Goldman Sachs and JPMorgan are looking at trading futures contracts tied to the rental price of GPUs, the chips that power most AI work, according to The Information. The two banks are early in the process. They may not go through with it at all.
That last part matters and tends to get lost. "Early stages of exploration" is bank-speak for an idea someone is allowed to keep working on, not a product with a launch date. Treat the whole thing as a signal about where the money is pointing, not a done deal.
Why a bank wants this
The pitch is straightforward. Banks are lending enormous sums against AI infrastructure, and the value of that collateral swings with GPU rental rates. If those rates crater, the loans look worse and the hardware backing them is worth less. A futures contract lets a lender take a short position to offset the damage. Cloud providers get the mirror image: a way to lock in pricing instead of riding the volatility.
Goldman's own research is part of why this is surfacing now. The firm projects data center power demand will climb 175% by 2030 against 2023 levels, the kind of number that makes a derivatives desk start sketching contracts. Whether GPU rental prices actually stay volatile enough to need hedging is a separate question, and not everyone agrees they will.
The exchanges already moved
This is the part the banks are reacting to. CME Group said on May 12 it plans to launch compute futures later this year, pending regulatory review, in partnership with Silicon Data, a GPU benchmarking firm backed by trading house DRW. The CME announcement leans hard on the oil comparison.
"Compute is the new oil of the 21st century." That is CME chief Terry Duffy, and it is exactly the line you would expect from the exchange that wants to run the market. The comparison is doing a lot of work, because oil gets burned once and a GPU can be re-rented until it depreciates.
Intercontinental Exchange followed with its own plan for GPU compute futures built on Ornn's index. So by the time Goldman and JPMorgan are "exploring," the venues are already lined up. The banks would be customers and market-makers, not pioneers.
The benchmark problem nobody has solved
Here is the catch. You cannot trade a futures contract without a price everyone trusts, and compute pricing is a mess. Rates for the same Nvidia H100 vary by provider, by region, and by contract length. Silicon Data and Ornn are both trying to be the reference price, but a contract is only as good as the index under it, and these indexes are young.
The other thing worth flagging: the so-called alternative venues abroad described in some coverage are mostly prediction-market and on-chain experiments. Polymarket has touted an institutional block trade settled against the Ornn Compute Price Index, which tracks H100 rental pricing. Promising, sure. A liquid global market, not yet.
What to watch
The real milestone is regulatory sign-off on the CME and ICE contracts, expected to come to a head later this year. If those clear and trade with actual volume, the bank question answers itself. If the benchmarks stay thin, Goldman and JPMorgan have a convenient reason to keep "exploring" indefinitely.
