Paper compute

On October 5, a trader in Chicago who has never trained a model, never signed a power contract, never waited a week for a GPU allocation to clear, starts helping set the price of the thing every AI company runs on. He does not need to know what a B200 does. He needs to know whether the number goes up or down, and he needs someone on the other side of the bet. That is what CME Group’s new compute futures market does. The two contracts it launches that day, one tracking H100 rental prices and one tracking B200, each stand in for roughly a month of GPU rent. They take the cost of compute out of the builders’ hands and give it to the market. The price stops being a thing you negotiate with a data center. It becomes a thing strangers bet for and against, and every day those strangers answer, in public, the question the industry has spent two years dodging: is the buildout scarce and worth it, or oversupplied and a bubble? The builders no longer get to decide that. The pit does.

Compute is making a crossing older commodities made long before it. First it was something you bought, an input, a line on an invoice. Then it became something you held, an asset with a price that moves whether or not you use it. Now it becomes something you bet on, a financial instrument that trades on its own, detached from any particular server in any particular building. Every commodity that made this crossing was sold the same way, as transparency and as a hedge, and for the largest players it delivered exactly that. For everyone downstream, the price of their own work turned into a football.

The pattern repeats with the loyalty of a law. Wheat farmers learned that a room full of traders in Chicago could move the price of a harvest before it left the field. Oil producers watched paper contracts vastly outnumber the physical barrels, until the price of a real thing pumped out of the ground swung on the mood of people who would never touch it. And in 2008, once it was traders rather than lenders or the families in the houses who set the price of housing risk, the people actually living there became collateral in a game they had never agreed to enter. You won’t find it in a single number. You’ll find it in the repetition.

This time it’s easy to read, because the winners are named. CME collects a fee on every contract that trades, whoever turns out to be right. That’s the oldest and safest position in any market: the house. But the sharpest detail sits one layer down. The benchmark the entire market prices against, the reference index the contracts settle to, is owned by a company called Silicon Data, and Silicon Data is backed by DRW, a Chicago proprietary trading firm. The people who built the ruler are the people best placed to profit from what it measures. Everything else here is downstream of that fact.

And the exposed are just as easy to name. The last two years produced hundreds of “neocloud” startups whose whole business is the spread between what they pay for GPUs and what they charge to rent them out. That spread was always thin and always private. Now it is a public number that can move against them in an afternoon, on a rumor, with the loss visible to everyone, including their lenders. Below them sit the small labs and solo builders, the ones too small to staff a desk of people whose job is to hedge. Their single largest cost becomes a speculative instrument, whipsawed by traders arguing about whether AI is a bubble, an argument that has nothing to do with the model they are trying to ship this quarter.

The case for all of this is strong, and I’ll state it at full force. Every real commodity has futures, from jet fuel to electricity. Airlines hedge the price of fuel, and passengers get steadier fares as a result. A regulated exchange is a plainly better thing than the market it replaces, where companies, in CME’s own words, “have often paid vastly different prices for the same computing capacity with no way to compare deals.” That is not a small problem. Rent an H100 today, and the hourly price runs from a couple of dollars to more than ten depending only on where you happen to click. A public reference price is a genuine public good, and pretending otherwise would be dishonest.

But the 2008 instruments were also sold as spreading risk and making everyone safer, and they were, right up until they weren’t. The hedge isn’t the danger. The danger is what happens once a deep, liquid, speculative layer settles on top of a real input. The paper price stops merely reflecting the physical one and starts to drive it. A wave of traders acting on an AI-bubble story can bend the curve, and a bent curve reprices real companies’ costs and their access to credit, no matter what is actually happening on the floor of the data centers. Transparency for the whales who can hedge is exposure for the minnows who cannot. Price discovery is a public good and a loaded gun at the same time, and which one you experience depends entirely on your size. The question worth asking is who is left holding the risk on the day the pit decides the whole buildout was a bubble. The answer has been the same every time before. Never the exchange. Never the trader.

I’ve already made the case that compute became an energy business, a fight over who owns the megawatts and pours the concrete. This is the next stage in the same commodity’s life, and it’s a different fight: not who owns the megawatts, but who owns the price of them.

The buildout is either scarce and worth it or oversupplied and a bubble. From October 5, traders bet that argument up and down every day. DRW owns the index they settle against. CME collects on every trade. The labs and neoclouds renting the GPUs pay whatever number the traders land on, right or wrong.