MTS

Wiki

AI Memory & the DRAM Cycle — Research Wiki

A working corpus on the supply-demand mismatch in DRAM and HBM during the 2024–2026 AI buildout. Built from primary sources (10-Ks, 10-Qs, earnings call prepared remarks, DOJ press releases, DART filings, court documents, vendor spec sheets) rather than commentary.

The wiki has no thesis. It is structured to let a reader form one — by surfacing the disclosures that the "memory is sold out" narrative skates over.

The motivating dispute

The headline story is simple: SK Hynix and Micron have both publicly said their HBM supply is sold out through calendar 2026 on volume and price (Kwak Noh-Jung, SK Hynix Q3 2025 call; Sanjay Mehrotra, Micron Q1 FY26 call). DRAM contract prices were +95% Q1 2026 quarter-over-quarter, with another +63% projected for Q2 (TrendForce). On its face: the boom is real, the producers are safe, demand is durable.

The pushback the wiki exists to test: "sold out" is allocation talk, not enforceable take-or-pay. The disclosed binding commitments sit somewhere very specific in the chain — and not where most coverage assumes.

  • Nvidia 10-Q (Oct 26, 2025): $50.3B in manufacturing/supply/capacity commitments. An additional $26B in multi-year cloud-service commitments. $3.5B prepaid to suppliers. $2.77B in excess-inventory purchase obligations already accrued on the balance sheet, plus a $4.5B single-quarter charge in Q1 FY26 for H20 excess inventory when USG export-license requirements collapsed Nvidia's China product demand — proof of the mechanism by which contracted obligations convert into recognized losses.
  • AMD 10-K (Feb 4, 2026): $12.2B in unconditional purchase commitments, up from $4.3B fifteen months earlier.
  • Micron Q1 FY26 10-Q (Dec 18, 2025): Customer prepayments are a material funding source — $146M outstanding mid-2025, with $777M recognized into revenue over the prior nine months.
  • OpenAI–Samsung & SK Hynix Stargate LOI (Oct 1, 2025): Targets 900,000 DRAM wafer starts per month — roughly 40% of global DRAM output. Disclosed as a Letter of Intent, not a binding offtake.

What the chain shows: the entities with the largest signed take-or-pay-style exposure are Nvidia and AMD, not the hyperscalers. The producers — having watched Qimonda die in 2009 and Elpida in 2012 — are responding to AI demand with capex restraint that creates the scarcity rather than relieves it. The hyperscalers built optionality into their position (LTAs and LOIs, not equity, not binding wafer commitments). Where the loss lands if AI demand misses is the question this wiki tries to answer with primary documents.

Single most load-bearing datapoint: Micron disclosed on its Q1 FY26 call (Dec 17, 2025) that HBM consumes ~3× the wafer area per GB of DDR5, and that Micron is currently meeting only 50–66% of core-customer HBM demand. This is the physical link between AI accelerator demand and commodity DRAM scarcity: every wafer routed to HBM removes roughly three wafers of equivalent DDR5 from the market — without any of the Big 3 having "underinvested" in the conventional sense.

Sections

1. Demand

  • demand/hbm-per-accelerator.md — HBM content per GPU across H100 (80GB), H200 (141GB), B200 (192GB), B300/GB300 (288GB / 576GB), MI300X (192GB), MI325X (256GB), MI355X (288GB). The 3.6× memory-per-socket jump from H100 to B300 in two generations.
  • demand/hyperscaler-buying.md — Nvidia/AMD purchase obligations, hyperscaler LTAs (Microsoft 3-year DDR5, Google up-to-5-year), the regime shift from <5% prepayments historically to 10–30% (some 30–40%) today.

2. Supply

  • supply/capex-discipline.md — Verbatim "discipline" quotes from Mehrotra, Kwak, and SK Hynix CFO Kim Woo-hyun. Micron FY26 capex $20B; SK Hynix 2024 capex ~16T KRW (still below 2022's 19T KRW peak); Samsung Memory's late-cycle restraint.
  • supply/hbm-packaging-bottleneck.md — Why the binding constraint is TSV packaging and CoWoS interposer capacity at TSMC, not wafer fab. The 3:1 HBM-to-DDR5 wafer trade ratio decomposition: ~1.85× from die area alone (SK Hynix D1z DDR4 0.296 Gb/mm² vs HBM3 0.16 Gb/mm²), the remainder from TSV yield compounding.
  • supply/fab-pipeline.md — Every announced fab and packaging line, with realistic ramp dates. Micron Idaho (ID1, 2H 2027), Micron NY (delayed), SK Hynix M15X (Cheongju, pilot May 2026), SK Hynix Cheongju advanced packaging ($13B, output 2027), SK Hynix West Lafayette IN (mass production 2H 2028), Samsung Pyeongtaek P4/P5 (1c conversion, not net new wafers). None of these relieve the 2025–2027 squeeze.

3. History — prior DRAM cycles

  • history/prior-cycles.md — The 1995–96 crash, the 2001–02 DOJ price-fixing cartel ($731M aggregate fines across Samsung, Hynix, Infineon, Elpida; 18 executive jail sentences), Qimonda insolvency Jan 23, 2009, Elpida bankruptcy Feb 27, 2012 → Micron acquisition July 31, 2013, the 2018–19 crash + China SAMR antitrust probe, the 2022–23 downturn (Micron's largest annual loss ever at -$5.83B FY23; SK Hynix KRW 7.7T operating loss; Samsung's April 27, 2023 "meaningful production cut" capitulation). The consolidation arc from ~20 DRAM makers in 1995 to three.

