Ornn, the compute exchange that publishes GPU price indices, argues that the way a data center sells its chips often decides whether it can borrow the money to buy them. The company makes the case in an article posted on X on 5 November, the first in a series on buying, selling, and financing GPU capacity.
Ornn builds the argument on a made-up provider. It is offered 1,024 Nvidia B300 chips but cannot pay for them outright. A lender sees a business with no signed customers and uncertain rental income, files it as “high-yield” (a polite word for junk), and demands an interest rate that kills the project.
The fix in Ornn’s example is a reservation. A hyperscaler, one of the giant cloud companies with plenty of cash, commits to 640 of the chips for five years at $4 per GPU-hour and pays even if the chips sit unused. That “take-or-pay” deal is worth $22,425,600 annually. The provider promises the chips will be available 99.9 percent of the time, and owes service credits, or risks losing the contract, if it falls short. With that signed, the lender offers a term loan, drawn in stages, priced at SOFR plus 5 percent and secured by the GPUs themselves.
Certainty costs upside. Locked in at $4, the provider earns nothing extra on those 640 chips if B300 rental rates climb. Ornn adds that longer commitments generally earn the customer a lower hourly rate, so the buyer takes a discount in return for carrying the risk that its own demand shrinks.
The other 384 chips go out on demand. Customers rent them with no fixed term, are billed by the second or minute, and hand them back when finished, at an hourly price above the reserved rate. Ornn’s example buyer is a company that adapts open-source models for clients. One client, a legal startup, rents two eight-GPU servers to run experiments side by side, drops one after comparing results, and releases the other when the project ends. A multi-year reservation would leave that provider paying for chips long after the work dried up. Ornn compares the eventual blend to electricity retailers, who cover part of their expected demand with long contracts and buy whatever remains at short-term prices.
Spot is the third route, for hardware that would otherwise sit idle. In the example, a server goes unrented at the on-demand price, so the provider offers it to an academic researcher at, say, half that rate. The catch is that the work can be stopped at short notice. Days afterward, an inference provider asks for eight B300s on demand, the researcher is cut off, and the chips move to the paying customer.
Ornn has a stake in all of this. It sells price indices, and the article closes its explanation by describing them. The Ornn Compute Price Index tracks only on-demand rental prices in dollars per GPU-hour, drawn from executed transactions, and leaves out reservations and spot. Its H100 and B200 benchmarks underpin GPU compute futures that settle in cash each month on ICE Futures U.S.
The scope of that index shapes how to read it. In Ornn’s own example, 640 of 1,024 chips are priced by a contract that no index can move, and the index follows the uncommitted remainder, the portion a provider keeps open because it expects rates to rise. The $4 reservation and the half-price spot discount are illustrations Ornn constructed, and the article does not present them as observed market rates.
The series continues with a twist on the example: twelve months into the five-year deal, the hyperscaler no longer needs all 640 chips but still owes the payments, so it resells the surplus to someone else. That secondary market is where a contract signed at $4 first meets whatever the going rate is on the day.
Ornn, in an article posted on X on 5 November 2026.
