AI infrastructure faces bankability challenges

To help fund that buildout, long-term compute offtake agreements (COAs) are emerging. The concept borrows from the power purchase agreements (PPAs) that have underpinned energy project finance for decades. Customers commit to buying or reserving computing capacity over a defined period, creating contracted revenue that can support investment in the infrastructure needed to deliver it.

But the comparison only goes so far. A COA may convert future demand for computing power into contractual revenue, yet not all contractual revenue is financeable. Technology cycles are shorter, performance obligations are more complex and risks often extend far beyond the hardware itself.

Several factors will influence that assessment:

Contract term

Unlike power projects, where PPAs are often aligned with the useful life of the generating asset, COAs must account for the far shorter technology cycles of GPUs and other computing equipment. Shorter contract terms increase renewal, repricing and redeployment risks. Even under longer arrangements, parties need to clearly allocate responsibility for maintenance, replacement and performance upgrades.

Pricing mechanisms

Just as PPAs use different pricing structures, COAs can be based on utilisation, availability or a combination of both. However, payment obligations under a COA are often linked to technical performance metrics, service levels and availability requirements, as well as remedies for degradation or outages. That can make revenue predictability more difficult to assess.

COAs also bring technology and data-related risks that have no direct equivalent in many traditional infrastructure projects. Compute services depend not only on GPUs but often on software licences, data centre hosting arrangements, power supply, cooling systems and network connectivity. Data security, intellectual property, cross-border data transfers and technology regulation may also affect performance and revenue generation.

For lenders, the result is a more complex due diligence exercise. Alongside the customer's ability to pay, they must assess whether the contracts underpinning the compute platform are aligned in terms of duration, performance standards, risk allocation and remedies.

Publicly disclosed transactions suggest the market is experimenting with a range of structures that combine long-term compute contracts with customer prepayments, debt financing, strategic equity investments and residual-capacity purchase arrangements. Two COAs may therefore have a similar headline value while offering very different levels of support for financing.

Contract-supported debt

A five-year agreement between a major hyperscaler and an AI infrastructure provider offers a clear example.

The provider agreed to deliver dedicated GPU capacity from a US data centre in several phases. Subject to deployment and availability milestones, the contract is worth approximately US$17.4 billion through 2031, rising to around US$19.4bn if additional services or capacity are provided.

The provider said cash flow from the agreement would help fund associated capital expenditure and that it expected to raise debt secured against the contract. The parties' obligations only begin once the provider has secured the financing needed to fund the relevant infrastructure.

The structure creates a familiar cycle. Long-term contracted demand improves revenue visibility and can support debt financing for the assets needed to deliver the service. However, payment remains dependent on successful deployment and operation of the contracted capacity.

Customer funding

Another transaction involving a major hyperscaler and a vertically integrated AI cloud provider takes the model a step further.

Under a contract worth approximately US$9.7bn, the provider agreed to make advanced GPUs available from a US data centre campus over five years. The arrangement included a customer prepayment equal to 20% of the contract value. At the same time, the provider entered into a separate agreement to acquire around US$5.8bn of GPUs and related equipment from a technology supplier.

The provider said capital expenditure would be funded through a combination of existing cash, customer prepayments, operating cash flow and additional financing.

Long-term commitments improve revenue visibility, while a prepayment provides immediate liquidity.

In effect, the customer becomes part of the funding package. But the value of that support depends on the legal terms. Lenders will focus on whether the prepayment is refundable, whether the customer has set-off rights, what happens if delivery is delayed and what rights survive termination, insolvency or default.

The analysis therefore extends beyond the amount prepaid to the contractual protections attached to it.

Offtake plus strategic equity

Another model combines a compute contract with an equity investment.

One leading AI company entered into an arrangement with an AI cloud infrastructure provider under which dedicated compute capacity was contracted at an initial value of up to $11.9bn. Alongside that commitment, the customer invested US$350 million through a share issuance. Subsequent agreements increased the publicly announced value of contracts between the parties to approximately US$22.4bn.

