Two AI Financing Contracts, Two Different Credit Risks

Microsoft says OpenAI contracted to buy another $250 billion of Azure services. CoreWeave says NVIDIA agreed to absorb unsold cloud capacity under an order form with a $6.3 billion initial value. Both suppliers also hold equity in their customers or partners. Yet the purchase terms differ sharply. One is described as contracted spending. The other depends on capacity remaining unsold, available and delivered. For bondholders, face value alone says little about the cash available for debt service.

The IMF reviewed 14 AI financing transactions since 2023. It estimates about $300 billion of equity, $650 billion of debt and $1.4 trillion of future service or chip revenue commitments. These measure financing and prospective sales across different periods. They cannot be added into a cash-flow total. The IMF's chart anonymizes its firms, and its public summary does not reveal how much of the $1.4 trillion is firm, contingent or optional. The two named contracts below test its risk mechanism; the chart does not identify them as nodes.

Clarqo has covered AI bond issuance, NVIDIA's financing role and purchase-linked chip warrants. This audit asks a narrower question: which future orders are enforceable, and which financing claims depend on the same counterparty paying?

Microsoft's exposure has two sides

In its October 2025 filing, Microsoft said OpenAI had contracted to buy an additional $250 billion of Azure services. The same filing put Microsoft's total funding commitments to OpenAI at $13 billion. These are a purchase contract and an investment, respectively. Neither is $263 billion of cash already exchanged.

The June 2026 annual report sharpens the exposure. Microsoft said $11.9 billion of its $13 billion funding commitment had been advanced by June 30. It recorded $24.1 billion of fiscal-year revenue from commercial arrangements with OpenAI, including revenue-sharing payments, and $6 billion of related accounts receivable at year end. That revenue is not a clean measure of Azure purchases, because the filing combines commercial flows. The $250 billion is a future contract value, not recognized revenue.

The same report treats OpenAI as an equity-method investee. If OpenAI could not perform, Microsoft could face a weaker investment, difficulty collecting receivables and lower future service revenue at once. That is the IMF's wrong-way-risk path in a named pair. It is a stress test, not a default forecast. The public filing does not give the payment schedule, minimum take, termination triggers or security for the Azure agreement. A bond analyst cannot turn its $250 billion headline into a recovery value without those terms.

Nor does Microsoft's lost right of first refusal make the order optional. The filed partnership announcement says the additional Azure purchase was contracted while OpenAI gained freedom to use other compute providers. Exclusivity and a purchase commitment answer different questions. The commitment can remain while Microsoft's share of OpenAI's later demand falls.

NVIDIA's backstop is a different promise

CoreWeave's September 2025 Form 8-K names NVIDIA in an order form with an initial $6.3 billion value. NVIDIA agreed to buy residual CoreWeave cloud capacity when other customers do not fill it, through April 2032. This is a backstop on unsold capacity, not proof that NVIDIA will buy $6.3 billion of service irrespective of utilization, delivery or availability. The filing allows either party to terminate after a qualifying breach, or after an insolvency proceeding remains unresolved for 90 days.

In January 2026, NVIDIA also bought $2 billion of CoreWeave shares, according to CoreWeave's June 2026 quarterly report. NVIDIA supplies CoreWeave with GPUs, the order-form filing says. Distress at CoreWeave could damage NVIDIA's equity position and future chip sales. Public documents do not quantify NVIDIA's remaining chip orders from CoreWeave, so the second exposure cannot responsibly be added to the $2 billion stake.

The backstop also cuts the other way: NVIDIA is a buyer of last resort for capacity. CoreWeave's failure could reduce NVIDIA's purchase obligation if the contract terminates, even while its shareholding loses value. Calling every leg an equal revenue promise would conceal that asymmetry.

For CoreWeave creditors, the counterparty map is more important than the backstop headline. At June 30, 2026, the company reported $35.6 billion of total debt. Of that, $13.6 billion was outstanding under delayed-draw facilities backed by contributed contracts and pledged cash flows, generally from investment-grade customers. The filing does not allocate that $13.6 billion to NVIDIA's backstop or any one named customer.

Its DDTL 4.0 facility illustrates another trap in the numbers. The facility limit was $8.5 billion, but only $2.9 billion was drawn at June 30: $1.4 billion floating rate and $1.5 billion fixed rate. The facility is non-recourse to CoreWeave apart from limited carve-outs and is secured by the borrowing subsidiary's equity and assets. The public filing does not say its collateral is NVIDIA's backstop. A lender needs the borrower and collateral schedule before treating CoreWeave's debt as one undifferentiated claim on all customers.

The decision is about contract quality

The IMF's aggregate is a map of potential links, not an audited receivables ledger. The two named chains already show why. Microsoft's Azure commitment is described as contracted, but its public terms leave recovery uncertain. NVIDIA's capacity purchase depends on unused, delivered service and contains an insolvency exit. CoreWeave's debt sits in facilities with different collateral and recourse. A dollar of contract face value is not a dollar of collectible cash, and a debt commitment is not always a drawn loan.

Before buying an AI-linked bond or equity position, map each revenue leg to its payer, delivery date, termination right and legal entity. Then map the borrowing to the exact collateral and guarantee. If a single customer's failure weakens both pledged cash flow and a sponsor's investment, the apparent diversification is thin. If the order can be cancelled or capacity never arrives, the headline backlog is a poor denominator for debt service. The IMF has identified the overlap. Contract disclosures determine its size.