The coming futures market in AI compute

The writer is a former hedge fund manager and author of the Net Interest newsletter

You can trade all sorts on the Chicago Mercantile Exchange: soyabeans, precious metals, lean hogs, crude oil, interest rates — all the staples that underpin the world’s economy. Now, CME Group plans to introduce a new commodity to its roster: compute, which it dubs “the new currency for the AI economy”.

“Just as oil fuelled the 20th-century economy and evolved from spot trading into a global derivatives market, our futures contracts will now turn compute into a standardised, tradeable commodity,” said Pete Keavey, the CME executive responsible, when the product was announced in August. Others on Wall Street are equally bullish. “A new asset class will be buying futures of compute,” said BlackRock chief executive Larry Fink in May.

For an exchange, it makes sense to tack on new products, and compute is a potentially large one. Boston Consulting Group estimates that the market for AI compute will climb from $360bn in 2025 to roughly $2.3tn in 2030. If the CME can capture even a fraction of those flows, it will add a lucrative revenue stream.

But just because a market is large doesn’t make it ripe for trading. Research on the origins of futures markets finds that between two-thirds and three-quarters of new contracts fail to attract and sustain a profitable level of trading volume. For compute derivatives to flourish the way oil derivatives have, they must satisfy a number of conditions.

First, underlying prices need to be volatile enough to make hedging worthwhile. Speculators may supply liquidity to a futures market, but it is commercial hedgers that give them their economic purpose. On this test, the case for compute is fairly strong. Long lead times bringing new capacity on stream can make pricing quite volatile.

Oracle’s Project Jupiter data centre build illustrates the point. The facility is due to come online in 2028 but has faced setbacks, ​including delays to a planned natural-gas pipeline and legal challenges over water and air-quality permits. More generally, when chip bottlenecks were tightest at the beginning of 2024, hourly rental rates for one of Nvidia’s H100 chips spiked as high as $8, before falling below $2 in late 2025. That kind of variance is something participants may want to hedge.

Second, participants need to agree on a standardised contract. When oil futures listed in 1983, they were designed to reference West Texas Intermediate, with Cushing, Oklahoma chosen as the physical delivery point. Although WTI accounts for only a small percentage of global crude output, it serves as a pricing benchmark for the rest of the market. In contrast, pioneers of bandwidth trading in 1999 struggled to converge on a standard. Reliability, quality and bandwidth services varied too widely and attempts to fashion a market collapsed with its champion, Enron.

Working with index provider Silicon Data, CME Group has chosen Nvidia’s H100 and B200 chips as the reference for its contracts. According to Silicon Data, capacity on a B200 can be rented for $5.86 per hour; $2.77 on the older H100. CME plans to list contracts that reflect the future value of these rental rates, going out 36 months.

But even a GPU-hour is not fully standardised. Performance varies with cluster configuration, networking, software and location. Rapid obsolescence fragments the market further: an H100-hour is not interchangeable with a B200-hour, which is why CME needs separate contracts for each. Nor is there consensus over how to measure even those rental rates. Another index provider, Ornn, has its own benchmark, and the two do not always align. Until a consensus forms, a derivatives market may be slow to build.

Finally, a vibrant derivatives market requires a diverse group of buyers and sellers. Onion futures were explicitly banned in 1958 in the US after two traders cornered the market. Compute is unlikely to be cornered in quite the same way, but it is already highly concentrated. Nvidia dominates chip supply, hyperscalers control much of the available capacity and a handful of AI laboratories account for much of the demand.

CME Group’s compute futures will attract plenty of attention when they list. The price of compute increasingly bears on the economics of AI, the value of its infrastructure and the vast investment boom surrounding it. This gives investors good reason to want a benchmark for it. Bears may recall the ABX subprime index, which illuminated a previously opaque corner of mortgage finance — and became a focal point for the unwind that followed. Whether it validates the boom or punctures it, compute futures may ultimately matter more as a signal than as a market.