Wall Street’s Biggest Exchanges Are Racing to Turn AI Compute Into a Commodity
ICE and CME both want to trade GPU hours like crude oil. The analogy is more revealing than its promoters intend.
In May, within days of each other, the two largest futures exchanges in the world announced plans to list contracts on the price of AI compute. ICE, which owns the New York Stock Exchange, partnered with a startup called Ornn, whose Compute Price Index, or OCPI, tracks the open-market cost of renting Nvidia’s H100, H200, B200, and RTX 5090 chips. CME Group had announced its own compute futures with a rival index provider, Silicon Data, the week before. A smaller venue, Architect Financial Technologies, had already said in January it would launch futures on GPU and memory prices using Ornn’s data. Kalshi will let you bet on Nvidia rental prices today. More recently, Ornn raised $33 million led by Andreessen Horowitz and says the first compute swap against its cleared prices was executed in December.
The pitch, repeated almost word for word across every announcement, is that compute has become a trillion-dollar market without the pricing and risk-transfer tools every other major commodity relies on. CME’s chief executive Terry Duffy went further: “Compute is the new oil of the 21st century.” The comparison is deliberate. ICE’s Brent contract and CME’s WTI contract emerged from a similar race between exchanges in the 1980s and went on to define the global reference price for crude.
The oil analogy seems correct, but perhaps not in the way the exchanges mean it.
The case for the contracts
Start with the strongest version of the case. GPU rental prices swing violently. Ornn’s own index showed the rental price for Nvidia’s Blackwell chips rising 48 percent between mid-February and mid-April, from $2.75 to $4.08 per GPU-hour. For a lab budgeting a training run in the tens of millions of dollars, that kind of swing can destroy a plan mid-execution. Data center operators carry the mirror-image risk: they borrow billions against future rental revenue whose price they cannot lock in. Lenders financing the buildout are exposed to both.
A futures contract solves this in principle. We know the rough shape of these instruments from oil and wheat: you fix a price today for something delivered later. What’s worth working out is what that machinery means when the underlying thing is compute. These particular contracts are cash-settled, meaning nobody ships a GPU anywhere; when the contract expires, the two sides simply exchange the difference between the agreed price and the index price. Which puts enormous weight on the index. It is the number the entire market settles against, so the question worth sitting with is not whether you can trade compute but who constructs the number.
Ornn’s answer is that its index is built only from completed trades, not surveys or dealer quotes. That is a genuine virtue, and a sign the industry has learned from LIBOR, the interest rate benchmark that quietly governed hundreds of trillions of dollars in contracts until it turned out to be manipulable. LIBOR was built from bank estimates, and estimates are cheap to shade. Real trades are harder to fake. But an index built on real trades is only as sturdy as the market it samples, and the open market for GPU hours is a small one. Most compute never touches it: the bulk of capacity is locked up in long-term private deals between the big cloud providers and the largest AI labs, invisible to any index.
What the commodity frame conceals
Commodity markets work because the underlying good is fungible and storable. A barrel of Brent-grade crude is interchangeable with another barrel; if prices drop, you can leave it in the ground or park it in a tank in Rotterdam. Compute is neither of these things.
An H100-hour is not an H100-hour. The same chip delivers wildly different effective performance depending on interconnect, cluster topology, networking, and the software stack around it. A GPU-hour on a tightly integrated cluster built for frontier training is a different product from the same nominal GPU-hour on a fragmented marketplace of loose capacity. The index has to flatten this heterogeneity to produce a single price, which means the benchmark describes a stylized commodity that the actual buyers of compute, the people the hedging tools are supposedly for, do not purchase.
And compute cannot be stored. An idle GPU-hour is gone. There is no tank farm for unused capacity, no way to buy cheap and warehouse it for a tighter market. Storability is what anchors futures prices to physical reality. Without it, the price of future compute floats free, tethered to expectations alone.
