A hedge is only as good as its fit. If the price an organization hedges moves differently from the price it actually pays, the hedge can protect it less than expected, or occasionally work against it. This gap is called basis risk, and managing it is the first job in designing any hedge.
This article walks through how a CFO or finance team buying AI inference might design a hedge that fits: what to measure, which reference to choose, how long and how large the hedge should be, and how to keep it working as things change.
The airline example
Airlines are a familiar case. Their exposure is to the price of jet fuel, but for many years the deepest markets were in crude oil and heating oil, so airlines often hedged with those instead. Most of the time the prices moved together. When refining margins shifted, they did not, and some hedges protected less than planned.
The lesson is not to avoid hedging. It is to know exactly what the organization is exposed to, choose the reference that tracks it most closely, and understand where the two can drift apart.
What a hedge does, and does not do
A price hedge makes a gain when the price the organization is exposed to rises, and a loss when it falls. Held alongside the underlying cost, it smooths the total: when costs rise, the hedge pays; when costs fall, the hedge costs something, but the underlying bill is lower.
A hedge does not make costs lower. It makes them more predictable. For a company with a budget to meet, a product to price or lenders to satisfy, predictability is often worth more than the chance of paying a little less.
A price hedge also does nothing about volume. If usage doubles because a product succeeds, the bill doubles at any price. That is a business outcome, usually a welcome one, and it is managed through planning and supplier agreements rather than through price hedging. Keeping the two apart is one of the most important steps in designing a hedge.
Step one: name the exposure
For AI, the question for the CFO is what the organization's costs really follow. Two broad cases cover most buyers.
- Paying for a level of capability. The organization runs a defined workload, such as support, document processing or coding, and will move to whichever model does it well at the best price. Its exposure is to the cost of that capability over time.
- Paying what the market pays. The organization uses a broad mix of models, often the newest, and its spend tracks prevailing token prices across the market.
It helps to measure cost per unit of work, such as cost per resolved support ticket or per document processed, rather than total spend. Total spend mixes price and volume together. Cost per unit of work shows how the price of what the organization buys has moved, which is the part a hedge can address.
Step two: choose the reference that tracks it
Carry publishes two indices built for these two cases.
- CICI, the flagship benchmark. It tracks the price of AI inference at constant capability, and fits buyers whose exposure is to a fixed level of capability. See CICI.
- CRTI. It tracks the realized cost of AI inference in US dollars per million tokens, and fits buyers whose costs follow what the market actually pays. See CRTI.
Many organizations sit between the two. In that case it is worth looking at how the organization's own historical costs have moved against each index before deciding.
Three kinds of basis risk
Even with the right reference, a hedge and the cost it protects will not move in perfect step. For AI inference, the gaps tend to come from three places.
- Model basis. An organization's costs depend on the particular models it uses and when it switches between them. If it moves to a cheaper model faster than the market as a whole, its costs fall faster than the index; if it stays on an older model, they may fall more slowly.
- Contract basis. Supplier terms may differ from published rates, through committed-use discounts, fixed prices or tiered pricing. A discount that is fixed for a year means the organization's costs do not follow the market in that year at all.
- Timing basis. Costs and the index may move at different times. A price change may reach the organization's invoices a month or a quarter after it appears in the market.
None of these is a reason not to hedge. They are reasons to measure how closely the organization's costs have followed the index in the past, and to size the hedge accordingly.
Step three: match the term
A hedge should cover the period over which the organization is exposed. If it has committed to a product that will run for eighteen months, a three-month hedge leaves most of the exposure uncovered. If its plans change every quarter, locking in two years may overshoot. The term should match the commitment being protected, and the hedge can be rolled or extended as plans firm up.
A common approach is to hedge in layers: a larger share of the next few months, where usage is well understood, and smaller shares of later periods, adding to them as forecasts become firmer.
Step four: size it
The size of a hedge should reflect how much of the organization's cost moves with the index, not its whole budget. A simple illustration, with hypothetical numbers:
Suppose a team expects to spend $1 million on inference next year, and its past costs have moved about 80% as much as the relevant index. A hedge covering roughly $800,000 of exposure would offset most of the price movement that the index captures, leaving the rest to the team's own choices of models and volumes.
Follow the numbers through. If the index rises by 25%, the team's costs would be expected to rise by about 20%, or $200,000. A hedge on $800,000 of exposure gains 25%, also $200,000, and the budget holds at about $1 million. If the index falls by 25%, costs fall by about $200,000 and the hedge loses about the same. Either way, the team spends close to what it planned.
The 80% figure is called the hedge ratio. It can be estimated by comparing past changes in the team's cost per unit of work with past changes in the index. A higher ratio means costs follow the index closely; a lower one means more of the movement comes from the team's own choices.
In practice, firms rarely hedge all of an exposure. Many start with a portion, often the volume they are most confident of, and adjust as they learn how well the hedge tracks.
Step five: review the fit
Exposures change. Workloads grow, teams move between models and new products launch. The finance team should check periodically how actual costs have moved against the chosen index, and resize or switch the reference if the fit has drifted. A hedge that fitted last year may not fit next year.
Common mistakes
- Hedging volume with a price instrument. A price hedge does not protect against higher usage. Hedging a usage forecast that then falls short leaves a position with nothing behind it.
- Hedging the wrong reference. A hedge on a measure that does not track the organization's costs adds risk rather than removing it.
- Hedging too much, too early. Covering all of an uncertain exposure for a long period leaves little room to adjust as plans change.
- Setting and forgetting. A hedge designed once and never reviewed drifts away from the exposure it was meant to protect.
- Hedging without a policy. Hedging works best inside a written policy, approved by management, that sets what may be hedged, how much, with which instruments and who reviews it. Accounting for hedges also has its own rules, which are worth settling with auditors before the first trade.
Name the exposure, pick the closest reference, match the term, size it to what moves, and check the fit.
Carry Indices exist so that this exercise is possible for intelligence. A trusted, published reference for the cost of AI inference is the first thing a well-fitting hedge needs. How enterprises manage AI inference costs looks at the wider set of tools companies use alongside hedging.
Frequently asked questions
What is basis risk in an AI inference hedge?
It is the risk that the organization's own inference costs and the index used for the hedge move differently. It comes from the particular models the organization uses, the terms of its supplier contracts and the timing of price changes. Measuring how closely past costs have followed the index shows how large it is.
How much of its inference spend should an organization hedge?
Usually less than all of it. A common starting point is the share of spend that has historically moved with the index, applied to the volume the organization is most confident of. Many start with a portion and adjust as they see how well the hedge tracks.
Can a hedge protect against higher usage?
No. A price hedge protects against changes in the price of inference, not against using more of it. Higher usage is managed through planning, budgets per unit of work and supplier agreements.
Which index should be used, CICI or CRTI?
CICI tracks the price of inference at constant capability and fits organizations that buy a defined level of capability from whichever model offers it best. CRTI tracks what the market actually pays per million tokens and fits organizations whose spend follows market prices. Comparing past costs with both is a good way to choose.
How often should a hedge be reviewed?
At least every quarter, and whenever workloads, models or supplier terms change materially. The review checks how closely costs have followed the index and whether the size and term of the hedge still match the exposure.
This article is for information only and is not investment advice. Derivatives involve substantial risk of loss and are not suitable for all investors. The figures above are hypothetical illustrations.