Take-or-Pay Contracts Rewrite Who Owns AI Demand Risk

There's a sentence buried in a securities filing that quietly decides who eats the loss when AI demand cools. It says the customer pays for reserved computing capacity whether or not it actually uses it. That one clause is why a business selling a product that barely existed five years ago can borrow like a utility.

Strip the jargon and take-or-pay is simple. You reserve a block of compute for a set term, and you pay for it anyway. Use every GPU-hour and you got what you bought. Use none and the invoice still lands.

The clause exists because the campus gets built before the buyer needs it, and a GPU hall has no second act if the tenant walks. So the seller locks in a floor. Buyers sign because capacity, not price, has been the real constraint.

CoreWeave's June quarter filing draws the line cleanly. It splits revenue into committed contracts, which are take-or-pay, and on-demand work, which is pay-as-you-go. In that quarter, 98% of revenue came from the committed bucket. Some agreements add a make-up right, letting a buyer bank an unused slice and draw it down later, so the money is deferred rather than forgiven.

Now sit in the lender's chair. Nobody writes a billion-dollar construction loan against a revenue forecast. They write it against a contract. CoreWeave says it plainly in its annual filing, funding infrastructure development through asset-level debt supported by take-or-pay customer contracts.

Buyers often hand over cash before they ever get service. Across CoreWeave's active contracts, the weighted-average prepayment runs 15% to 25% of total contract value. That money shows up as deferred revenue on the seller's books and as cash out the door for the buyer.

The same logic is spreading to power. Duke Energy has proposed a large-load tariff for North Carolina data centers. Under it, they would pay a minimum bill for at least a decade regardless of actual use.

At the project level, campuses increasingly sit inside special purpose vehicles. Meta's Hyperion campus is a $30B example, co-owned by Meta and Blue Owl. A residual value guarantee of roughly $28B sits in the footnotes with no liability recorded against it.

Then there's the number everyone quotes. Remaining performance obligations are the contracted revenue a company has not delivered or billed yet. Backlog is a looser house definition, usually RPO plus extra amounts management expects from committed deals. That's why the two rarely match.

CoreWeave (CRWV) reported $103.7B of unsatisfied RPO at June 30, 2026. Only 41% is expected to convert within 24 months, with another 39% falling in months 25 through 48. The rest sits years out, closer to an aspiration than a receivable.

Now flip to the buyer. Microsoft (MSFT) discloses $194.1B of purchase commitments, and $169.0B of that comes due in FY2027. The footnote is explicit that these primarily relate to datacenters and include take-or-pay contracts.

Prepayments can get big enough to look like borrowing in reverse. Oracle's deferred revenue from customer prepayments with a significant financing component rose $11.4 billion in its latest quarter. CoreWeave's deferred revenue climbed to $9.7B from $8.2B over six months. That's customer money up front against a promise to deliver later.

So who actually eats the risk? The buyer, on paper. But the buyer's ability to pay is the hinge nobody prices properly.

CoreWeave's June quarter disclosed that three customers made up 72% of revenue, with the largest at 36%. For FY2025, Microsoft alone was about 67%.

The credit market is already marking this. S&P cut Oracle to BBB-/A-3 in July 2026, citing aggressive AI spending and customer concentration. Analysts estimate OpenAI accounts for roughly half of Oracle's $664B RPO, a figure Oracle has not disclosed.

None of this is a default story, and it would be wrong to dress it up as one. CoreWeave trimmed its FY2025 outlook over a customer-related capacity delay, not a customer exit. Nvidia (NVDA) agreed to buy residual unsold capacity through April 2032, an initial $6.3B backstop on utilization risk. Microsoft's own risk factors warn about overestimating AI demand and impairing infrastructure.

Take-or-pay doesn't kill demand risk. It moves it to whoever signed the dotted line. The honest test of any AI infrastructure story is three things: the conversion schedule, the counterparty's balance sheet, and what the footnotes disclose.

Read the contract, not the backlog headline. Utilities, private credit funds, and hyperscalers are all copying the same clause.

Disclosure: The Signal holds no position in CRWV. Positions may change. This is not financial advice.