The questions a lender should ask before financing an AI data-centre build
Portfolio sample. Neutral analyst voice. Numbers marked (illustrative) are round figures used to show the shape of a calculation, not statistics.
AI data-centre lending has moved faster than the underwriting discipline around it. The asset class is new enough that many credit teams are applying a colocation playbook to a business that behaves differently: shorter revenue contracts, faster-depreciating collateral, a single dominant vendor, and a power constraint that decides the schedule. The questions below are the ones that separate a financeable build from a well-presented one. None of them require an engineering degree to ask. All of them require an answer before committee.
Offtake and revenue
1. Who is the customer, and can they actually pay for the term they signed? Take-or-pay sounds absolute until you read the remedies. Ask for cure periods, credit support, termination rights, volume step-downs, and whether the obligation survives the customer's insolvency. A commitment larger than the counterparty's balance sheet is a letter of intent with better formatting.
2. What percentage of revenue is contracted, for how long, and at what price? Lenders are usually told the contracted share. They are less often told the shape of the uncontracted share, which is where the equity return sits and where the debt service in the later years comes from.
3. Does the contract start date move with the power date? If the customer's obligation begins on a fixed calendar date while energisation is a third-party dependency, the borrower carries an unhedged delay risk. Check whether liquidated damages flow through the chain or stop at the borrower.
Technical and delivery
4. Does the cluster match the vendor's reference architecture, and what exactly is "equivalent"? Deviations from a supported configuration change support entitlements, performance, and in some cases warranty coverage. The follow-up question is who bears the cost of re-integration if a variant fails to perform.
5. What does the acceptance test measure, who witnesses it, and what happens when it fails? A test without a defined throughput threshold, a defined duration, an independent witness, and a contractual consequence is not a milestone. It is a conversation.
6. Are spares, lifecycle coverage, and replacement lead times contracted? An accelerator that fails with a multi-month lead time is a revenue outage with a maintenance ticket attached. Ask for the coverage terms and the assumed failure rate used in the model.
7. Is the building ready for the rack density the operator says it will run? Rack densities in AI halls have moved from the tens of kilowatts well past a hundred (illustrative, widely reported). A shell designed for air cooling and a liquid-cooled cluster are not the same asset, and retrofits are capital events nobody budgeted.
Power
8. Is the interconnect secured, and what is the utility's track record on that queue? The single most common source of schedule slippage in this asset class is not construction. It is energisation. Ask for the interconnection agreement, the studied capacity, and the utility's stated milestones, then model the downside from those rather than from the EPC schedule.
9. What is the power price, and is it hedged? Power is a fixed operating cost against variable revenue. An unhedged energy assumption in a market with volatile pricing turns a technical asset into a commodity trade the lender did not intend to make.
10. What happens at 70 percent utilisation? Not 100 percent. The base case should show debt service coverage at a realistic utilisation and a realistic price, and the covenant should survive both.
Structure and documentation
11. Which assumptions in the model have no independent evidence behind them? Every deal has a handful: residual value, re-contracting rate, failure rate, PUE, ramp curve. The lender's job is not to remove them but to name them and price them.
12. Is the milestone schedule tied to funding draws, and are the draws achievable in sequence? Delayed-draw facilities fail quietly when two milestones land in the same quarter and the cash need arrives before the test is witnessed.
13. What is the residual value assumption, and who would buy this hardware in year five? If the answer is "the same vendor's customers", the lender is underwriting a relationship, not an asset.
14. Where does the sponsor's equity sit in the waterfall, and how much of it is cash versus in-kind? Contribution of an existing asset or of contracted capacity is not the same as cash at risk. Both can be legitimate; only one absorbs loss.
15. What does the technical adviser say, and who engaged them? Adviser alignment matters. A report commissioned by the sponsor answers the sponsor's question. Lender-side technical due diligence exists to produce the three or four findings that change the structure, and a severity-graded list of what remains unresolved.
What good underwriting looks like
The strongest credit memos in this asset class do three things. They separate the timing risk from the demand risk, because on most of these builds the schedule is the binding constraint. They treat uncontracted revenue as equity risk rather than as a base case line. And they can state, in a sentence, the single event that would impair the loan and the clause that responds to it.
The weakest memos share a tell: a model where the power arrives on time, the acceptance test passes first time, the customer performs for the full term, and the price holds. Any one of those can be true. All four being true is not an underwriting assumption. It is a hope with a discount rate.
Lenders who ask these fifteen questions will still finance imperfect deals, because every deal is imperfect. What they will not do is discover the imperfection after the first drawdown, when the only remaining lever is a waiver.