Article

Reading NVIDIA's $390 Billion Purchase Commitments and $118 Billion Guarantee Exposure

The part of NVIDIA's earnings report most readers skip is not revenue but three sets of commitments: $390 billion in supply purchase commitments, $102 billion in cloud service and data center lease commitments, and $118 billion in guarantee exposure. None of them flow through the income statement, yet together they determine the next several years of cash flow and risk. This analysis explains their accounting treatment, the three paths through which risk propagates, and the three metrics worth tracking.

Key takeaway

Revenue of $115.8 billion is this quarter's result; $390 billion in purchase commitments is the obligation for the years ahead. The first determines current profit. The second determines how much freedom the next cycle allows — and it never appears on the income statement.

1. Start with the three numbers ​

NVIDIA's CFO commentary for the third quarter of fiscal 2027, ended October 25, 2026, disclosed three sets of figures:

ItemEnd of Q3End of Q2Change
Future supply and capacity purchase commitments$390B$279B+$111B
Cloud service agreements and not-yet-commenced data center leases$102B$74B+$28B
Maximum aggregate exposure under guarantees$118Bnot disclosed—

The largest single guarantee item is a credit support guarantee of up to $105 billion signed with SB Energy in August 2026, covering a data center project of roughly 4.25 gigawatts in Ohio across nine buildings, with OpenAI holding a 20-year lease. The company said guarantees phase in as buildings are completed and reach service conditions, with the first tranche expected in fiscal 2029, and that no guarantee payments have occurred to date.

In one quarter, purchase commitments rose $111 billion and lease commitments rose $28 billion. That increment exceeds the annual revenue of most companies.

2. Why none of this is on the income statement ​

To read these figures, start with the accounting.

Purchase commitments are contracted but undelivered purchase obligations with suppliers. They do not hit the income statement because goods have not shipped and costs have not been incurred. They appear in the notes as off-balance-sheet commitments.

Lease commitments fall into two buckets: leases that have commenced are recognized on the balance sheet under lease accounting as right-of-use assets and lease liabilities, while not-yet-commenced leases — part of the $102 billion disclosed here — appear only in the notes until the lease term begins.

Guarantees are contingent liabilities. A loss provision is recorded only when a loss is probable and estimable; otherwise the maximum exposure is disclosed in the notes.

The common thread: all three describe obligations that must be met or may be borne in the future, not costs already incurred this period. The income statement therefore looks clean, while these obligations enter it later through depreciation, rent and purchase costs.

3. Three paths through which the risk travels ​

These figures are not isolated. They form a chain.

Path one: purchase commitments to inventory to write-downs. The $390 billion in supply commitments is mainly multi-year memory purchasing. If inference demand disappoints and committed memory and packaging capacity cannot be absorbed, it first piles up as inventory — which has already risen from $21.403 billion to $39.724 billion this fiscal year — and then turns into valuation reserves. This is the classic loss mechanism of a hardware cycle.

Path two: lease commitments to depreciation to margin pressure. The $102 billion in cloud and data center lease commitments will enter the income statement as rent and depreciation once projects operate. Fixed costs are fixed regardless of utilization. If compute demand declines and utilization falls, those costs hit gross margin directly.

Path three: guarantees to contingent liabilities to credit events. Of the $118 billion in guarantee exposure, $105 billion is tied to the SB Energy Ohio project, whose cash flow depends on OpenAI's 20-year lease. The ultimate risk of that guarantee therefore rests on whether OpenAI can keep paying rent. This is a compute-financing-lease-credit loop: if one link breaks, the guarantee moves from contingent to actual.

4. Why the market calls this circular financing ​

Placed together, the three paths explain concerns about circular financing in AI infrastructure.

The typical structure runs like this:

  1. NVIDIA sells GPUs to cloud providers and neoclouds, generating revenue and profit;
  2. NVIDIA also provides credit support or investment to those customers' data center projects;
  3. Customers rent that compute to model companies, which pay rent to the cloud providers;
  4. Cloud providers service the debt taken on to buy GPUs using rental income;
  5. Part of the model companies' funding comes from investments NVIDIA participates in or from customer prepayments.

