The WTI of Compute Doesn't Exist Yet

Last week I told you the traders in the compute market are going to be agents, and that the household is the last adopter of governed AI, not the first. I stand by both. But I skipped a step. The market those agents are going to trade in doesn't have a price yet.

Compute is becoming a commodity. You already knew that. What you might not know is how unsettled the commodity still is.

The News: Four Exchanges, One Price Nobody Agrees On

Oil has WTI. Gold has the London fix. Wheat has a Chicago contract that has been clearing trades since the 1870s. Every mature commodity has one reference price the whole market argues around. Compute has at least four references, they are all a few weeks old, and not one of them has cleared a single exchange-listed trade.

Run the tape from the last two months. In May, CME Group and Silicon Data, the market-data firm backed by the trading house DRW, announced a compute futures contract, pending CFTC review. A week later, on May 19, ICE announced GPU-hour futures referencing Ornn's Compute Price Index. In June, Ornn raised $33 million from a16z crypto and Galaxy to keep building that index, which already runs on the Bloomberg Terminal. On July 1, ICE announced a second compute futures product, this one with a firm called NATIVX, referencing an energy-normalized index named COIL. Then on July 8, Architect Financial Technologies announced a network that converts a compute futures position into actual delivered GPU capacity.

Read that again. ICE is running two different compute benchmarks at the same time. The exchange placing the bets is hedging against its own product.

That is not a market that has decided what it is. That is a land grab.

The headlines skip the part that matters. As of today, none of it is trading. CME's contract is pending. Both ICE contracts are pending. Architect's U.S. exchange is pending. The only thing actually live is Ornn's spot marketplace, running under a de minimis exemption while the bigger venues wait in the CFTC's line. The benchmark war is real. The futures market it is fighting over is still a stack of press releases.

Why Nobody Agrees on the Price

Part of why nobody agrees on the price is that compute isn't one thing. An H100 hour is not a B300 hour. A GPU sitting next to cheap hydro power in Washington is not the same asset as one drawing on a strained Texas grid in August. Goldman Sachs put a number on the scale last month: $7.6 trillion of AI infrastructure spending through 2031, of which the compute layer alone is roughly $5.1 trillion. Goldman's own framing is that the binding constraint isn't the chips or the money. It's power. That is why NATIVX normalizes its index for energy, why Ornn tracks specific chip classes, and why the two don't line up. They are measuring different things and both calling it compute.

The demand underneath is not theoretical. Public bitcoin miners have announced more than $70 billion in AI and high-performance-computing contracts, turning their power and their shells into compute supply. TeraWulf signed a $19 billion data-center lease with Anthropic this month. Real money is moving at spot today. The futures market that would let anyone hedge that exposure is the piece still stuck in the regulatory queue.

The Assumption I Had Backwards

When I started writing about agents in commodity markets, I assumed the market would settle first and the agents would arrive after. A benchmark would win, a contract would clear, liquidity would pool in one venue, and automated desks would plug into a finished thing. That is how I would have drawn it up.

It is happening in the opposite order. The algorithmic desks, the ones that behave like agents whether or not you call them that, are showing up in a market that has not standardized. Four indices. Several venues. Settlement rails that don't match, some cash, some physical, some pointed at crypto. No single price to anchor to.

I used to think an unsettled market made the governance question premature. You can't govern trades in a market that barely exists. I now think that's exactly backwards. The mess is the reason the authorization layer matters sooner, not later.

What an Agent Actually Has to Do

Think about what an agent actually has to do to trade compute right now. It has to choose which benchmark to price against, because there is no default. It has to pick a venue, because there is no consolidated tape. It has to settle across rails that don't agree. Every one of those is a decision a human handed to it, under some mandate, with some limit, on someone's behalf.

In a settled market you can lean on the market's own machinery to catch a bad trade. There's a clearinghouse, a reference price, a paper trail everyone recognizes. In a market with four benchmarks and no clearing, that machinery doesn't exist yet. The only thing that travels with the trade is the authorization the agent carried into it. Accords governs what agents do on the open internet.

This is what that sentence means in practice. Not a compliance report you assemble after the quarter closes. A cryptographic record, created at the moment of the trade, of who authorized the agent, for what, within what limits, and on whose behalf. AccordsVault produces that record, and it doesn't care which index wins the benchmark war. Whichever one does, an agent will consume it, and someone will have to prove the agent was allowed to.

Why This Isn't a Someday Problem

The regulators are already circling the general shape of this. In late June the CFTC and SEC jointly asked for public comment on harmonizing their rules. There's a CFTC proposed rule out on event contracts. None of it names compute futures yet. It will. And when a regulated venue finally lists a contract and an automated desk starts trading it at machine speed, the first question an examiner asks won't be about the price. It will be who told the machine it could do that.

The Falsifiable Close

Here's what would prove me wrong. If the CFTC clears one contract in the next quarter and the market coalesces fast around a single benchmark, the way equities settled around a handful of reference indices, then standardization beats agent adoption to the line and the market's own plumbing absorbs a lot of what I'm describing. The urgency softens. Two things worth watching: the first real CFTC clearance, and whether ICE picks one of its two horses instead of riding both.

I'd bet the other way. Standardization is slow and political. Agents are fast and indifferent to politics. This market is going to get automated participants before it gets a settled price, and the distance between those two dates is exactly where the authorization layer earns its keep.

Next week I want to walk through the strangest piece of all of this, the one Architect just announced: the moment a paper compute hedge turns into physical GPUs changing hands. That seam is where the whole thing stops being abstract.

Max