Quick, what's the world's hottest new asset class? How about "compute," the computational or processing power required to create artificial intelligence, which, perhaps predictably, some are calling the new oil or the new crypto.
Naturally, then, compute (yes, increasingly a noun) is getting securitized. CME Group, partnering with Silicon Data, which tracks pricing of graphics processing units, or GPUs, and trading firm DRW, will begin trading compute futures on Oct. 5, pending regulatory approval. A futures contract allows investors to buy or sell an associated asset at a certain time and price, and these securities would allow investors to hedge and speculate on the cost of renting chips that machines need to train and run AI models.
Companies usually rent these chips from cloud providers like Amazon.com's AWS or Microsoft's Azure, or so-called neocloud companies like CoreWeave and Nebius Group, because it reduces costs and allows for flexibility.
Still, the cost to rent these chips has been prone to wild swings. Early rental prices of the Nvidia H100 GPU, the workhorse of the AI economy, launched in mid-2022, run from $7 to $10 per GPU hour. By last fall, prices fell to as low as $1.70. Now, supply has tightened, and prices are up some 40% to about $2.35 per GPU hour. A typical AI data center might have about 100,000 H100s, so that volatile pricing can wreak havoc.
"The amount of money that's going to go into compute is outrageous," says CME CEO Terry Duffy. "I don't know the last time I've seen so much money go into something like this. We think a futures contract is a natural for companies to mitigate and manage risk. There's a lot of people clamoring for this."
Opacity is an issue, too. Carmen Li, CEO of Silicon Data, says that two companies buying the exact same GPU capacity could pay wildly different prices with no way to know who got the better deal. Now, she says, they will have a benchmark to check that against. Silicon Data's benchmarks appeared in SpaceX 's initial-public-0ffering prospectus, which made note of its rental-rate data.
Specifically, Silicon Data tracks prices of Nvidia H100 GPUs as well as Nvidia's newer B200 or Blackwell chips (no easy task), with its indexes based on this pricing. As such, the CME will be offering two compute futures contracts -- Silicon Data H100 and Silicon Data B200. Li believes that these futures will become as common as pork belly contracts, telling CNBC that the value of compute futures will someday exceed that of oil futures.
For those in the hyperscaler/data-center business, these futures might be used to lock in prices of GPUs or hedge their exposure. For sophisticated investors, compute futures might be a way to speculate on a belief that the supply of, or demand for, this processing power is going up or down.
While trading these futures may be more for seasoned traders, plain-vanilla retail investors who are interested should keep an eye out for exchange-traded funds based on compute. ProShares, for example, filed a registration statement for an exchange-traded fund tied to the proposed contracts.
But is compute really an asset class?
"I think so," says Duffy. "Look at the growth of the cryptocurrency markets when we came out with the first futures contract there in 2017. I don't see anything different here when it comes to compute. AI is in everyone's face every day, all day long. Why not have a risk management contract for something that people feel is game-changing for the world?"
So quick -- are compute futures the real deal, or a sign of peak AI? Stay tuned.
Comments