Wiki/Primary documents
Nvidia 10-Q for period ended October 26, 2025
Date: Filed November 19, 2025; period ending October 26, 2025.
Why it matters: Nvidia is the load-bearing balance sheet in the AI memory chain. The October 2025 10-Q quantifies the load with four disclosed figures: $50.3B in manufacturing/supply/capacity commitments, an additional $26B in multi-year cloud-service commitments, $3.5B in prepaid supply and capacity agreements, and $2.77B in accrued excess-inventory purchase obligations on the balance sheet at quarter end. Most importantly, the 10-Q discloses a $4.5 billion Q1 FY26 charge for excess inventory and purchase obligations triggered by USG export-license requirements on Nvidia's H20 China product — a fresh writedown, in a single quarter, on commitments Nvidia could not recover.
Primary source: Nvidia 10-Q for period ended October 26, 2025 (Note 11 — Commitments and Contingencies).
Load-bearing figures — verbatim from Note 11
$50.3 billion in manufacturing/supply/capacity commitments
Verbatim:
"Manufacturing production, long-term supply and capacity, and other related commitments reflect long lead and cycle times for our current and future product architectures... As of October 26, 2025, these commitments were $50.3 billion, of which substantially all will be paid through fiscal year 2027."
Why: Largest single disclosed unconditional purchase obligation on any balance sheet in the AI memory chain. Roughly 4× AMD's $12.2B. Nvidia is the entity carrying the most binding contractual paper in the entire memory–accelerator–compute supply chain.
$26 billion in multi-year cloud-service commitments
Verbatim:
"Multi-year cloud service agreement commitments as of October 26, 2025, were $26 billion for which $1 billion, $6 billion, $6 billion, $5 billion, $4 billion, and $4 billion will be paid in fiscal years 2026 (fourth quarter), 2027, 2028, 2029, 2030, and 2031 & thereafter, respectively."
Why: A separate obligation from the manufacturing commitments — Nvidia has additionally committed $26B to multi-year cloud-service agreements (reflecting its consumption of compute capacity from hyperscalers and neoclouds for internal use and circular Nvidia-financed buildouts). The 2026–2031 pay-down schedule is disclosed; the counterparty mix is not.
$3.5 billion in prepaid supply and capacity agreements
From the Other Assets and Prepaid Expenses footnote:
- Long-term prepaid supply and capacity agreements: $1,536M at Oct 26, 2025 (vs. $1,747M at Jan 26, 2025).
- Short-term prepaid supply and capacity agreements: $2,000M at Oct 26, 2025 (vs. $3,300M at Jan 26, 2025).
- Total: ~$3.5 billion, down from ~$5.0 billion at fiscal year-end Jan 2025.
Why: Cash on the table at suppliers — including memory partners — that does not recover if products do not ship. The decline from $5.0B → $3.5B over nine months reflects the H20 writedown absorbing some of the short-term prepayments. The 10-Q does not itemize counterparties.
$2.77 billion in excess inventory purchase obligations (accrued liability)
From the Accrued and Other Current Liabilities footnote:
- Excess inventory purchase obligations: $2,770M at Oct 26, 2025 (up from $2,095M at Jan 26, 2025).
Why: Sitting on Nvidia's balance sheet as an accrued liability is $2.77B of already-recognized losses on purchase obligations Nvidia knows it cannot recover. The line grew $675M over nine months. This is the leading-indicator metric the rest of the AI supply chain should track.
$4.5 billion Q1 FY26 H20 charge — the single most-pointed writedown
Verbatim:
"In April 2025, the U.S. government, or USG, informed us that a license is required for exports of our H20 product into the China market. As a result of these new requirements, we incurred a $4.5 billion charge in the first quarter of fiscal year 2026 associated with H20 for excess inventory and purchase obligations, as the demand for H20 diminished."
Why this is the standout finding: A single regulatory decision — USG export-license requirements — triggered $4.5B of writedowns in one quarter on a single product line. This is the test of the purchase-obligation structure. When demand on a contracted product disappears, the $50.3B in commitments is not phantom paper — it converts into real, recognized losses. The H20 charge is the proof of the mechanism, and it is the leading indicator of what scale of writedown a broader AI demand miss would produce.
Nvidia subsequently received some licenses in August 2025 ("to date, we have generated approximately $50 million in H20 revenue under those licenses" — i.e., a small fraction of the foregone demand).
What the 10-Q does not contain
- Itemized HBM vs. CoWoS vs. logic-wafer composition of the $50.3B. The 10-Q reports the aggregate; the breakdown is private.
- Counterparty names. No itemization of how much is owed to TSMC vs. SK Hynix vs. Micron vs. Samsung. The same is true for the $26B in cloud-service commitments and the $3.5B in prepayments.
- Cancellation terms. Whether the $50.3B is enforceably "unconditional" or has carve-outs is not disclosed at the line-item level.
These are the same omissions Micron's 10-Q has on the supplier side (see micron-q1-fy26.md). Both ends of the contract are partially obscured. The aggregate numbers, however, are firm.
Reading this 10-Q against Micron's
The two filings, read together, frame the cycle:
| Micron Q1 FY26 | Nvidia Oct 26, 2025 | |
|---|---|---|
| Customer prepayments received | $146M | — |
| Consideration payable to customers | $1.64B | — |
| Prepaid supply/capacity (paid to suppliers) | — | $3.5B total ($2.0B ST + $1.5B LT) |
| Manufacturing/supply/capacity obligations | — | $50.3B |
| Multi-year cloud service commitments | — | $26B |
| Excess inventory purchase obligations (accrued) | — | $2.77B |
| Single-quarter writedown (H20, Q1 FY26) | — | $4.5B |
Micron is partially financed by its customers ($146M of prepayments). Nvidia is partially financing its suppliers ($3.5B prepaid) and has $76.3B in combined contractual commitments. Only Nvidia is taking visible writedowns — $2.77B already accrued plus the $4.5B H20 charge. The producer-side balance sheets are clean; the OEM balance sheet is where loss first appears.
This is the asymmetry structure/who-eats-the-loss.md walks through.
What to read next
documents/amd-10k-fy25.mdfor the second-largest accelerator-OEM balance sheet.documents/micron-q1-fy26.mdfor the supplier-side counterpart.structure/who-eats-the-loss.mdfor the synthesis.