AI Infrastructure Is a Credit Story

A data center is easy to draw as a building full of chips. A lender sees a bundle of promises that must hold at the same time: the site will receive power, construction will finish, the tenant will pay, the hardware will remain useful, and the debt will mature on terms the cash flow can bear. The AI boom has made the building urgent. It has not made those promises interchangeable.

The financing market is moving to meet the build-out. Moody's says data center developers and landlords are raising equity, bank loans, corporate and securitized bonds, and project finance. It expects leverage to rise for hyperscale projects planned for completion in 2026-28. That is a credit story before it is a technology story. The asset may be built for computation, but its funding depends on real-estate collateral, contracted rent, utility access, and a capital stack that can survive delays.

Start with power. A lease for compute capacity has little value if the grid connection arrives after the customer needs the servers. Ropes & Gray's 2026 account of the investment market describes grid interconnections that can take up to four years. It also warns that a utility "will-serve" letter is not the same as power delivered by a date certain. Those are not engineering footnotes. They determine when rent begins, whether construction interest keeps accumulating, and who absorbs the gap.

The answer can be to bring generation to the site, but that creates a second project with its own fuel, permits, capital, and completion risk. Financing the power plant separately from the building may attract different pools of money. It does not abolish the dependency between them. A lender holding the building needs to know whether the plant will be ready; a lender holding the plant needs to know whether the building and tenant will be there to buy the output.

Tenant credit is the next hinge. A long lease from an established hyperscaler is easier to lend against than an equivalent promise from a thinly capitalized GPU reseller. The price per megawatt may look attractive in both cases. The lender cares about who writes the check after a workload moves, a model architecture changes, or hardware values fall. Guarantees, security, termination rights, and rent-commencement dates can matter more than an optimistic capacity forecast.

As the project matures, the financing changes shape. Ropes & Gray describes land-cost facilities and GPU financing arriving earlier in development, construction and project finance for vertical build-out, and asset-backed securities, bonds, or commercial mortgage structures for stabilized assets. The sequence has a reason. A piece of powered land, an unfinished hall, and a leased facility have different risks. A single headline number called "AI infrastructure investment" hides the transitions between them.

The central question is who holds the stranded-asset risk. If power arrives late, a tenant can seek delay protection while the developer carries interest. If the tenant signs a weak lease, the developer may struggle to refinance. If the building works but its chip-heavy use case changes, the value of the shell and the value of the machines can diverge. A financing that treats all three as one durable asset can discover the distinction at the worst possible time.

None of this argues that data centers are a bad business. It argues for reading them in the same disciplined way one reads any financed infrastructure: contracted demand, completion risk, counterparty strength, collateral value, and the date cash actually starts to flow. AI names the demand. Credit decides who can build enough capacity to serve it and who carries the cost when the timetable slips.

Sources: Moody's, Data centers credit risk insights: https://www.moodys.com/web/en/us/insights/credit-risk/data-centers.html

Ropes & Gray, Data Center Investment in 2026: https://www.ropesgray.com/en/insights/viewpoints/102mvfl/data-center-investment-in-2026-ai-demand-power-constraints-and-private-equity