Compute Is Already Trading

Two audiences tend to be surprised by that sentence. People who work in compute every day, who operate fleets, buy capacity, or lend against hardware, and have never thought of any of it as tradable. And people who trade everything else for a living and did not know compute had prices worth trading. This piece is for both.

Figure 1, below, is the starting point. One chip, three prices, and the ordering is the whole story.

Why compute has a curve

Three structural facts produce the term structure, and each one should look familiar to someone from an adjacent market.

Everyone who owns compute has to sell term. Hardware is financed, and lenders size loans to contracted revenue, so operators need to sell one to three years forward to borrow at all. Very few buyers want to commit that long. The result is a persistent imbalance. Term capacity is offered at a discount to short because the sellers are structural and the buyers are scarce. When a market pays you to take the other side of a structural flow, that payment has a name. It is a risk premium, and risk premiums attract capital.

Compute cannot be stored. An unsold GPU hour expires at the end of the hour. You cannot buy cheap capacity today, warehouse it, and deliver it next year, which means no cash-and-carry arbitrage ties the forward price to the spot price. The curve is expectation plus risk premium, with no arbitrage anchor underneath it. Traders from power markets will recognize this immediately. Traders from oil should be careful assuming their intuitions transfer.

Prices jump rather than drift. Hardware does not lose value smoothly the way a truck does. It loses value on event dates, when a new chip ships or a new model changes what workloads need. Anyone who has traded single-name CDS will recognize the structure immediately. The risk is jump to obsolescence rather than jump to default. The number underneath everything is residual value rather than recovery, and in both markets it is the assumption everyone prices and nobody can verify until the event. Even the calendar rhymes. Chip releases are telegraphed the way maturity walls are telegraphed, so the date is known and the outcome is not, which is precisely what makes the date tradable. A GPU can go from premium to marginal in a quarter on a single release, the same way a bond can trade at par into a refinancing everyone saw coming and gap the day it fails.

Put those together and you get an asset that borrows one trait from each market and belongs to none of them. It carries like credit, expires like power, and gaps like a single name, and the combination is its own asset class. Markets like that do not stay untraded for long. This one already hasn’t.

The same map, side by side

Who is in the market

Before the trades, the participants. Every one of them has a position today. Most of them acquired it by accident.

The physical trade

The first trade in any young market is the simple one, and in compute it is tenor transformation.

As an illustration, multi-year blocks of current-generation capacity have been clearing in the high threes per GPU-hour, while one-year commitments on comparable hardware clear in the mid fives and shorter reserved blocks higher still. Buy the three-year, sell the one-year against it, keep reselling the tail as it shortens. The spread is not free money. It is payment for holding three risks the short buyer refuses to hold. The capacity can sit idle between contracts, the re-let price can come in lower than expected, and a release date can make the hardware obsolete before the term ends.

Figure 2, below, lays out the shape. The first year is sold and the spread is locked. The back two years are the position, and the dashed row is where the cash-settled market enters, a short on the index against the unsold tail.

This trade is on right now. Firms are buying multi-year capacity and reselling it shorter, running the book the way a desk runs any carry position. We broker this flow at Liquid Compute, and blocks are trading. What the trade currently requires is operational reach, real relationships with data centers and a sales effort to place the short end. What we are building removes that requirement. Standardized contracts that transfer mean a position can be entered and exited by assignment rather than by owning a rack relationship and a sales team. That is the difference between a trade available to operators and a trade available to capital, and it is where this market is headed.

The second physical trade is block liquidity. Sellers of capacity often need speed more than they need the last nickel, and a buyer who can price a block quickly gets paid for that service. Every OTC market rewards the participant who can say a number fast. Compute is no exception, and today almost nobody can.

The cash-settled trade

The second channel needs no racks at all. These are contracts that settle in cash against a reference price for compute.

This is what our index exists for. Liquid Compute maintains a reference price for compute, licensed today to institutional counterparties, built from the ground up to be settled against. The shapes that trade against it are the ones every OTC market converges on.

