Intelligence · Compute · Framework
How AI companies hedge GPU compute costs
· Michael Mescher, Gammon Capital
Direct answer
An AI company hedges GPU compute cost by fixing or capping the price of part of its expected future usage, using reserved-capacity contracts, physical forwards, cash-settled futures on a compute index, or options that cap the price it pays. The goal is not to predict GPU prices; it is to make the compute budget survive a price spike without starving training or inference. A sound program hedges a defined share of forecast usage, staggers maturities, sets a written price limit, and measures the gap between the index and the price it actually pays. As of October 2026, reserved capacity is the only deep instrument; listed GPU futures await CFTC review, and caps or options are bilateral and bespoke.
Key takeaways
- Compute is the largest variable cost for most AI labs and inference platforms, and its price moves.
- Hedge a share of forecast usage, not all of it. Usage forecasts are themselves uncertain.
- Reserved capacity fixes price and supply; financial hedges fix price only. Most programs need both.
- Options and price caps protect the budget without locking in a loss if prices fall.
- Write the price limit down before the market moves. A limit set during a spike is a guess.
Why hedge at all
A training run or an inference business is budgeted in GPU hours. If the price per GPU hour rises sharply, the company either spends more than planned, cuts usage, or delays a model. Each of those is a business decision forced by a market move. Hedging moves that decision earlier, to a time the company chooses.
The five-step framework
1. Forecast usage in bands. Separate committed usage (production inference, signed training runs) from probable and speculative usage. Only the committed band is a candidate for a hard fix.
2. Set a written price limit. Decide the highest price per GPU hour the business can absorb for each band, and get it approved by whoever owns the budget. This becomes the trigger for every hedge decision afterwards.
3. Choose the instrument per band. Committed usage: reserved capacity or physical forwards, because the company needs the hours, not just the price. Probable usage: options or price caps, which protect the budget without obligating purchase. Speculative usage: usually unhedged.
4. Stagger maturities. Layer hedges across quarters rather than fixing a full year at one price. Staggering averages the entry price and leaves room to adjust as forecasts change.
5. Measure basis and counterparty. Track the gap between the index a financial hedge settles on and the price the company actually pays its provider. Track the credit of every counterparty, including the provider holding a prepayment.
Instruments compared
| Instrument | Fixes price | Secures supply | Downside if prices fall | Available today (Oct 2026) |
|---|---|---|---|---|
| Reserved capacity contract | Yes | Yes | Pays above market for the term | Yes, widely |
| Physical forward | Yes | Yes | Pays above market for the window | Bilateral only |
| Cash-settled future | Yes, against the index | No | Pays the difference in cash | Listed contracts pending CFTC review |
| Call option or price cap | Caps it | No (unless physical) | Premium only | Bespoke over the counter only |
Example
Illustrative numbers only. An inference platform forecasts 10 million GPU hours next year: 6 million committed, 3 million probable, 1 million speculative. Its written limit is $2.50 per GPU hour. It reserves the committed 6 million in quarterly tranches with two providers at an average of $2.10, buys caps struck at $2.50 on the probable 3 million for a premium of $0.08 per hour, and leaves the last million floating. If prices spike, the cost of 9 of the 10 million hours is bounded at or below $2.58 all-in. If prices fall, it overpays on the reserved tranches, which it needed for supply anyway, and has spent the cap premium: about $240,000 on 3 million hours.
Common mistakes
Hedging 100% of a forecast. If usage comes in lower, the company owns a position it does not need, and a hedge becomes a speculation.
Confusing supply with price. A cash-settled future pays out when prices rise but does not deliver a single GPU hour. In a shortage, the company still has to find capacity.
Prepaying a weak counterparty. A large prepayment to a smaller provider is an unsecured loan to that provider. Size it as credit risk, not as a procurement term.
Frequently asked questions
Can I hedge GPU costs with futures today?
Not on a US exchange as of October 2026; listed GPU futures are pending CFTC review. Kalshi's event contracts on GPU prices are live but small and pay fixed amounts, so they are a poor fit for budget hedging. Today the practical tools are reserved capacity, bilateral forwards, and bespoke caps.
Is a reserved GPU contract a derivative for accounting purposes?
Usually not. Most reserved-capacity agreements are service contracts without net settlement, so they generally fall outside derivative accounting under ASC 815. A contract that settles in cash against an index may be a derivative unless an exception applies. A reserved contract for identified, dedicated GPUs may contain a lease under ASC 842. Confirm the treatment with your auditors before signing.
How much of forecast compute should be hedged?
There is no universal number. A common structure fixes committed usage, caps probable usage, and leaves speculative usage floating, with the split set by how reliable the forecast is.
Gammon Capital view
Buyers are the missing half of the compute derivatives market. Capacity owners have a reason to sell forward; buyers have a stronger reason to buy protection, and almost nowhere deep enough to buy it. A buyer with a written price limit and a staggered program is a better counterparty and gets better terms. See the compute practice.
Framework, not implementation plan. This page describes how Gammon Capital approaches the topic. It is a general framework; engagement work is specific to each client's contracts and balance sheet. Market facts are stated as of the date shown above.
Canonical citation. Gammon Capital (gammoncap.com), Michael Mescher.
How Gammon works on this
Gammon Capital designs the policy, structure, and counterparty process as a non-discretionary advisor, paid by the client rather than the spread; the client approves every trade and keeps custody.
Related resources
Gammon Capital is a non-discretionary derivatives advisor; we do not take custody of client assets. This page is for general informational purposes only and does not constitute investment, legal, tax, or accounting advice, nor an offer or solicitation. Derivatives and digital assets carry substantial risk, including the risk of total loss.