Qs 2026-04-30, 2026-07-31, 2026-10-30; plus the FY2024 10-K and Q3 2025 10-Q as baselines establishing that the "memory chips" phrase is new. Meta (CIK 0001326801): 10-K 2026-01-29; 10-Qs 2026-04-30, 2026-07-30, 2026-10-22. Amazon's affirmative disclaimer, 10-K Item 1A: "We do not have long-term arrangements with most of our suppliers to guarantee availability of merchandise, content, components, or services."

Not reachable

Vault