算力期货市场将启，能否为AI经济定价？

Oliver Kemp for Transformer

Earlier this year Oxford University’s St Hugh’s College needed to update the computer lab used by its math department, but it found the price had nearly doubled. While for the college’s administrators that posed an annoying but manageable IT bill, for one of the university’s professors, it was a symbol of a far bigger challenge facing anyone sourcing computing power.

Rama Cont, an Oxford professor of mathematical finance and adviser to central bankers on financial risks, blames the soaring costs on the data centers powering models such as ChatGPT and Claude. “This is affecting everyone. All good universities, all corporations, are going to be affected this year,” Cont says. But while it’s a mostly manageable budget problem for many institutions and businesses, for the companies building and financing the AI economy, it could be an existential risk.

McKinsey estimates that globally, data centers will require nearly $7t in capital by 2030, including $5.2t for AI. The cost of that infrastructure, roughly the price of 22 Apollo space programs in today’s dollars, must be financed at the same time that the world figures out what its output — so-called compute — will eventually be worth.

Cut through the noise.

A handful of startups and exchanges now want to create markets trading in future compute usage, allowing data center operators to set clearer prices, and lenders to better value the assets they are financing. If they work, they could bring the transparency and price discovery that such futures — effectively a tradable promise to sell or buy something in the future for a set price — have given oil, wheat and electricity. If they don’t, they risk creating a new channel for losses to flow through, connecting an already highly leveraged AI sector more closely to the wider financial system.

BlackRock chief executive Larry Fink, whose investment firm is pouring billions into data center construction, predicted this spring that the demand for compute is so large that traders will soon be able to trade bets on it as its own asset class. In fact, those markets could launch within weeks, pending regulator approval. CME Group, one of the world’s largest trading markets for futures is partnering with tech startup Silicon Data on plans to start a market in compute futures this year. New York Stock Exchange owner Intercontinental Exchange is preparing a rival in partnership with Ornn, another tech startup. Others in Shanghai and elsewhere are working on them too.

Brett Harrison hopes to beat them all. The former president of FTX’s US business — he left shortly before the crypto exchange blew up — runs Architect Financial Technologies. It plans futures tied to the price of renting GPUs on the open market. “I do believe this will be a multiple-trillion-dollar market,” Harrison says.

What you are actually buying

Futures are basically a legal agreement between a buyer and a seller which allows the two parties to lock in the price of something. A farmer, for instance, can offer wheat at a set price months before it’s ready, “hedging” against the possibility of a bountiful harvest causing an abundant crop and prices to plummet.

Marketplace operators have often created new futures markets for niche products. One of the most famous agricultural futures markets on CME was for pork bellies. It traded from 1961 until 2011, but was delisted as bacon became a year-round consumer good with more predictable prices, instead of a seasonal food which needed to be frozen in the winter. There are futures markets for all sorts of goods, from orange juice (the subject of 1983 comedy film Trading Places) to sunflower seeds, all doing the job of evening out volatile prices for farmers and buyers of goods.

Today’s compute “farmers” are the hyperscalers such as Alphabet, Amazon, Oracle and Microsoft, and the numerous “neoclouds,” companies which build data centers and rent out AI chips to others. AI futures markets, in theory, will allow neoclouds to protect themselves if the price of computing power crashes in the future. It could also allow AI companies worried about rising costs to protect against the price soaring higher than expected. No chips actually change hands in this purely financial transaction.

To build their data centers, many neoclouds borrow money against the future revenue they can extract from renting servers, which contain cards like Nvidia’s Blackwell chips. They, and their lenders, have to consider the risks to the value of their assets and their revenue. For instance, if a much better chip arrives sooner than expected making the ones they have installed prematurely obsolete, or if a more efficient AI training or inference algorithm makes the same work much cheaper, that will reduce the price the neocloud can charge.

Without the ability to hedge that risk, Harrison says, neoclouds have become “their own broker, their own insurance company, their own hedger, their own financier and their own pricer,” as well as a data center construction company. A forward price which everyone agrees on, and can trade, could lower their financing costs by giving lenders a way to value future revenue.

That hedge requires someone to take the other side, effectively betting that future value of the commodity, in this case compute, will be higher or lower than what it’s priced at now. “There have to be enough people willing to bet against you,” says Frank Partnoy, a Berkeley law professor and former financial product designer for Morgan Stanley who chronicled the collapse of Enron. That means future markets must welcome speculators who have nothing to do with data centers, and who are simply trying to make money. “That seems to be what all of them are talking about. No one’s talking really about speculation, although if you create the market, that will happen, right? You create the market, inevitably people will use it to bet on these prices,” says Partnoy.

The trouble with pricing an AI cloud

Before you get to predicting supply and demand for computing power, there is the problem of defining what “compute” is and how to compare one unit of it to another. An oil future specifies a grade of crude, a place and a delivery date. Compute has no settled equivalent.

Carl Anthony, the co-founder of Symmetric Research, a startup which aims to create a “standard compute unit” so that banks can better understand their investments in data centers, says you need to look at the overall “computing fabric” of data centers to truly understand their value. That fabric includes memory, networking, cooling, location, software and workload, and whether those thousands of chips can communicate efficiently while training a model or carrying out an AI inference task. Paper specifications like FLOPs, a standard unit of computing power, or even standardized stress test benchmarks of GPUs, don’t give you a true picture of their real-world capabilities, he says. Research by Silicon Data, CME Group’s partner, found that the performance of rented Nvidia GPUs used in AI data centers can vary as much as 38% in real-world conditions.

Architect’s compute futures get around this by setting prices against an index created by its partner Compute Desk, which uses private rental transactions from cloud providers. It adjusts the details for configuration and location, and normalizes them to an anchor such as an hourly rate for a GPU rental in one year’s time. It also looks at the offer prices for computing contracts, and produces what Harrison calls a “smooth, averaged index.”

The ongoing construction of Meta’s massive AI data center in El Paso, Texas. Credit: Brandon Bell/Getty Images

Harrison admits his index will not be a 100% accurate picture of every GPU rental at any particular time, similar to, he says, futures markets for other commodities. That still leaves traders some risk to manage.