MTS

Wiki/Structure

LTAs, prepayments, and the new contract regime

The previous project (ai-capex) found that the load-bearing risk-allocation structures in AI infrastructure are buried in SPV agreements, lease terms, and rate-class orders that most coverage doesn't read. The memory equivalent lives in ASC 606 footnotes, balance-sheet "customer prepayment" lines, purchase-obligation tables, and Letters of Intent.

This page collects what is disclosed about the new contract regime. The companion page who-eats-the-loss.md reads the chain top-to-bottom.

What Mehrotra has actually said

The strongest public characterization of the new contracts comes from Mehrotra's Q1 FY26 prepared remarks (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."

Contracts "involve specific commitments and stronger contractual structures" — language he called "unprecedented."

Source: Micron Q1 FY26 prepared remarks PDF; Motley Fool transcript.

Two things to notice:

  1. Price is locked, not just volume. In prior cycles, LTAs covered volume; price renegotiated quarterly. Mehrotra is saying that for calendar 2026, Micron has agreed both axes with all six HBM customers. This removes producer downside to a spot price crash. It also transfers inventory risk to buyers.

  2. "Specific commitments and stronger contractual structures." This is as close as Mehrotra gets to admitting take-or-pay. He does not use the phrase. He does not have to: the underlying ASC 606 footnotes (revenue recognized over time, with constraint) and the balance-sheet "consideration payable" line tell the rest of the story.

What the balance sheet shows — Micron

From the Micron Q3 FY25 10-Q (filed June 26, 2025; period ending May 29, 2025):

"Customer prepayments made to secure product supply in future periods and other contract liabilities were $146 million."

Source: Micron Q3 FY25 10-Q.

From the same filing: $777M of revenue was recognized over the prior nine months against the prepayment balance. This is not a rounding item — customer prepayments have become a material source of working-capital funding for Micron.

From the Q1 FY26 10-Q (filed Dec 18, 2025; period ending Nov 27, 2025):

"Consideration payable to customers" line item: $1.64B (up from $1.19B at Aug 28, 2025).

Source: Micron Q1 FY26 10-Q.

This is the symmetric risk Micron retains on the agreed prices — pricing-adjustment and return reserves owed back to customers. It rose ~$450M in one quarter.

What hyperscalers signed (coverage-disclosed)

The hyperscaler-side terms are not in any 10-Q — they sit in private agreements summarized in trade press during 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 hyperscalers at 30–40%. Historical norm: below 5%.

Sources: KED Global Apr 6, 2026; TrendForce Apr 9, 2026; TrendForce May 4, 2026.

The 10–30% prepayment is the regime change. In the 2018–19 cycle, hyperscaler memory contracts were essentially annual with prepayments well under 5%. Customers were not putting cash on the table years in advance. They are now.

What Nvidia and AMD signed (10-K / 10-Q disclosed)

This is where the hardest paper sits in the chain:

Nvidia 10-Q for period ended October 26, 2025 (Note 11), verbatim:

  • $50.3 billion in "manufacturing production, long-term supply and capacity, and other related commitments... of which substantially all will be paid through fiscal year 2027."
  • $26 billion in multi-year cloud-service agreement commitments, with pay-down schedule disclosed through fiscal 2031.
  • $3.5 billion in prepaid supply and capacity agreements ($2.0B short-term + $1.5B long-term) — cash on the table at suppliers.
  • $2.77 billion in excess-inventory purchase obligations already accrued on the balance sheet (up from $2.10B nine months earlier).
  • $4.5 billion single-quarter charge in Q1 FY26 for H20 excess inventory and purchase obligations after April 2025 USG export-license requirements collapsed H20 China demand.

Source: Nvidia 10-Q Oct 26, 2025.

AMD 10-K filed February 4, 2026:

  • $12.2B in unconditional purchase commitments at Dec 27, 2025.
  • $8.5B due in fiscal 2026.
  • Up from $4.3B at Sept 28, 2024 — roughly 3× growth in 15 months.

Source: AMD 10-K Feb 4, 2026.

Combined: $62.5B of unconditional purchase obligations sitting on Nvidia and AMD balance sheets — and that's just the manufacturing-and-supply line. Nvidia carries another $26B in cloud-service commitments on top. Nvidia has $2.77B already accrued as recognized excess-inventory losses, plus the $4.5B Q1 FY26 H20 charge — combined evidence that the contractual paper converts into real losses when demand on a contracted product disappears.

What no one signed: equity

A negative finding worth foregrounding: no hyperscaler holds a disclosed equity stake in any DRAM producer. No Apple-TSMC precedent. No JV. No co-investment.

This is the central asymmetry of the cycle. The buyers wanted optionality and got it — multi-year LTAs with prepayments, yes, but no equity exposure. The producers absorbed the capex risk. If demand misses, the capex sits on Samsung / SK Hynix / Micron balance sheets, not Microsoft / Google / Meta's.

The Stargate LOI — the outlier

The exception that proves the rule is Stargate, announced October 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.

Sources: OpenAI announcement; SK Group press release.

But it is an LOI, not a binding offtake. OpenAI has not committed cash. Samsung and SK Hynix have not committed wafers. The structure is a deferred negotiation dressed up as a partnership.

If OpenAI funding falters before the LOI is converted to binding terms, the capacity producers built against it is stranded on producer balance sheets — exactly the failure mode that took Qimonda down in 2009 and Elpida in 2012. See history/prior-cycles.md.

Summary — where the binding paper sits

Counterparty Binding paper Disclosure
Nvidia $50.3B purchase obligations, $4.2B prepayments, $3.7B/yr writedowns SEC 10-Q
AMD $12.2B unconditional purchase commitments SEC 10-K
Microsoft / Google / Amazon / Meta LTAs, 10–30% prepayments Trade press only
Micron $146M customer prepayments, $1.64B consideration-payable SEC 10-Q
OpenAI LOI for 900K WSPM (~40% of global DRAM) Joint press release
Hyperscaler equity in memory makers none n/a

The pattern: the further upstream from the data center you go, the more binding the paper. Nvidia/AMD have signed the most. Hyperscalers have signed less. Memory producers have received prepayments but retained capex discretion. OpenAI's commitment is the loosest of all.

See who-eats-the-loss.md for what that means when AI revenue catches up to the obligations.