Evan Conrad has an everyday image for the strangest part of selling computer time. "If I have milk and it is about to expire, I will lower the price on it," said Conrad, the CEO of SF Compute. An hour on a GPU cluster can't be sold once it has passed, so capacity nobody has bought gets cheaper as its start time approaches.
Conrad made the point on Latent Space, in an episode published October 20, 2026 and recorded on the sidelines of the Compute Trading track at the AI Engineer conference in New York. SF Compute runs a market where companies buy and resell time on GPU clusters, the banks of graphics processors used to train and run AI models. The conversation skipped the basics. The questions came in derivatives-trading terms, and Conrad answered with an operator's view of how GPU time is priced.
A small market with concentrated trading
Asked how mature, deep or liquid the compute market is, Conrad answered plainly: "not very mature or deep or liquid." A liquid market is one where you can buy or sell quickly without moving the price much. SF Compute first had to bring a lot of hardware online, which he said "just takes a while", and then build the platform to trade it, which he said had only just started working.
About 5% of the company's total capacity now sits on its order book, the running list of buy and sell offers that a matching engine pairs up. Conrad said that share keeps growing, and that all newly added capacity goes onto the book. Trading volume is small, he said, but liquidity is concentrated in that slice.
He tracks it with one number. A customer who spends $100 buying capacity makes back $25 on average by reselling time they don't need. Conrad said he prefers this to "high-minded" measures such as counting market makers because it shows whether the market is "solving the problem" for the customer. The company's September 9, 2026 post reports the same 25% average and works an example: a GPU hour bought for $4.50 ends up costing $3.375 after resale proceeds. The post also describes an automated capacity manager that buys missing capacity, extends nodes already in use and lists spare time for sale.
SF Compute's order documentation shows the mechanics. Prices are quoted per eight-GPU node-hour. Buy orders by default either fill completely right away or are cancelled. Preemptible purchases, which can be cut short, are automatically listed for resale at a specified price multiple.
How far out the contracts go
The answer depends on which part of the market you mean. In the broader market, Conrad said, clusters are financed over three to five years, and payments run over that period. SF Compute's own order book takes blocks of any length: "You can buy right now to the next hour, or you can buy right now to the next three years and anywhere in between," he said, adding "or five years technically."
For now, buyers pay for those blocks entirely upfront. Longer-term leases are handled differently, and Conrad said the cash-flow side of those was something he would need a whiteboard to explain.
The company's recent deals show what the long horizons look like. In September it announced $245 million in sales to an unnamed AI company: two dedicated NVIDIA B300 clusters, each worth $122.5 million over 36 months. The company argues that lenders need durable commitments while customers need a way to adjust them when demand changes. Its settlement ledger, Abel, handles that adjustment. In one example a customer buys 64 nodes from March through November, finishes early and resells September through November.
Paying for compute in advance can weigh heavily on a young company. At the same conference, Jump Trading's compute trading co-leads gave an illustrative example: a company raises $40 million and signs a $100 million, three-year compute contract, and a 30% prepayment leaves it $10 million for everything else.
A market for breaking leases
Asked whether real-time trading or longer horizons interested him more, Conrad picked the longer ones. SF Compute is "trying to build a forwards curve that has a lot of liquidity on it," he said. A forward curve shows today's prices for capacity delivered at different future dates.
What a buyer will pay depends on how painful an interruption would be. A batch inference job, which works through queued requests with no user waiting, can be cut off almost any time. Conrad said such a customer might need an hour's warning, or even a minute or two. A company serving live inference to its own customers is different: "you can interrupt me, but in a month, maybe two months, maybe three months," he said, "but you cannot interrupt me in a minute." A live service that breaks is probably paying penalties under a service-level agreement, the uptime promise it made to its customers. Because the cost of being interrupted differs, he said, the two kinds of capacity are priced differently.
The discussion arrived at a name for this, a market for breaking leases, and Conrad agreed: "These are the market for breaking leases basically." There is also a market for future leases, he added. He compared it to electricity, which has a day-ahead market and a live market. Compute likewise has capacity about to arrive and capacity already in use, and "these have different price dynamics."
A humped curve
Then came a trader's question. Is the market in contango, where future capacity costs more than capacity now, or in backwardation, where it costs less? "Both at different times," Conrad said. "It depends on who you are." Pressed on the shape of the curve, he said: "We see like a hump."
The milk explains the first part. Anything still on the market close to its start time is capacity no one has bought, and since it "can't be sold in the past," its price drops until it clears. Further out, the price rises "to where it wants to be." Beyond that, compute may be worth less again, he said, because people want compute now and pay for the shorter lead time. "This is what I think a normal healthy curve looks like." SF Compute's capacity-manager examples follow the same logic, cutting asking prices on unused reservation time as delivery approaches, though the company presents these as illustrated trading policies rather than a measured market-wide curve.
The B300 gap
What struck Conrad was the gap between what a top chip earns and what it rents for. A B300, one of NVIDIA's newest data-center GPUs, running inference at full use can bring in "$40 an hour or so," he estimated, or about $25 an hour by his rough figure at 50% utilization. "Meanwhile, you can buy B300s at like five to $6 an hour. One of these numbers has to change." These are his round estimates, not published market data.
Carmen Li, presenting Compute Exchange, a marketplace for reserved and future GPU capacity, at the same conference, showed something in a similar direction: her contract-duration curves had moved upward and lost the discounts that used to come with longer commitments. Li also cautioned that the forward rates implied by such curves are market expectations and can be wrong.
The gap produces the other side of Conrad's curve. Buyers who expect prices to rise, he said, may be willing to pay more than they had planned for capacity that comes online later, just to lock it in. "So we see the other side as well," he said. "Kind of depends on who you are."