An equity investment differs from a prepayment. Rather than paying in advance for future services, the customer acquires a stake in the business.

That alignment of interests may strengthen the commercial relationship, but it does not replace credit analysis. A contract with an investment-grade hyperscaler presents a different risk profile from a contract with a privately held AI company that remains dependent on continuing access to capital markets.

Lenders must also distinguish between economic support and enforceable payment obligations. Strategic relationships may provide comfort, but they do not necessarily create rights that lenders can rely upon in a restructuring or enforcement scenario.

Residual capacity risk

Another transaction focuses on a different challenge: what happens when capacity is built but not fully sold?

Under an existing master services agreement, a leading GPU manufacturer obtained access to an AI cloud provider's residual unsold computing capacity. The initial order was worth approximately $6.3bn and covers residual capacity through April 2032, subject to termination rights and service-availability requirements.

The arrangement functions less like a conventional COA and more like a residual-capacity backstop.

Such structures can reduce utilisation risk, but they can also increase concentration risk. The same counterparty may simultaneously be a major equipment supplier, investor and purchaser of surplus capacity. Risks that appear diversified on paper may therefore remain economically linked.

Different assets, different financing structures

COAs are only one part of the evolving AI infrastructure finance market.

In one large US project, a technology company and funds managed by a leading alternative asset manager formed a joint venture to develop and own an AI data centre campus. The funds hold 80% of the venture and the technology company 20%.

Together they committed approximately US$27bn towards development costs covering buildings and long-term power, cooling and connectivity infrastructure.

Much of the investment relates to infrastructure likely to outlast several generations of computing equipment.

A separate US development illustrates a different approach. An AI infrastructure developer, an alternative asset manager and a digital infrastructure investor announced a US$15bn joint venture for a 1.2GW campus. The project's second phase later secured approximately US$7.1bn in construction financing from a banking syndicate, while a major cloud customer committed to a long-term lease.

Further projects being developed for large cloud customers suggest the scale at which the market could evolve. Reports in late 2025 indicated banks were preparing approximately US$38bn of debt financing for developments in Texas and Wisconsin. Although these reports related to prospective financings rather than completed transactions, they point to growing lender interest in the sector.

Taken together, these examples suggest a layered approach to financing. Land, buildings, power systems and cooling infrastructure may support long-term property-backed or project-finance debt. GPUs and servers may be funded through equipment leases, asset-backed facilities or shorter-term debt. COAs may support revenue-backed financing, while customer prepayments, guarantees and equity capital strengthen different parts of the capital structure.

The goal may not be to finance an entire AI data centre through a single structure. Rather, it is to align each source of capital with the assets, contractual rights and cash flows that can support it.

From contract value to bankable cash flow

A US$10bn agreement may support less debt than a smaller contract if payments depend on deployment milestones, service-level performance or a weaker counterparty. Conversely, a smaller agreement with a strong customer, meaningful prepayments and robust minimum-payment obligations may be more attractive to lenders.
For lenders, the analysis extends beyond projected revenues. Key considerations include:

• whether the customer's creditworthiness and the contract term support the proposed debt tenor and repayment profile;
• how much of the stated contract value represents a binding minimum payment or purchase commitment;
• how technology-refresh, replacement and obsolescence risks are allocated;
• whether the contract provides predictable cash flow after accounting for deductions, damages, termination rights and termination payments;
• whether lenders benefit from assignment rights, cure rights and step-in protections; and
• how different creditor groups would interact if the project encountered financial distress.

This is where the comparison with traditional power project finance starts to break down.

Power projects are generally financed against long-lived assets whose useful lives broadly match their financing profiles. AI infrastructure requires lenders to assess assets with very different economic lives within the same project.

The challenge is to develop financing structures that can accommodate successive generations of technology while preserving the contractual cash flows and legal protections on which lenders depend. The structure must be robust enough to not only survive an exuberant market, but also the aftermath.

That is likely to be the defining test of whether AI infrastructure can achieve the same level of bankability that project-finance markets have long associated with conventional infrastructure assets.