Then there is the problem no other commodity has: the reference asset is decaying on a schedule set by a single company. Every Nvidia generation reprices the last one. An index of H100 rental prices is structurally a melting ice cube, and any contract built on it embeds the market’s guess about Nvidia’s release schedule as much as any supply-demand fundamental. Trading compute futures is, in meaningful part, trading Nvidia’s roadmap. That is not a commodity market. That is a bet on the strategic decisions of one firm, wearing a commodity market’s clothes.
Benchmarks make markets, they don’t just measure them
There is a well-developed literature in the social studies of finance on how financial models and benchmarks don’t passively describe markets but actively constitute them. Donald MacKenzie put it in the title of his 2006 book: finance theory is “an engine, not a camera.” It doesn’t photograph the market. It drives it. An index becomes the price. Contracts settle against it, loan terms get written around it, procurement gets justified by it, and the messy underlying reality reorganizes itself around the number.
This is the part of the story that matters most and gets discussed least. Once the OCPI prints on a Bloomberg Terminal every day, “the market price of compute” exists as a social fact. It becomes citable. A government procuring sovereign compute capacity can point to it. A cloud provider defending its pricing to a regulator can point to it. A subsidy program can peg its disbursements to it. The number carries an aura of neutrality because it emerges from a market process rather than a boardroom.
But look at what generates the number: a small open market trading leftover capacity, sitting on top of a supply structure controlled by one chipmaker, a handful of hyperscale buyers, and long-term contracts nobody outside the negotiation ever sees. The benchmark takes concentrated market power as an input and emits an objective-looking price as an output.
Who gets to hedge
The democratization story says these instruments open compute price risk to everyone. Small AI labs can finally lock in training costs the way airlines lock in jet fuel. In principle, true. In practice, using futures markets at scale requires things small players don’t have: a relationship with a clearing member (the intermediary firms that guarantee trades on an exchange), the balance sheet to post margin (the collateral a position requires), and the cash reserves to survive the position moving against you. When prices swing hard, the exchange demands more collateral on short notice; fail to produce it and your position gets closed for you, at a loss, at the moment of maximum volatility. A hyperscaler’s treasury desk absorbs that without blinking. A twenty-person lab can get forced out of its own insurance policy by the very turbulence it bought the insurance against.
The entities that have all of this are the hyperscalers, the largest specialist GPU clouds, and the trading firms. The trading world is not arriving late, either: Silicon Data, the index provider behind CME’s contracts, is backed by DRW, one of the world’s largest trading houses, whose founder expects compute to “become the largest commodity in the world.” These are also the players that already lock up most global capacity through multi-year private purchase deals before it ever reaches the open market. For them, a futures market is not a replacement for privileged access but a second instrument layered on top of it: they can hedge their buildout risk, play their private contract prices against the public benchmark, and, as the largest participants in the small open market the index samples, lean on the number itself.
Brent and WTI created a financial layer whose returns accrued to the majors and to trading houses like Vitol and Glencore, firms whose edge came from seeing physical flows the paper market couldn’t. Everyone else got price transparency, which is not nothing, but transparency about a price you cannot influence and can barely afford to hedge is a consolation prize.
The fourth layer
I’ve been mapping how control over AI consolidates through infrastructure layers: the hyperscale compute layer, the CDN and edge layer that decides who gets scraped, the data layer being enclosed through licensing deals. Financialization is not a separate story. It is what happens to a choke point once it matures: the control becomes liquid. You can price it, lend against it, securitize it, and, crucially, defend it with a market’s legitimacy rather than a monopolist’s.
None of this means the contracts shouldn’t exist, and there’s a certain inevitability to them. If the futures work, some mid-sized players will genuinely benefit at the margin. But the question to ask about any new market infrastructure is not whether it is useful. It is who designed it, who settles against it, and whose position it consolidates. Compute is being turned into a commodity by the same actors who spent the last three years making sure it would never trade like one.
The 1980s oil race gave us a world where a benchmark price set in a financial market governed the physical economy of energy for forty years. The compute race is a bid for the same prize. Watch the index methodology documents, not the launch announcements. That is where the next forty years are being written.