A Compute Provider's Guide to Arbitrage: When Reserve Contracts and GPU Futures Disagree (Part 2)

Updated September 8, 2026. Term-structure values as of September 7, 2026.

Part 1 showed how an AI company can fix next year's GPU bill with the new NYMEX compute futures. This part answers the question that comes next: you can also fix a rate by signing a reserved contract with a provider, so which one is cheaper, and what does it mean when the two markets disagree about the same months?

Two prices for the same months

A 6-month reserve and a 12-month reserve both cover months one to six. Only the longer one covers months seven to twelve. So the difference between what the two contracts cost in total is what the market is charging for months seven to twelve, and dividing by six gives it as a rate.

On September 7, Silicon Data's term structure for the H100 put the 6-month term rate at $2.52 an hour and the 12-month at $2.30, against a spot index of $2.63.

So this is equivalent to months seven to twelve being charged at $2.08 an hour. That is the reserve-implied forward rate: the price for a six-month rental starting six months out, bootstrapped from the physical reserve contracts, sits 55 cents below spot, a "backwardation" term structure in commodity lingo.

The hypothetical futures market quotes the same six months directly: one contract per month, each settling on that month's average index. Buying all six and renting on-demand in those months gives a fixed rate equal to their average. Suppose the six contracts average $2.25. (Futures prices are hypothetical until listing.)

So for months seven to twelve the physical market says $2.08 and the financial market says $2.25. They disagree by $0.17 an hour.

What a gap is worth, and who can close it

On one GPU over the six months, $0.17 times 730 hours times 6 is about $745. On a hundred GPUs it is $74,460. That is the size of the disagreement before anyone touches it.

When the futures strip sits above the reserve-implied forward, a potential arbitrage opportunity involves executing two legs.

Leg 1

• Setup: buy the cheaper physical capacity for months seven to twelve, which in practice means taking the 12-month reserve and using or subletting the first six months.
• At the end of month 6: sublet the capacity at the on-demand price.
• Revenue: index − $2.08

Leg 2

• Setup: sell the six futures contracts at the strip.
• Revenue: $2.25 − index

Note that the futures pay out the difference between the strip and the index, because by construction the strip converges to the index at expiration. In the final P&L, the two index legs cancel, and what remains is the $0.17.

The natural person to do this is a provider, who already owns capacity and prices it across terms every day, or a large buyer who can commit to the longer reserve and has a use for the early months.

When the strip sits below the reserve-implied forward, the trade runs the other way: sell reserved capacity forward at the rich physical rate, cover it by renting spot during the interval, and buy the cheap futures strip.

Bandi and Su call this the hard direction, and it is worth understanding why. The reserve you sold is for specific hardware in a specific place with a specific provider. There is no guarantee that you will be able to rent the specified capacity at the index price during months 7 to 12 to fulfill the original 12-month rental you sold. This is not only the financial basis risk described in Part 1 but also a risk of contractual breach. Only participants owning large compute capacity are positioned to execute the arbitrage in this direction.

In frictionless markets, arbitrage would quickly close any gap between reserve-implied forwards and the futures strip. But in practice the trade is asymmetric: one direction requires access to specific physical capacity, balance sheet, and the ability to take delivery and perform, while the other is much easier to express financially. Because the physical leg carries residual risks (basis, operational constraints, and even breach or fulfillment risk), only a limited set of participants can enforce convergence. As a result, a persistent spread can remain and does not necessarily represent a free arbitrage.

Implication for users, beyond arbitrage

For an AI company that is not in the business of subletting GPUs, none of this needs to be executed. It needs to be read. The comparison is a decision rule: for the months you actually need, put the reserve-implied rate next to the futures strip plus the basis you expect against your own provider, and take the cheaper route. In the example, a builder who expects to pay index plus $0.15 gets $2.40 through the futures and $2.08 through the reserve, and signs the reserve. If instead the strip printed at $2.00, the futures route would cost $2.15 against the reserve's $2.08, and the seven cents would be the price of guaranteed hardware, the premium that the asymmetry above lets persist.

What this does and does not establish

The arithmetic establishes how to read the two markets side by side: adjacent term rates imply a forward rate, a strip of monthly futures implies a term rate, and the difference between them is a number you can compute the day the contracts list.

Bootstrapping a forward price from term rates is standard practice from fixed-income to commodity desks. A common misconception is that the forward price is the market's prediction of where the price will be in the future. It is not. While it does provide some indication, there is often a gap between a crowd-sourced prediction and a term rate bootstrapped forward. A persistent discount or premium can be embedded in the forward rate due to various frictions and opportunity cost.

The current analysis does not establish what that number will be, which direction it will lean at launch, or how much of it is the access premium rather than mispricing; those are empirical questions the first months of trading will answer. Nor does it establish that any particular buyer should prefer one route, since the answer depends on basis against a specific provider and on whether guaranteed access is worth paying for in that buyer's plans. The Term-Implied Forward Curve publishes the term structure and the implied forward rates this comparison starts from; the futures will supply the other side from the launch.

References

• Bandi, F., & Su, J. (2026). (Early) AI Compute Asset Pricing (arXiv:2607.12156). https://arxiv.org/abs/2607.12156

This article is an illustration of pricing mechanics, not a recommendation to trade. Futures involve leverage and daily cash requirements; contract terms are subject to the NYMEX rulebook and pending CFTC review. Prices for contracts that have not yet listed are hypothetical; term rates are from Silicon Data's published term structure as of September 7, 2026. Silicon Data is the index sponsor; the indices are benchmarks, and the futures are the instruments that settle against them.