Outright. Express a view on where forward compute prices are going, long or short, without ever touching hardware. The financing pressure described above gives the curve a persistent shape, and persistent shapes invite opinions. In June 2025 Blackwell Ultra volume was confirmed for the second half of the year and wafer supply was moving toward it. LCI-H100-US printed $2.95 at the end of that month. Anyone who believed new supply reprices the old generation was a seller, and by the last week of December the index printed $1.84, down 38 percent. Then the same fact pattern inverted. Wafers leaving Hopper meant legacy supply stopped growing just as inference demand accelerated, and by late December on-demand capacity was effectively sold out across the market. The index bottomed that week and printed $2.57 on August 29, up 40 percent from the low. The same chip paid the short in 2025 and the long in 2026, and both entries were public information at the time. The market was writing the H100’s obituary at the exact moment the contract market was tightening, which is also the argument for having a reference price at all.

Relative value. One index against another. Chip generations do not move together, and release dates are the catalyst that reprices the spread between them. A new generation shipping in volume changes what the previous generation is used for, and the ratio between the two moves accordingly. Anyone who has traded credit curves or calendar spreads will recognize the structure of the trade immediately. At the end of December 2025, LCI-B300-US printed $4.23 against LCI-B200-US at $3.81, a spread of roughly 40 cents between two chips separated mainly by memory. Reasoning and agentic workloads were pushing inference toward models that need the larger part, and the 288GB tier was where the new wafer supply was going. By late August the B300 printed $5.97 against $4.97, and the spread was a dollar. Both legs rallied, which is the point. The trade never needed a view on the direction of compute, only on the ordering of the two chips.

Hedges. This is where the inquiry actually comes from, so be specific about who has the risk. An operator with unsold forward capacity is long future compute prices whether they like it or not, and can lock revenue on the unsold portion. A lender secured by hardware is long residual value, the single assumption in every compute credit that nobody can currently verify, and a hedge against it changes how much they can advance. An enterprise or AI company that knows it must buy compute again at renewal is short future prices and can cap the cost. None of these participants is speculating. Each is reducing a position they already hold.

That last point is the reframe this market needs. If you operate a fleet, buy capacity on contract, or lend against hardware, you already have a position on this curve. You did not choose to put the trade on, but it is on. The only question is whether you have any way to manage it.

Over the past months we have fielded inquiry on exactly these structures, referencing our index, from dealers, market makers, proprietary trading firms, lenders, and corporates. The interest is not hypothetical and it is not confined to the financial side. Some of the most natural users are the people closest to the hardware.

The basis

Where a physical and a cash-settled market coexist, a third trade appears between them, and it is what makes the first two liquid.

A corporate bond desk will bid a block of bonds below the market because it can hedge with CDS the moment it gets hit, then work out of the position at its own pace. The hedge is what lets the desk provide liquidity at size without betting the book on direction. The same mechanic works in compute. A physical desk that can lay off price risk in the cash-settled market can bid a block of GPUs aggressively, in either direction, and get paid for providing immediacy rather than for guessing right. Divergences between the physical and cash-settled price become tradable in themselves, and every basis trader who steps in tightens the relationship between the two markets.

This is the quiet reason the physical and derivative markets need each other. Neither becomes deep alone.

Why this matters even if you never put on a trade

It would be easy to read this as a sideshow for financial tourists. The opposite is true, and the logic runs through financing.

A lender who can hedge residual value can advance more against the same fleet. An operator who can lock forward prices on unsold capacity can sign the long lease a data center requires. A buyer whose exit is assignment rather than breach will sign three years instead of one. Every one of those changes shows up in the physical market as more capacity financed, on better terms, with longer commitments.

Where we come in

We came at this market from every side of it. Our team has traded and structured credit, sold compute at the platforms that define it, and built the hardware underneath it, and that mix is deliberate. A market like this needs people who have priced risk, people who have carried a book of buyers, and people who have racked the machines.

We are building the infrastructure this piece describes at Liquid Compute. Physical brokerage is live and blocks are trading. Our reference index is licensed to institutional counterparties, with live inquiry from both the financial and the physical side of the market. We are building toward cash-settled contracts referencing term compute pricing, and the data behind the index already tracks the spreads and relationships described above.

If you operate or own capacity and want to see where term paper is actually clearing, write to us. If you lend against hardware and want to talk about residual value, write to us. And if you trade for a living and this looked familiar, that is because it is. The instruments are new. The trades are not.