Affaan Mustafa and Alejandro Soumah Fernandez told AI Engineer why the GPU hour still trades like a commodity in the 1990s. Mustafa, who leads compute trading at Jump Trading alongside Soumah Fernandez, called the current market a dark pool.
He did not mean an exchange. He meant unlit, bilateral, chat-driven deals. Most tenured physical compute trades as a commercial agreement, usually an MSA plus an SLA plus an order form with hardware-specific language, and pricing is argued around the skew, the base board like H100, H200 or B300. There is no widely adopted exchange for physically traded compute, Mustafa said, so quotes move over Slack, X messages and Telegram group chats. Mustafa likened the setup to Ice Chat and Bloomberg Chat in commodities, while public on-demand marketplaces exist only as price references that update daily or faster.
The macro picture explains why that matters. Mustafa pointed to an a16z chart showing the cost per capability falling sharply while total hardware demand rises, a classic Jevons paradox. Token prices have gotten far more efficient on new boards, yet teams burn more tokens for the same budget, so inference demand keeps compute prices robust even as price per million tokens drops. That decoupling sits on top of a heavy upstream stack of chips, power, buildings, grid and capital. Mustafa cited the IEA April update that global data center electricity will rise from 485 terawatt hours in 2025 to 915 terawatt hours in 2030, which pushes colocation, land and grid access straight into the compute quote.
The choke point is financing. Mustafa mapped the chain from chip and system vendors through sites and capacity providers like hyperscalers and neoclouds to users such as inference clouds, labs and enterprises, then showed how lenders underwrite buildouts against contracted revenue. CoreWeave filings were his large-scale example of infrastructure financed by take-or-pay contracts, and CoreWeave has quoted a median prepayment of 15 to 25% on a $100 million total contract value, Mustafa said. For a Series B startup raising $40 million, a 30% prepayment on that $100 million three-year cluster would leave $10 million to run the business. The alternative is turning away customers or bridging the gap with on-demand.
That is where microstructure bites. Every deal is still bespoke, Mustafa said. Spec sheets, bill of materials, delivery dates and even the word memory can mean different things between VRAM, CPU memory or storage, and a price often applies only to that specific cluster. Against that, the market has layered price discovery indexes and benchmarks, physical matching venues and aggregators, bare-metal resellers, desks and brokers who source, negotiate and carry cluster risk for a spread, plus nascent futures and derivatives on the GPU hour. Mustafa said ownership versus rental only works if rental stays below the cost of building and financing your own stack, and fungibility is partial at best, centered on the skew, while the rest trades like a product listing.
Soumah Fernandez closed on policy and the opening for intermediaries. He said export controls now sit directly on this market, noting the US Commerce Department moved H200-class license reviews for China and Macau to case-by-case review with conditions in January. He also tied open-weight models to the US AI Action Plan, which he said explicitly encourages open source and open weight AI and recommends improving the financial market for compute so startups and academics can access scale. The gap is the middle market, both speakers argued, where buyers want smaller, flexible blocks and resale or upgrade rights and suppliers want long commitments to justify buildouts. That mismatch, plus the fact that even listed compute proxies trade thinly, with Hyperscale Data's GPUS quoted with support around $0.12 and resistance around $0.13 as of the latest one-day standard deviation model per Market Chameleon, leaves room for balance sheets uncorrelated to crypto or AI beta to intermediate, but only if the market agrees on a standard capacity contract, common specs and common definitions.
