Hedging Compute Prices
A short primer for anyone who buys, sells or finances compute
If you buy compute, own GPUs, or lend against them, you are exposed to the price of compute whether or not you intended to be. This explains what a hedge is, how the ones quoted today work, and what they do and do not fix. No derivatives experience assumed.
Which side are you on
| Long · hurt when prices fallYou want to lock in a price you can sell at | Short · hurt when prices riseYou want to lock in a price you can buy at |
|---|---|
| Operators with unsold capacityHours not yet under contract get sold at whatever the market pays then | Enterprises and AI companiesYou will need capacity again at renewal and do not know what it will cost |
| Lessors and owners at lease endYou own the hardware at maturity and its value tracks the rental price | Buyers who cannot commit longShort contracts mean repricing often, with budget set a year ahead |
| Anyone whose offtake is shorter than their financingThe uncovered years get re-let at an unknown price while payments stay fixed | Anyone growing into more computeYour future cost base is unhedged and rises with the market |
Two sides of the same price. One does better when compute gets cheaper, the other when it gets dearer.
Nobody chose these positions. They are the shape of the business. Traders call one side long and the other short, which only means one of you does better when prices rise and the other does better when they fall. The question is not whether you carry the exposure. It is whether you want to keep it.
What a hedge actually is
A contract that fixes a price for a future period.
Nobody delivers a GPU. Nobody moves a rack. Nothing about your operations changes. At the end of the period you compare the price you agreed to the price the market actually printed, and one side pays the other the difference in cash. If prices fell, the seller gets paid, making up the lower revenue they earned. If prices rose, the buyer gets paid, offsetting the higher cost they paid. Both end up where they agreed. That is the entire product.
Case one. An enterprise fixing what compute will cost
The simplest case, and the one most people recognise immediately.
Your current capacity contract expires in twelve months. You know roughly what you will need after that, but not what it will cost, and you are being asked to put a number in next year’s budget today.
Say you will need 128 GPUs for a year, about 1.12 million GPU-hours. At today’s price of $4.40 an hour that is $4.93M. If compute prices rise 25 percent before your renewal, the same capacity costs $6.16M. You are $1.2M over budget through no decision of your own, and the usual response is to serve fewer requests, throttle users, or push a launch. The product suffers because of a price move.
You buy the forward at $4.40. Twelve months later you go and buy your capacity in the normal way, from whoever you like, on whatever terms you negotiate. If prices went up, the hedge pays you the difference. If they went down, you pay it. Either way your all-in cost is $4.40, and you can plan capacity around what your users need rather than around what you can afford that quarter.
Your annual compute bill on 1.12 million GPU-hours, unhedged against hedged at $4.40.
Three things people get nervous about, and none of them are true.
You are not committing to buy from anyone. The hedge is separate from your procurement. You still run your RFP, still pick your provider, still negotiate. The hedge only fixes the price level.
You are not taking delivery of anything. No GPUs arrive. It settles in cash against a published index.
You are not speculating. You already have this exposure. Buying the forward removes it. Doing nothing is the position with more risk, not less.
The board version is short. We fixed next year’s compute cost. It is a line in the budget instead of a range, and we are not rationing capacity to defend a number.
Case two. A neocloud selling forward to lower its cost of capital
The mirror image, and the reason to do it is financing rather than a view on the market.
You have 256 GPUs coming off contract in six months with no buyer lined up. That is about 1.12 million GPU-hours you will sell at whatever the market pays then, roughly $4.9M at today’s six-month price of $4.40.
To a lender, those hours are worth very little. They are a forecast about your own sales performance, not contracted revenue, and no credit committee sizes a facility against a forecast. So that capacity either gets no advance at all or a punitive one.
You sell the forward at $4.40. The revenue is now fixed. You still market the capacity, still sign whatever customers you find, still run the business the same way. But the cash flow from those hours is a number rather than a range, and that is the form a lender can lend against.
| If the market ends up at | You earn selling capacity | The hedge pays you | Total |
|---|---|---|---|
| $3.80 | $4.26M | +$0.67M | $4.93M |
| $4.40 | $4.93M | nothing | $4.93M |
| $5.00 | $5.60M | −$0.67M | $4.93M |
Revenue against plan, in $M, against where the market is when you re-let, in $ per GPU-hour.
The blue line is your business unhedged, moving with the market. The orange line is the hedge, moving the opposite way. Together they give the flat green line. You have not made or lost money. You have converted uncertain revenue into contracted revenue, which is what changes the advance rate and the spread. In most cases that is worth more than the upside you gave away.
Case three. Sale-leaseback and the residual
What a sale-leaseback is. You buy hardware and immediately sell it to a lessor, who leases it straight back to you. The purchase price comes back as cash, they own the asset, you make monthly payments, and at the end you buy it back for a nominal amount and own it outright. A financing structure rather than a sale, used to take capex off the balance sheet without losing the equipment.
Where it goes wrong. The lease runs three years because that is what makes the payments work. Your customer signs for two, because that is as far ahead as they will commit.
Three years of lease payments against two years of contracted offtake, and the twelve months in between.
Where the risk sits. In month 25 the customer contract has ended and you still owe twelve months of lease payments. You re-let at whatever compute is renting for then, and at month 36 you own hardware worth whatever it is worth then. Both are the same underlying price, and today both are guesses.
How you handle it. Two ways, and most people use both.
Roll it. The market trades well out to about twelve months, so when month 24 arrives the 25 to 36 window is a twelve-month trade and you fix it then. You are not exposed for three years, only until the window comes into range.
Or buy protection now. A put on the year-three period pays you if prices fall below your strike and costs you nothing if they rise. You pay a premium for it, and because a dealer prices the option rather than a forward, it can be quoted today even where outright forwards are still thin.
Combining them. Sell a near-dated forward on capacity you already have, where liquidity is best, and put the proceeds toward the premium on the far-dated put. The liquid part of the curve funds protection on the part that is not. One trade, two legs, and it is how operators in other perishable markets bridged this gap before their long end filled in.
Why this helps your financing. Nobody can prove what a used GPU fleet is worth, because no one has repossessed and re-let one at scale, so lenders assume the worst. A residual you can point to a market for is a different conversation from one you can only model, and that shows up in the advance rate.
What a hedge does not do
It does not fill your racks or guarantee you find capacity. It fixes price, not volume. If your capacity sits idle the hedge still settles. Hedge the volume you are confident about, not your full theoretical capacity.
It is not a perfect match. Your actual price and the index will differ somewhat, because of chip, region, contract length and service level. A hedge removes most of the price move, not every dollar of it.
It is not free. A forward costs nothing upfront but you give up the move in your favour. If you want protection while keeping the upside, that is an option, and options have a premium.
What you would need
An ISDA with the counterparty, which is the standard master agreement for these contracts, a licence for the index it settles against, and a decision on how much of your exposure you want fixed. Most people start with a slice rather than the whole book, which is the right instinct.
Talk to us
Everything here is simplified to make the mechanics clear. Real positions have more moving parts, and the right structure depends on your book, your tenor and what your lender needs to see.
If you buy compute, send what you expect to need and when, and we will show you what fixing that cost looks like today.
If you own or operate capacity, tell us what is uncontracted and over what period, and we will price it and show what it does to your financing.
If you lend against fleets, we will walk through how a hedged residual changes what you can advance and where the numbers come from.
If you are just curious how any of it works, ask. We would rather answer questions than send documents, and there is no cost or obligation.