Compute Futures

Anyone developing large AI programs today needs enormous amounts of computing time on specialized machines. In the industry, this computing time is called “compute.” It is scarce, expensive, and its price fluctuates sharply. A compute future is a contract in which two parties agree today that a certain amount of computing time will be delivered and paid for at a certain price in a specific month. The contract itself can be resold, even by people who have no intention of using the computing time at all. Such contracts for future delivery are generally called futures or forward contracts on exchanges.

Compute time becomes a commodity with price risk

A company that wants to train an AI model often plans months in advance. It roughly knows how much computing time it will need next year. But it doesn’t know what that computing time will cost by then. If the price doubles, an entire project can become unaffordable. This exact risk can be hedged with a future: the price is fixed, no matter what happens in the market.

The other side has a mirror-image problem. Anyone building a data center invests billions before the first customer pays. If the operator sells computing time in advance for the years after completion, they secure guaranteed revenue. Banks are then more willing to finance such construction projects, because the future earnings can be demonstrated.

A third effect is interesting. Once such contracts are publicly traded, a visible market price for computing power emerges. Today, most contracts are negotiated individually and in secret. A futures market would show what the market expects computing time to cost in two years. This is also a signal for investors and journalists about just how strong the AI boom really is.

From contract to delivered compute hour

For something to be tradable on an exchange, it must be standardized. With wheat, you specify variety, quantity, and quality. With computing power, this is more difficult. You have to define which chip type is meant, how many hours it runs, how fast the chips are interconnected, and in which region the data center is located. A contract might read, for example: 1,000 hours on a specific graphics chip model, delivered in the third quarter, in Europe.

At the end of the contract term, there are two options. Either computing time is actually provided, which is called physical settlement. Or only money changes hands: if the current market price is above the agreed price, the buyer receives the difference. This second variant is called cash settlement and is the norm in financial markets.

One detail sets compute apart from wheat: computing time cannot be stored. An unused hour on a chip is irrevocably lost, similar to an empty seat on a plane that has already taken off. That’s why compute futures resemble electricity futures more than commodity futures. Electricity, too, must be consumed at the very moment it is generated.

Who trades them and where the problems still lie

In the news, compute futures usually come up in connection with data centers and chip manufacturers. Large cloud providers already sell computing time today in multi-year contracts that function economically like forward deals. In addition, there are young trading venues and exchange operators looking to set up standardized contracts on graphics chip hours. A regulated mass market like the one for oil does not yet exist.

A common misconception is confusing compute futures with buying shares in a chip manufacturer. A share is a stake in a company. A future refers directly to the commodity itself, in this case compute hours. The share price can rise while the price of computing time falls.

The biggest open problem is rapid obsolescence. A chip that is cutting-edge today can seem significantly slower than the new generation in two years. A contract for old hardware then loses value rapidly. Critics also warn that such markets could attract speculation and inflate prices further. Proponents counter that visible prices, especially during a boom phase, actually provide more realism.