Two weeks ago, I wrote about the hidden debt layer behind the AI infrastructure boom. Since then, readers have asked how I distinguish a sound private-credit structure from one that may become a refinancing problem later. I have spent 25 years analyzing credit structures where the investment story looks convincing until you examine the contracts and cash flows. Data centers have some of the same characteristics I saw during the fiber and shale booms. I think five questions provide a useful first filter. With the right documents, I can make that initial assessment in about 20 minutes. Missing documents are useful information too.
First, I want to know whether the power is actually secured. A data center without power is not a viable data-center project. I draw a sharp distinction between power being announced and power being contracted. I want an executed interconnection agreement, a credible position in the grid queue, and a realistic connection date. I also want to know who bears the cost if that date slips. A statement that power is “allocated” tells me little without contractual protection. Local permitting matters as well. A financed project with a site and prospective customer can still fail if rezoning, permits, or community opposition delay construction. I therefore ask for the interconnection agreement and the financial consequences of a two-year delay. I think a signed agreement, a credible connection date, and defined protections provide useful evidence. A letter of intent and a list of outstanding approvals do not.
Second, I examine whether the customer commitment is actually binding. When I lend to a data-center SPV, I am not lending directly to Microsoft, Google, or Amazon. I am lending to a vehicle whose cash flow depends on a contractual relationship with one of them, or sometimes on another intermediary that itself relies on a hyperscaler. That distinction matters. A 15-year lease with limited termination rights and a parent guarantee is very different from a capacity reservation that allows the customer to reduce its commitment as AI demand changes. I therefore focus on termination, change-of-control, assignment, and payment provisions. I want to know whether the obligation is genuinely “hell or high water.” Oracle illustrates why looking through the structure matters. Its growing role in AI infrastructure does not make it an equivalent counterparty to the largest hyperscalers. I therefore assess where the economic risk ultimately sits, rather than relying on the logo on the presentation.
Third, I stress the cash-flow model and the capital structure. I typically ask what happens at 60% utilization, with electricity costs 30% above plan, or under another project-specific downside case. I want sufficient debt-service coverage before assumptions become heroic. A 1.4x DSCR at 60% utilization gives me more comfort than 1.1x at 95%, even if the latter looks better in the base case. I also examine technology risk. GPUs may sit outside the SPV, with separate financing, ownership, and useful lives. The building can therefore outlast the equipment, but adapting it for a new generation of chips may require additional capital. I think lenders must understand that mismatch before relying on long-dated contracted cash flow.
Fourth, I focus on refinancing, liquidity, and visibility. Many private placements mature in five to seven years while the underlying lease runs for 15 years or more. That maturity mismatch leaves the project dependent on a future lender. Refinancing becomes harder when credit spreads widen, or private-credit funds become more selective, even if the project continues to perform. I therefore want to see amortization, cash sweeps, interest-rate protection, and sufficient headroom to absorb higher refinancing costs. I also want regular construction reporting, independent oversight, and credible valuation evidence. A project that requires refinancing in 2027 or 2028 needs a realistic source of capital, not an assumption that the next lender will accept the same terms. For now, however, my biggest concern is still whether the asset can actually be built and whether the contracted cash flow can support the debt.
I think the simplest filter is whether both power and the customer commitment are sufficiently secured. The broader industry pipeline may eventually prove much smaller than announced capacity, because projects can fail to obtain electricity, permits, financing, or customers. A proposed 7-gigawatt project without a signed tenant is, in my view, a development opportunity rather than a financed data-center asset. I remain constructive on the need for computing infrastructure, but I am cautious about debt structures built on high leverage and optimistic assumptions. I would rather establish the quality of the power contract, customer obligation, downside cash flow, refinancing plan, and valuation evidence in the first 20 minutes than discover the weakness when the loan comes due.
Alessandro Rovelli
Rovelli Research
Independent credit investment research