Within this loop, the same demand is counted as revenue at several stages. As long as end demand — inference revenue at model companies — grows genuinely and sustainably, the loop holds. Once end demand slows, multiple stages come under pressure at the same time.

This quarter's report provides evidence in both directions:

  • Supporting: operating cash flow of $48.253 billion and free cash flow of $43.777 billion were close to net income, arguing that current growth is backed by real cash rather than receivables.
  • Cautionary: accounts receivable rose from $38.466 billion at the start of the fiscal year to $81.033 billion and inventory from $21.403 billion to $39.724 billion, while off-balance-sheet commitments expanded sharply. Working capital and commitments are expanding together, which means future cash flow depends more on whether downstream buyers pay on time.

5. Three metrics worth tracking ​

First, the ratio of purchase commitments to revenue. At $390 billion against $115.8 billion of quarterly revenue, the ratio is about 3.4 times. A rising ratio means the company is locking capacity for a more distant future; a falling ratio means it is retrenching. Two consecutive quarters of a rising ratio alongside decelerating revenue growth is the most dangerous combination.

Second, receivables turnover. Receivables rose from $38.466 billion to $81.033 billion, roughly 1.1 times, while revenue grew 103%. The two moved roughly in step, which is normal. But if receivables growth clearly outpaces revenue growth next quarter, collections are slowing and customers — especially financing-driven neoclouds — may be straining.

Third, the pace at which guarantees take effect. The company says the first guarantee tranche becomes effective in fiscal 2029 and that no payments have occurred. That statement needs quarterly re-verification: once guarantees begin phasing in, contingent liabilities convert into actual liabilities on the balance sheet.

6. Three implications for the industry ​

First, the picks-and-shovels model is becoming picks, shovels and road-building. NVIDIA does not just sell chips; it also carries supplier prepayments, data center leases and credit support. That extends its influence and its risk surface at the same time.

Second, the real constraint on AI infrastructure is not chip capacity but end demand. The $390 billion commitment can only be absorbed if inference demand keeps scaling. If inference revenue growth lags compute supply growth, price and utilization fall together.

Third, the way we read AI company financials has to change. An income statement alone badly understates risk; commitments, leases, guarantees and working capital in the notes must be read together to judge how much freedom a company has over the next two years.

7. An easily missed comparison: commitment growth versus cash growth ​

Putting two sets of figures side by side is instructive:

ItemThis quarterPrior quarterChange
Total off-balance-sheet commitments (purchase plus lease)$492B$353B+$139B
Operating cash flow$48.253B$24.118B+$24.1B
Net income$65.287B$59.7B+$5.6B

Commitment growth of $139 billion is about 5.8 times the increase in quarterly operating cash flow. This is not alarming, because commitments are multi-year obligations spread over time while cash flow is a single-quarter figure. But it shows one thing clearly: the company is using current cash strength to lock in future capacity far larger than its current scale.

That is the most rational move in an upcycle, and the one that most requires end demand to materialize. The moment the commitment is signed, the risk shifts to whether the output can be sold.

8. Boundaries ​

  • The $390 billion purchase commitments, $102 billion cloud and lease commitments, $118 billion guarantee exposure and SB Energy project details come from NVIDIA's CFO commentary in its November 17, 2026 earnings materials, verified here through media transcription.
  • All three are commitments and maximum exposure figures, not incurred spending or recognized losses; the company states no guarantee payments have occurred to date.
  • No prior-quarter comparable figure was disclosed for total guarantee exposure, so the table leaves it blank rather than estimating.
  • Circular financing is a descriptive term for industry structure, not an accounting term, and is not an allegation against any company.
  • The three tracking metrics are an analytical framework; the company does not report on that basis.
  • This article makes no valuation judgment, does not forecast the share price, and is not investment advice.

9. Sources ​