The Practitioner's Guide to the Compute Future: How to Hedge, Step by Step

A 50-GPU company pays about $105,000 a month for H100 rental at today’s index. It knows this month’s bill and nothing about next June’s. Two NYMEX contracts, GPU1 and GPU2, are built to answer that question.

They are not trading yet. CME targeted October 5; the CFTC extended its review by 45 days on September 21, and the listing date on the exchange notice now reads TBD, no earlier than November 9. CME has filed a liquidity provider program that starts November 10 or on approval, so market makers are lined up to quote both sides from the first session.

What follows is how a company that has never traded a future would use one the week it lists.

The contract, in five facts

The exchange notice sets out two contracts: H100 Rental Index Futures, code GPU1, and B200 Rental Index Futures, code GPU2. Five things matter for a user.

One contract is 730 GPU-hours. That is 365 days times 24 hours divided by 12, the hours in an average month, so one contract is one GPU for one month. Sizing a hedge is counting GPUs.

Prices are quoted in dollars and cents per GPU-hour, the same unit as your invoice and the same unit as the index. The smallest move is one cent; a tick is $7.30 of P&L per contract.

Settlement is financial. No GPU changes hands; money does.

Trading in a contract ends on the last business day of its month, and 36 monthly contracts are listed at a time, so day one gives you a curve three years deep. Stack the monthly contracts for the months you need, and the stack is your hedge.

The exchange charge is $5.50 per contract to trade and $1.35 at cash settlement, non-member rates. Six hundred contracts, the size of the hedge below, costs $4,110 in exchange fees, 0.38% of the strip’s value. Your broker adds its own commission.

How it settles, and why an average

GPU1 for a given month settles to the arithmetic average of the index across every business day of that month, not the level on one fixed day.

The reason is that compute behaves like electricity, not like oil. A GPU-hour is consumed the hour it exists; it cannot be stored and delivered later. Your bill is not one price on one day, it is the sum of every day’s rate across the month, and a contract that settles on the month’s average tracks that bill. It is also much harder for one odd print, or one thin quote panel, to move a monthly average than a single settlement day, which is why the PJM Western Hub monthly power contract averages peak-hour prices over the month the same way.

Bandi and Su at Johns Hopkins, in their early asset-pricing framework for compute, make the same case for a first-generation compute contract, and name the tradeoff: averaging cuts the pull of a single print, and it widens basis risk for a user whose daily exposure runs near expiry.

Hedging a year of H100 rental step-by-step

Take a company renting 50 H100s through 2027 from a neocloud provider that bills the index plus $0.12. With the index at $2.88, its all-in rate is $3.00/hour. The company wants 2027 fixed and does not want to sign another three-year capacity commitment to get it.

It buys 50 contracts for each month from January to December 2027, 600 in all.

The contracts have not traded, so there is no curve to quote. The prices below are illustrative, chosen for arithmetic: a ladder running from $2.70 for January to $2.20 for December, averaging $2.45 against a $2.88 spot. If the curve that prints on day one looks nothing like this ladder, substitute your own numbers into the formulas; every step is shown.

Two things then happen every month of 2027. The provider bills the company at the index plus $0.12. The January contract settles at the January average of the index, and the company receives, or pays, the difference between that average and $2.70, times 730, times 50.

Here is the year under two paths for the index, one climbing from $2.85 in January to $3.45 in December, one falling to $1.75.

The net is identical to the dollar: the strip, $2.45 times 730 times 600, which is $1,073,100, plus the $0.12 basis on every hour, which is $52,560. Unhedged, the same year ranged $372,300 between those two paths. Hedged, it is one number, known in advance.

The second column is the part people skip. There the company pays $65,700 into the futures market and ends the year $65,700 worse off than if it had done nothing. That is not a flaw in the contract; it is the price of the certainty in the first column. A company that needs the upside more than it fears the downside should not hedge.

If CME lists options on these futures, a cap gets you part of both: the floor of the hedge, and the first slice of any fall in price, for a premium paid up front.

What happens day to day

Futures settle every day, not just at expiration. Buy the January contract at $2.70 and watch it settle at $2.73, $2.71, $2.76, $2.74, $2.79 over five sessions. On 50 contracts the account is credited $1,095, debited $730, credited $1,825, debited $730, credited $1,825. The running total is $3,285, which is the move from $2.70 to $2.79, nine cents, times 730 times 50.

Those flows are variation margin. They sum to the difference between final settlement and your entry price, so the daily accounting changes timing, not outcome. What it does change is cash: a run of debits in a month when the index falls has to be met with money in the account. Before trading, the exchange also requires an initial margin deposit per contract, set and published by the exchange and returned when you close.

The other side: the provider sells what you buy

A neocloud with capacity to lease runs the identical math in reverse. It fears falling rates, so it sells the futures a buyer would buy. Selling those 600 contracts locks its 2027 revenue to the same $2.45 strip: if the index falls to $1.75, rental income drops and the short position pays $65,700 of it back. If the index climbs to $3.45, the provider hands over $306,600 of the gain to the clearing house. Lenders underwriting GPU-backed debt ask for exactly this: revenue fixed for the life of the loan, not a bet on the spot market.

Where the hedge still fails

A cash-settled compute future removes price uncertainty. It does not remove these.

Basis risk

The worked example holds the provider’s spread at $0.12 for twelve months. That is an assumption, not a contract term. The index tracks the aggregate neocloud market; your provider prices its own region, network, and customer mix on top of it. If your spread moves to $0.32 in July while the index stays flat, the futures payout does not see the extra $0.20/hour, which is $43,800 across the second half of the year on 50 GPUs. Hedging fixes the index, not your supplier’s markup.

The hyperscaler mismatch

GPU1 and GPU2 settle against neocloud on-demand rates. If your workloads run on AWS, Azure, or Google Cloud, you are not buying the thing being hedged: hyperscaler pricing bundles compliance, redundancy, and support, runs higher, and moves less. A neocloud index hedge covers the direction and none of the premium.

Execution friction

Your cost of entry is the spread. At $2.69 bid and $2.71 offered around a $2.70 fair value, lifting the offer on 50 January contracts costs a cent, $365. In a wide market, $2.64 bid and $2.76 offered, the same market order gives up twelve cents, $4,380 on that one month. Use the block trade facility, which starts at five contracts with a 15-minute reporting window, and work limit orders with your clearing broker instead of crossing the screen.

Index quality

The delay matters here. The extension was reported as a review of novel or complex issues, with manipulation risk in the underlying rental market the stated concern. A 45-day review is not a finding of fault, and the exchange has an incentive to get the index right. Until the fix has a track record, size the hedge to the part of your exposure you must fix, not the whole book.

Conclusion and up next

Before these contracts, locking in compute meant a provider’s reserved agreement, with its own term, deposit, and hardware, or nothing. The future adds a third route that is separate from the hardware: rent wherever you like, fix the price on the index, and pay for the difference in cash. Which route is cheaper depends on how the reserve market and the futures curve price the same months, and that comparison is the thing to run before you sign either. Part 2 takes up how to read the curve and where fair value sits in it.