MTS

Wiki/Demand

Hyperscaler and accelerator-maker buying

The unusual feature of the 2024–2026 memory cycle is not raw demand. Demand cycles come and go. The unusual feature is the structural change in how the largest buyers are contracting — multi-year LTAs with prepayments at 10–30% (some 30–40%) of contract value, versus a historical norm below 5%.

What "sold out" actually means

On the SK Hynix Q3 2025 earnings call (Oct 29, 2025), CEO Kwak Noh-Jung said HBM was "essentially sold out" through 2026. Mehrotra echoed it on Micron's Q1 FY26 call (Dec 17, 2025): "Our HBM for 2026 is sold out in terms of volume and our negotiations with customers have been completed for calendar year 2026 for volume as well as pricing." (Micron Q1 FY26 prepared remarks; transcript.)

What "sold out" means in practice:

  • Six HBM customers at Micron as of Q1 FY26 (Mehrotra). The plausible six: Nvidia, AMD, Google, Microsoft, Amazon, Meta.
  • Micron meeting only 50–66% of core-customer HBM demand. (Same call. See Tom's Hardware.)
  • SK Hynix: ~27% of H1 2025 consolidated revenue from a single (un-named, market-identified-as-Nvidia) customer. SK Hynix semi-annual DART filing 2025, per TrendForce Aug 18, 2025.

The accelerator-maker side: signed paper

The strongest evidence of binding commitment lives upstream of the hyperscalers, on Nvidia and AMD balance sheets:

  • Nvidia 10-Q for period ended Oct 26, 2025 (Note 11): "These commitments were $50.3 billion, of which substantially all will be paid through fiscal year 2027." Plus an additional $26 billion in multi-year cloud-service agreements. Plus $3.5 billion prepaid to suppliers ($2.0B short-term + $1.5B long-term).
  • Q1 FY26 H20 writedown: Nvidia "incurred a $4.5 billion charge in the first quarter of fiscal year 2026 associated with H20 for excess inventory and purchase obligations" when USG export-license requirements collapsed demand for the H20 China product. Through October 26, 2025, only "approximately $50 million in H20 revenue" had been generated under the August 2025 licenses that partially reopened the channel.
  • $2.77 billion in excess-inventory purchase obligations already accrued on the balance sheet at quarter-end (up from $2.10B at Jan 26, 2025).
  • AMD 10-K filed Feb 4, 2026: $12.2 billion in unconditional purchase commitments at Dec 27, 2025, of which $8.5 billion due in fiscal 2026 — up from $4.3 billion at Sept 28, 2024. Roughly 3× growth in 15 months.

This is where the hardest contractual paper sits in the chain. Nvidia and AMD have committed cash and unconditional purchase obligations. The hyperscalers have not — at least not in disclosed form.

The hyperscaler side: LTAs, not equity

What hyperscalers have signed, per a wave of coverage in April–May 2026:

  • Microsoft: 3-year DDR5 LTA with SK Hynix, "tens of billions of dollars."
  • Google: up to 5-year general-purpose DRAM contract with possible 2-year HBM extension.
  • Prepayment structure: 10–30% of contract value, with some at 30–40%, vs. historical norm below 5%. Coverage: KED Global Apr 2026; TrendForce Apr 9, 2026; TrendForce May 4, 2026.

This is regime change relative to the 2017–18 cycle, when hyperscaler memory contracts were essentially annual and prepayments rare. But — critical for the risk-allocation argument — none of the disclosed structures involve equity in the memory producers (no Apple-TSMC-style stake) and the take-or-pay terms have not been published. The functional substance of the LTAs sits inside ASC 606 footnotes (see structure/ltas-and-prepayments.md) and in Korean-language DART filings that are easier to caveat than to enforce.

The Stargate LOI

The most aggressive signal comes from OpenAI:

  • Oct 1, 2025: OpenAI signed a Letter of Intent with Samsung Electronics and SK Hynix targeting up to 900,000 DRAM wafer starts per month — roughly 40% of global DRAM output — for the Stargate data-center program. Sources: OpenAI announcement; SK Group press release.
  • Disclosed as an LOI, not a binding offtake. If OpenAI funding stalls, the capacity Samsung and SK Hynix build against the LOI is stranded on producer balance sheets — exactly the structure that took Qimonda and Elpida down in prior cycles.

See documents/stargate-loi.md for the annotated text.

The price effect

What the demand-side regime change has done to prices, as of Q1 2026:

  • DRAM contract prices +~95% Q1 2026 QoQ (TrendForce 2026-02-02).
  • +63% projected Q2 2026 QoQ (same source).
  • NAND +75% Q1 2026 QoQ.

These are not gentle climbs. The Q1 2026 print is the largest single-quarter DRAM contract-price jump in the post-cartel era. Whether it's durable depends on the binding paper described above — and on what the buyers do when AI revenue catches up to (or falls short of) the obligations they have signed.