How to Control Your AI Compute Budget
AI compute prices are notoriously volatile. What if there were a way to insure yourself against the volatility?
This post is a collaboration with Ornn, a startup that is building an exchange for AI compute. They structured a compute swap with two counterparties exposed to AI compute price fluctuations. What follows explains what was done, why it matters, and what it unlocks.
This is written for people who are already accountable for AI compute costs or revenues, but don’t yet have a standard way to manage the price risk. Speculators and arbitrageurs might find it too high-level and hand-wavy.
Most teams buy AI compute the way you buy plane tickets during the holidays: you refresh a dashboard, watch prices drift, and hope you don’t get punished for waiting a week.
That works when compute is a rounding error. It stops working when compute is the business. If you’re managing an AI compute budget, your problem isn’t simply that compute is expensive. The problem is compute is unpredictable. This is a volatility problem, and the industry has no standard way to reduce that uncertainty.
Ornn’s compute swap is an attempt to fix that. It’s not a new scheduling system or a new GPU marketplace. It’s price insurance for compute, with the important caveat that it doesn’t pay in GPU time. It pays in cash. This is key: you don’t have to change where you buy your GPU time; you only get a cash payment designed to adjust your effective cost.
The core idea in one sentence
You can lock in an effective compute price for the next month by agreeing on a fixed $/GPU-hour price, then settling the difference in cash against a benchmark index.
This contract, essentially a Compute Swap, separates your need for the physical resource (GPU time) from your exposure to its price fluctuation.
If the market price goes up, the hedge pays you cash to offset your higher compute bill. If the market price goes down, you pay cash—but your compute is cheaper anyway, so your total outlay ends up more stable.
Think of it like a fixed rate mortgage vs a variable interest one. You’re trading uncertainty for predictability.
What Ornn actually built
Ornn describes their first compute swap as:
A cash-settled contract (money settles; nothing is physically delivered).
Based on the USD price of compute quoted as $/GPU-hour.
The benchmark is their read-only OCPI index (for specific GPU SKUs like H100 SXM).
The contract runs for 30 days.
The floating price is the average of the index’s daily values over the month.
The position updates daily (marked to market daily) and money moves daily to keep it safe.
Ornn’s memo describes the same “average-over-the-period” settlement choice explicitly: at expiration, settlement uses the average of the index value over the duration of the contract. That averaging detail is important: compute can’t be stored like oil. It flows like electricity.
A toy model
Let’s say you expect to use 10 GPUs continuously for the next month. A month is ~720 hours, so that’s 10 GPUs * 720 hours = 7,200 GPU-hours.
Now imagine you want price certainty. You enter a 30-day hedge:
You lock $1.80 per GPU-hour
At the end of the month, the benchmark index says the average market price over the month was something other than $1.80
Case A: compute got more expensive
Average market price over the month: $1.86
Average market price over the month: $1.86
Difference: $1.86 - $1.80 = $0.06 per GPU-hour
Cash settlement: $0.06 * 7,200 = $432 paid to you
That $432 is designed to offset the fact that your actual compute bill went up.
Case B: compute got cheaper
Average market price: $1.74
Difference: $1.74 - $1.80 = -$0.06
Cash settlement: -$0.06 * 7,200 = $432 you pay
But your compute bill was cheaper during the month. The hedge did its job: it reduced variability.
This is budgeting equipment, not a money-making machine.
“But how is this safe?”
When you do a cash-settled hedge like this, the obvious fear is: What if the other side can’t pay when I’m owed money?
The standard answer, in all serious markets, is collateral. Ornn describes a setup where each side puts some cash up front (”initial margin”), and then money moves daily as prices move (”variation margin”), with margin calls if balances fall too low.
You don’t need to become a derivatives trader to understand the intuition:
If the price moves against you today, you have to pay today, not at the end of the month.
That reduces the risk of blowing up at the end of the month.
In their memo, they also mention a third-party custodian structure intended to reduce counterparty risk.