4. Structural deals — who eats the loss

  • structure/ltas-and-prepayments.md — Mehrotra's "unprecedented" LTA framing on the Q1 FY26 call; six HBM customers; Micron's $146M outstanding customer prepayments; Nvidia's $4.2B supply/capacity prepayments; Microsoft/Google LTAs at 10–30% prepayment (some 30–40%) vs. historical <5% norm.
  • structure/who-eats-the-loss.md — The risk-allocation walk-through. The asymmetry: Nvidia/AMD signed $62.5B of purchase obligations; hyperscalers signed LTAs and one LOI; producers retained capex risk; no hyperscaler equity stake in any memory maker. Samsung is the marginal Nvidia supplier (qualified only Oct 10, 2025 for GB300) — first to be cut on any softening.

5. Documents — primary sources, annotated

These pages take the most consequential documents and read them honestly, with extended verbatim excerpts. Where the document is freely accessible we link to it; where we mirrored locally we note the path under data/source-pdfs/.

  • documents/micron-q1-fy26.md — Micron's Q1 FY26 prepared remarks and 10-Q (Dec 17–18, 2025). The "3:1 trade ratio," "50–66% of demand," "calendar 2026 fully booked on price and volume," "specific commitments and stronger contractual structures" quotes. The customer-prepayments line on the balance sheet.
  • documents/sk-hynix-q3-2025.md — SK Hynix Q3 2025 earnings release (Oct 29, 2025) and the H1 2025 semi-annual DART filing. The "27% of revenue from a single customer" disclosure and the Kwak Noh-Jung "supply-demand imbalance" quote.
  • documents/nvidia-10q-oct-2025.md — Nvidia's October 2025 10-Q. The $50.3B purchase obligation, the $4.2B prepayments, and the $3.7B FY25 inventory/supply-obligation provision — Nvidia's already-booked writedowns on commitments it can't recover.
  • documents/amd-10k-fy25.md — AMD's FY25 10-K (Feb 4, 2026). $12.2B unconditional purchase commitments, of which $8.5B due in fiscal 2026 — nearly 3× the level disclosed 15 months earlier.
  • documents/stargate-loi.md — The OpenAI / Samsung / SK Hynix Stargate Letter of Intent, Oct 1, 2025. The 900,000 wafer-starts-per-month target (~40% of global DRAM), the LOI-not-binding-offtake framing, the stranded-capex problem if OpenAI funding falters.
  • documents/doj-dram-cartel.md — The 2004–2007 DOJ DRAM antitrust case. Press releases dated Oct 13, 2004 (Infineon, $160M), May 10, 2005 (Hynix, $185M), Oct 13, 2005 (Samsung, $300M), Jan 30, 2006 (Elpida, $84M); $731M aggregate; 18 executive jail sentences. The conspiracy period April 1, 1999 – June 15, 2002.
  • documents/qimonda-insolvency.md — Qimonda AG insolvency filing, Amtsgericht München, Jan 23, 2009. ~10% of global DRAM bit supply removed from market. The single most-cited prior-cycle datapoint for "what producers remember."
  • documents/elpida-bankruptcy.md — Elpida's Feb 27, 2012 Corporate Reorganization Act filing in Tokyo District Court (largest manufacturing bankruptcy in Japanese history at the time); Micron's July 2, 2012 sponsorship 8-K; the July 31, 2013 acquisition close (~$2.5B). The moment the industry consolidated to three.

6. Reference

  • reference/glossary.md — HBM, TSV, CoWoS, WSPM, bit growth, LTA, take-or-pay, ASP, MSRP, 1α/1β/1γ/1c node nomenclature, EUV, etc.
  • reference/primary-sources.md — Where to find the underlying data: SEC EDGAR (Micron, Nvidia, AMD), DART (Samsung, SK Hynix Korean filings), DOJ Antitrust Division archives, PACER (DRAM MDL 1486 docket), TSMC investor materials, Amtsgericht München and Tokyo District Court records for the bankruptcies.

How to read this wiki

  • Every quantitative claim should have a citation in-line. If you find a number without one, it's a bug — flag it.
  • Numbers are dated. Memory prices and HBM allocations are moving fast enough that even three-month-old figures may be stale. Each page notes when its sources were pulled.
  • Coverage of the debate is balanced by design. Where there's a bull and bear case, both get sourced. Where there's a third view (e.g., "the producers' restraint is the bottleneck, not the demand"), it gets surfaced.
  • The wiki favors under-read primary corpora. Micron 10-Q prepayment footnotes, Nvidia's purchase-obligation line, Korean DART semi-annual customer-concentration disclosures, the original DOJ DRAM plea agreements — these get more attention than another sell-side note on HBM4.

Status

Snapshot built May 2026 from sources current through Q1 2026 earnings cycles (calendar Q4 2025 / fiscal Q1 FY26 for Micron and Nvidia; calendar Q1 2026 for SK Hynix and Samsung). See reference/primary-sources.md for how to refresh.