Renting an Nvidia chip for one hour is about to get something most commodities waited decades for: a futures market. Compute futures, contracts that let buyers and sellers agree today on what GPU rental will cost months from now, are scheduled to start trading on Oct. 5, when CME Group plans to list two contracts tied to Nvidia’s H100 and B200 chips, pending regulatory review. But before that market even opens, the first public price chart for AI compute has already been pulled offline.
Start with the basic idea. A bakery that goes through tons of flour hates surprises, so it locks in a wheat price ahead of time. An AI company renting thousands of chips has the same worry about its compute bill. On the other side, a cloud company that owns the chips worries about the opposite: rental prices dropping before the hardware is paid off. A compute futures contract lets both sides fix a number now. Each CME contract covers about one month of rent for a single chip and settles in cash against an hourly rental index published by a data firm called Silicon Data. No chip ever gets shipped. Only money moves.
A futures market also does something quieter. It publishes a forecast. Line up the prices for next month, three months out and a year out, and you get a curve showing where traders think the cost of compute is heading. Prediction market Kalshi built exactly that kind of curve. In July it began publishing “forward curves” for Nvidia B200, H200 and A100 rentals, stitched together from its own weekly and monthly GPU price markets. If you read my piece on how fed funds futures and Kalshi priced the same Fed decision differently, this is the same machinery, just pointed at chips instead of interest rates.
According to a Semafor report published Sept. 15, the Commerce Department told Kalshi in August to take those curves down, citing national security. Kalshi reportedly went along without an announcement, although the markets underneath the curves stayed open. Commerce disputes the story. A spokesperson said the department never asked Kalshi to remove any market. The same report says Commerce also pressed the Commodity Futures Trading Commission (CFTC), the agency that supervises futures trading, to hold off on approving new compute contracts for 60 days.
Why would a single chart make anyone nervous? Market participants pointed to two things. First, these markets are thin, meaning only a few people trade them, so a handful of orders can swing the price hard. Second, the prices that matter most are for older chips. A curve showing rental rates for older chips suddenly collapsing could shake confidence in the loans and bonds backed by data centers full of that hardware, because those debts get repaid out of rent. That is the same weak joint I described in stranded cost: the AI build-out gets paid for one way or another, and someone ends up holding the loss if the equipment loses value faster than planned.
So compute futures now sit in an awkward place. The tool the AI industry says it needs is weeks from launch, while parts of the government are reportedly uneasy about what its prices might reveal. Two dates are worth watching: Oct. 5, CME’s target launch, and the CFTC’s public comment process on compute derivatives, which the agency opened in August.
My take: Taking the price tag down doesn’t make the price go away.
Not advice. Just how I see it.