The financing behind an AI data centre rests on three moving parts: the equipment, the customer’s contract and the obligation to repay whoever funded it. Their timelines can differ.
The sums being discussed are large. In August, NVIDIA announced agreements in principle with six financial firms for platforms intended to mobilise more than $500 billion over time. The announcement concerned memorandums of understanding, subject to final agreements, rather than money already deployed.
Individual transactions reveal more about the risks than a fundraising target does. The relevant details include the assets pledged, the repayment schedule and whether the parent company guarantees the borrowing.
Two loans with different protections
CoreWeave’s March 2026 facility, DDTL 4.0, is a delayed-draw term loan. Funding can be drawn over time under agreed terms. It has a maximum size of $8.5 billion and matures on March 31, 2032. Its term sheet provides for monthly principal repayment after specified stabilisation triggers, with the debt fully repaid by maturity.
The collateral includes computing equipment, the customer contract and corresponding data-centre leases. The structure limits recourse to the parent, with exceptions including guarantees covering installation and related performance obligations. CoreWeave’s presentation explicitly includes the borrowing in consolidated debt.
The same company’s August DDTL 5.5 facility has a different design. The $2.6 billion loan matures on September 1, 2031 and has an unconditional parent guarantee. For SOFR loans, interest is the benchmark plus 5.5 percentage points, subject to a zero benchmark floor.
CoreWeave says this second facility’s customer contracts average around three years, while the loan lasts roughly five. The structure permits renewal or re-leasing capacity to other customers, subject to the credit agreement’s criteria.
These are asset-backed lending arrangements with specific borrowers and terms. Describing the collateral as AI infrastructure leaves important questions unanswered: how much principal remains when a contract ends, which future receipts will cover it and who else must pay if they fall short.
The next customer matters
Ratings agency KBRA’s June analysis makes a useful distinction between equipment continuing to function and remaining commercially competitive. Its assessment of re-leasing risk considers customer demand, suitable workloads, switching costs and the cost of power per unit of computing.
A working chip can still require a new price to attract a customer. If that price is lower than a financing plan assumed, the business has less revenue available for its costs and debt. That possibility makes repayment timing and contract renewals central to the financing decision.
A loan that steadily repays principal from contracted receipts presents a different exposure from one relying heavily on a later customer or resale. Guarantees can change who absorbs a shortfall. The label attached to the equipment alone cannot explain that allocation.
Where India fits
India has its own large funding requirements. Blackstone’s February announcement for Neysa envisaged up to $600 million of equity from investors and an intended additional $600 million of debt, subject to documentation. Those plans do not establish the terms or completion of the proposed borrowing.
IFSCA addressed a related question in an August 18 consultation: whether operating and hybrid leases of GPUs and connected data-centre equipment should be enabled as a financial product in India’s international financial centres. The paper sets out a proposal, rather than an operative leasing authorisation.
Its scope distinguishes leasing identified equipment controlled by the customer from buying computing capacity as a service, where hardware can be substituted. The proposed equipment-leasing route covers the first arrangement. That distinction matters because control of a physical asset changes the contractual relationship.
The consultation also discusses a lessor taking on obsolescence and residual-value risk. For that business model to work, the owner needs a credible view of what the equipment can earn from another customer or fetch when sold after its first lease.
For Indian operators and financiers, the useful comparison is therefore among actual contracts. Equipment access, customer commitments, running costs and repayment obligations need to fit together. A larger pool of funding creates possibilities; the terms determine how those possibilities are financed.
The Crunch
The financing of AI capacity depends on who pays, when the debt comes due and who bears the cost if the next customer pays less. India’s compute expansion makes those contract details increasingly important.