Renting Nvidia’s B300 chips now costs more than $24 an hour in some deals, and short-term rentals of under a year are running above $7, according to Eugene Ye, a trader who writes about compute markets on X. In a September 27 post, he said the number is stale by the time a reader finishes the sentence.
That speed is the problem Ye is trying to solve. Companies that rent out GPU fleets, often called neoclouds, frequently sell customers a fixed monthly price while their own chip costs float with the market. Every one of those contracts is a bet on where GPU prices go next, whether the operator meant to make that bet or not.
Ye’s proposed fix borrows directly from commodities trading: forward contracts and options written against a GPU rental price index, so an operator can lock in a ceiling price without committing to buy or reserve chips years in advance. He points to early signs the market is forming. FalconX, a crypto-focused trading firm, has executed what it calls the first over-the-counter swap tied to an H100 chip rental index. Wintermute, a market maker, has done a similar H100 forward trade. The Commodity Futures Trading Commission extended its review of a proposed compute futures contract on the New York Mercantile Exchange through November 9, per a filing dated September 21.
To show how the hedge would work, Ye built an illustrative example. A neocloud has already promised a video generation customer a flat monthly rate, and that customer might need 2,048 B300 chips within a year once its own launch is confirmed. Rather than reserve the chips outright, the neocloud instead buys a call option today, one that caps what it would owe if rental prices spike before the customer’s launch window arrives. He is explicit that the prices, volatility assumptions and dollar figures in the example are hypothetical, built to demonstrate the mechanics rather than to describe an actual trade.
Ye also ran his own simulation comparing purchasing strategies across 4,000 hypothetical five-year price paths. By his account, renewing GPU capacity annually without any hedge averaged $4.72 per hour, with the worst 5 percent of outcomes averaging $10.91. Adding the option hedge raised the average to $4.88 but cut the worst-case average to $7.20. Those are Ye’s own model outputs, not figures from an exchange or a counted set of executed trades.
The comparison that matters here is to how other volatile commodities matured. Oil, natural gas and electricity all went from spot-only markets to markets with liquid forward curves and options once enough buyers needed to hedge and enough dealers were willing to price the risk. Compute is following the same path, just years behind, and the CFTC’s ongoing review of a formal futures contract suggests regulators are treating that shift as more than a niche trading experiment.
For any company running an inference business on rented GPUs, the practical question Ye raises is not whether to hedge but what it costs to keep the option to change fleet size open. Operators locking in multi-year GPU reservations without pricing that flexibility against emerging forward and options markets are making a financial bet, whether their pitch deck calls it one or not.
Reported by Eugene Ye, a trader who writes about compute markets on X, in a post published September 27, 2026.
