AI compute is the new oil—and needs a new market

About $360bn was spent on AI compute last year, according to Boston Consulting Group, roughly 50% more than on copper; BCG expects $2.3trn of transactions by 2030, about the size of today’s crude-oil trade. Yet the market is primitive: hyperscalers and frontier labs strike bilateral deals worth tens or hundreds of billions that thousands of smaller firms can’t access, and there are few derivatives to hedge or speculate. Neoclouds such as CoreWeave and Nebius rent capacity for fixed, often multi-year periods, and informal brokers tout capacity on WhatsApp and Telegram. A cluster of startups is trying to build market plumbing: Compute Exchange lists GPUs by location, duration and price; CME Group and Intercontinental Exchange have teamed up with startups building rental-price indices and await regulatory approval for compute futures; Anera Labs arranged a six-figure pilot H100 rental-price trade with FalconX; Luxor offers swap-like contracts guaranteeing fixed cash flow per server-hour (tens of millions of dollars so far); and Architect Financial Technologies, founded by former FTX US president Brett Harrison, offers compute futures on a Bermuda exchange plus a platform for bidding to supply inference. Trading firms DRW and Jump Trading have invested in Silicon Data and built trading teams, and Jump bought Ito Markets this month. Many pioneers come from crypto. The Economist argues compute is too heterogeneous (chips, servers, locations, durations) to trade easily yet, and caveats that long-term contracts are hard to value when counterparties may not survive—CoreWeave’s rising bond yields and double-digit rates for smaller operators show the uncertainty. Jump’s David Olsen says clearing houses are needed and “we’re a long way from that”; many think maturity could take years or decades. Anera’s Vishwa Naik compares the moment to the end of oil’s “seven sisters” cartel.