Who would use this?
AI teams: stabilize costs so you can plan
If compute is your biggest variable expense, volatility leaks into everything:
hiring plans
product pricing
SLA commitments
delivery timelines
A hedge lets you say: “I don’t know exactly what compute will cost next month, but I can make it approximately fixed.
Data center operators: stabilize revenue so you can build
If you’re selling compute, your risk is the opposite: prices can fall.
Hedging can convert “future GPU revenue” from a volatile guess into something closer to a predictable stream—useful for underwriting builds and financing.
The non-obvious group: lenders and big buyers with long commitments
Ornn’s memo points out two interesting seller types:
Financiers shorting alongside borrowers to protect downside compute prices and improve credit profiles.
Large AI companies with long-term commitments shorting to protect the economics of excess capacity they can’t resell.
This is a big deal because it means compute hedging isn’t just “for traders.” It’s a tool for making large compute spend and large compute infrastructure financeable.
Why the futures curve matters
So far we’ve talked about hedging one month at a time. That’s the entry point. But for operators and investors, the real information content isn’t the spot price or the next 30 days. It’s the shape of expected prices over time. This is what commodities people would call the forward curve, but what you can think of more simply as a time-indexed price forecast that the market is willing to collateralize.
If you’re a data center operator, the curve answers questions your spreadsheets already care about: What do I expect GPU-hours to clear for six months from now? Twelve? Two years? Not in a vibes-based way, and not via a single point forecast, but as a continuous schedule of prices that can be referenced, hedged, and stress-tested. That matters because your cost structure is front-loaded and your revenue is spread out. Power, land, and GPUs are paid for now; compute revenue shows up later. The curve is the bridge between those two timelines.
This is where profitability forecasting improves. Instead of underwriting a build on a single assumed utilization rate and a hand-wavy price deck, you can anchor revenue assumptions to observable forward prices. If the curve slopes down, you know the market expects compression, and you plan accordingly: tighter leverage, faster amortization, shorter contract tenors. If it’s sloping upwards, you might accept more near-term volatility in exchange for longer-dated economics. Either way, future price forecasts no longer rely on a simple extrapolation from the current spot price.
The important thing is that none of this requires operators to become traders. You don’t need to arbitrage contango or debate roll yield. You just need a market-generated answer to a basic question: What does the world think compute will cost over time, and how confident is it? Once that exists, swaps and futures function like they do in mature infrastructure markets: planning instruments that make long-term capital deployment possible.
The important caveat: the benchmark is the whole game
A cash-settled hedge only works if the benchmark reflects the prices you actually face. The biggest residual risk here is basis risk: your real-world compute costs might not move exactly like the OCPI index.
Ornn says they’re building OCPI from executed spot transactions and performance benchmarking, and that regional benchmarks are coming (US-East, EU-West, etc.). That direction is correct, because region and quality differences are not rounding errors in compute markets.
If the benchmark becomes trusted and representative, hedging becomes practical at scale. If it doesn’t, you get a product that looks good on paper but doesn’t reliably offset real bills.
What this unlocks, if it works
If this swap instrument is successful, then:
AI teams can commit to longer-term customer pricing with less fear.
Operators can underwrite builds with less reliance on market vibes.
Lenders can finance infrastructure with less exposure to price swings.
It’s not glamorous. It’s foundational. Serious markets are mostly boring plumbing that enables big, non-boring things to happen. If Ornn’s instrument works, the industry can finally stop buying compute like holiday plane tickets and start building on a stable foundation.
If you enjoy this newsletter, consider sharing it with a colleague.
I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:
Solid breakdown of compute hedging mechanics. The futures curve angle is where this gets really intresting—data centers need forward visibility to justify capex, and right now they're basically flying blind. I saw a friend's team shelve a build last quarter becuase they couldn't underwrite revenue past 6 months. If Ornn's benchmark gets adoption, that's a genuine unlock, not just trading theater.