Underwriting a Neocloud Take-or-Pay Offtake Contract

A take-or-pay contract with a neocloud is only as good as the weakest of three things: how long the contract term outlasts the GPU depreciation schedule, whether the power actually shows up when the interconnection queue says it will, and how many other lenders are staring at the same borrower's revenue line. Miss any one of these and the credit that looked investment-grade on the term sheet turns into a workout. I've sat in enough of these credit committees to know which line item gets glossed over, and it's rarely the one that ends up mattering.

Start With the Term, Not the Rate

Everyone wants to talk about the rate first. Dollars per GPU-hour, dollars per megawatt, whatever unit the neocloud and its hyperscaler counterparty settled on, that number gets the press release and the LinkedIn post. Duration is the boring line nobody reads twice, and it's the one that decides whether the loan gets repaid.

Here's the mechanics of it. A take-or-pay contract is a bet that the borrower's revenue stream outlives the asset's useful life, or at least the debt amortization schedule. Nvidia's H100 generation is already giving way to Blackwell, and the refresh cadence at the high end has compressed to something like 18 to 24 months. A five-year offtake against a GPU cluster financed over five years looks matched on paper, but that match can break down if the hyperscaler counterparty has a two-year renewal option buried in Section 9 with no obligation to renew at the same rate.

So read the extension language. Actually read it, not the summary a banker put in the deck. What happens at renewal? Does the contract reprice to market, or lock in the original rate? A reprice-to-market clause protects the hyperscaler far more than it protects the borrower, since GPU compute pricing has been drifting down as supply catches up with demand in several regions. If the neocloud's debt service coverage got underwritten off year-one pricing and the renewal resets 20 or 30 percent lower, the coverage ratio that justified the loan in the first place disappears entirely.

The Interconnection Queue Is the Real Delivery Risk

Data center financing spent decades treating power as a solved problem. Sign a PPA, get an interconnection agreement, build. That assumption doesn't hold anymore in the markets where neoclouds are trying to scale. PJM's queue alone carries tens of gigawatts of generation and storage projects waiting on study completion, and four to six years from initial application to commercial operation has become routine in the most congested territories.

That matters to offtake underwriting in a very direct way: a take-or-pay contract with a hyperscaler is worthless if the facility never gets energized on schedule. The diligence question is where the project actually sits in the queue, whether it's cleared the system impact study, and whether the utility has a track record of hitting its own stated timelines in that territory. Utilities miss their own dates more often than borrowers like to admit.

ERCOT, PJM, and MISO are each running different queue reform processes, and each has a different practical effect on how confidently you can underwrite a 2026 or 2027 commercial operation date. ERCOT's connect-and-manage approach has historically moved faster than PJM's serial study process, though it trades that speed for curtailment risk once the project is online. A lender who doesn't distinguish between these regimes is underwriting a single delivery date instead of a probability distribution of delivery dates, which is a different exercise entirely. And every quarter of delay against a fixed take-or-pay commencement date is a quarter of debt service the borrower has to fund from somewhere else, usually a cash reserve that wasn't sized for that contingency.

Stress the interconnection timeline the way you'd stress an interest rate. Run the base case off the utility's stated schedule, then run a downside case using that ISO zone's actual historical slippage. If the base case has the facility energized in month 18 and the region's historical slippage runs another 12 months on top of that, your downside case needs to show the take-or-pay clock ticking for a full year before the asset produces a single kilowatt.

Power Pricing Is a Second, Separate Variable

Queue position tells you when the power might show up. Price tells you what it costs once it does. These get bundled together in casual conversation, and that's a mistake, because they're driven by completely different mechanisms.

Neoclouds signing colocation or build-to-suit deals in tight power markets, Northern Virginia being the obvious case, carry real exposure to real-time and day-ahead price volatility unless they've locked in a long-term PPA or a hedged retail supply contract. Loudoun County and the wider Data Center Alley corridor have already seen local utilities flag capacity constraints serious enough to affect new large-load interconnections. PJM's capacity auction prices have moved up sharply in recent cycles, pricing in exactly this demand growth from AI infrastructure buildout.

So ask the blunt question. Is the borrower's power cost fixed for the life of the offtake, is it a pass-through, or is it sitting exposed to spot with no hedge at all? If it's a pass-through against a fixed-rate offtake, the borrower has an unhedged position sitting in the middle of the deal, and that position gets bigger as capacity prices climb across PJM, ERCOT, and other constrained zones. A simple debt service coverage calculation won't catch this. The power cost curve has to get built into the model year by year; assuming today's rate holds flat for the contract term is how these models quietly lie to the people relying on them.

Counterparty Concentration Is the Quiet Killer

Most neocloud offtake structures run on one, maybe two, anchor counterparties. Microsoft, Meta, Google, and Amazon have all signed large compute capacity deals with neocloud providers to supplement their own buildouts, and CoreWeave's relationship with Microsoft has been public and material enough to show up in CoreWeave's own securities filings.

A single-counterparty offtake, no matter how tight the term and pricing, hands the borrower's entire revenue risk to one hyperscaler's continued willingness to pay. Hyperscalers carry investment-grade balance sheets, and that makes the whole thing look safe on the surface, but a strong counterparty rating says nothing about concentration risk, and I've watched credit teams conflate them more times than I can count. A hyperscaler doesn't have to default to hurt a neocloud's cash flow. It just has to decide, during its own capex reassessment, to bring workloads back in-house, or to renew at a smaller commitment. Take-or-pay language protects against non-payment on contracted volume. It does nothing for the borrower's business model if the counterparty simply declines to expand when the initial term rolls off.

This is project finance behavior dressed up in tech-sector language, and it should get underwritten that way, closer to a power purchase agreement with a single utility buyer than to a typical corporate credit. Lenders in that older world have long required take-or-pay guarantees backed by a parent credit rating, or minimum diversification thresholds across offtakers, before advancing real leverage. Neocloud financing is only starting to adopt that discipline, and frankly it should move faster. There are only a handful of companies on the planet capable of signing the kind of multi-year, multi-hundred-megawatt compute commitment that makes a neocloud's balance sheet work in the first place. That's not a large pool to build diversification out of.

Putting the Three Variables Together

None of this works in isolation. A contract with strong term coverage but shaky interconnection certainty has simply pushed the problem down the calendar. A project with confirmed power and a fast queue position but full exposure to one non-guaranteed counterparty has traded delivery risk for concentration risk, leaving the underlying exposure fully intact.

The only underwriting approach that survives contact with reality stresses all three at once: the downside interconnection timeline, against the downside power price, against a scenario where the anchor counterparty walks at year five instead of renewing. If debt service still clears under that combined scenario, the deal is financeable. If it only clears under the version where everything lands on time, at the modeled power price, with the counterparty renewing right on schedule, that's a bet wearing a credit's clothing, and this market has already been burned once for treating the two as the same thing.