CME Delays GPU Rental Futures a Second Time, Pushing Launch Into January

The contracts would be the first listed derivatives tracking the cost of renting AI computing power, a market that has ballooned as companies race to secure chips for data centres. CME had planned to list futures based on indices from Silicon Data that track rental rates for Nvidia's H100 and B200 chips. The exchange originally aimed to launch in early October before the regulator extended its review by 45 days, moving the target to November 9.

The CFTC has been examining what it has called novel or complex questions raised by the product. Chief among them is whether prices could be manipulated in a GPU rental market spread across many brokers, cloud platforms and one-off deals. Because the underlying cash market is fragmented, a weak benchmark could leave a wide gap between futures prices and what firms actually pay, which would deter the hedgers the contract is meant to serve.

The repeated delays point to the difficulty of building a transparent price for compute. Much of the capacity bought by the largest cloud providers is locked in private contracts with undisclosed terms, leaving little in the way of a public spot market to anchor a futures price. CME is not alone in pursuing the idea; Intercontinental Exchange has also been exploring GPU futures, and its chief executive, Jeffrey Sprecher, said this week that lenders could eventually require data centre borrowers to hedge GPU prices.

The second delay takes a decision past the CFTC's original timetable and into the new year, leaving the launch of the first compute futures contract uncertain.