THE TEARDOWN
You cannot store an hour of GPU time. That single fact is why this contract is hard to build.
The Structural Signal
CME Group and Silicon Data announced two compute futures on August 11, 2026. Trading is targeted for October 5, pending CFTC review.
Both are cash settled and listed under NYMEX rules. One tracks the Silicon Data H100 rental index, the other the B200 index.
Each contract stands for roughly one month of GPU rental. The underlying is an hourly rental rate for a specific Nvidia chip.
The CFTC moved in the same week. It sent a draft request for public comment on compute derivatives to the White House budget office in mid-August.
Bloomberg reported it on August 17. The request is pre-rule, and no comment deadline had been published as of late August.
Chairman Michael Selig framed the stakes bluntly. "America cannot win the AI race without a robust derivatives market for compute," he said.
CME is not alone. ICE plans a cash-settled contract on the NATIVX COIL index, expected later this year.
Architect Financial Technologies is building a third venue. It bought IMX Health, an already-designated CFTC contract market, to get there.
The Mechanical Breakdown
Start with what makes this hard.
An unsold hour expires. That kills the cost-of-carry link that anchors the curve in oil, metals, and grain.
Without carry, the forward price is pure expectation. No warehouse arbitrage keeps it honest.
Now the index. Silicon Data collects hourly on-demand rental prices for H100 and B200 chips across providers.
Those prices are not comparable out of the box. Region, network speed, storage, contract length, and uptime all change what an hour is worth.
So the index normalizes them. That normalization is the product, and everything else is packaging.
ICE went further down the same road. Its index measures tokenized compute and connectivity after adjusting capacity for energy use.
Read that carefully. Two rival benchmarks now define one unit of compute in two different ways.
Settlement is the last step. The contract pays against the index average, with no chips delivered anywhere.
Clearing sits at CME. Margin, daily marks, and a clearing house standing between the two sides.
Here is where a model replaces human judgment. In markets with no exchange spot price, analysts assess the number by hand.
Iron ore and LNG work that way. Compute does not.
The assessment is algorithmic instead. A normalization model decides what a scraped price from one provider means relative to another.
The analyst who used to make that call is not in the loop. The model is the benchmark.
Where should you invest $100 right now?
Elon Musk just invented and patented this new AI technology…
And he's predicting it will launch a NEW industry that will grow more than 7 million percent in the coming years.
Even if he's only 10% right, that would still be enough to grow $100 into more than $700,000.
Legacy vs. Autonomous
The legacy hedge is a contract, not an instrument. You sign a long-term capacity deal with a data center and lock the price.
Galaxy Digital did exactly that. It delivered 133 megawatts to CoreWeave under a 15-year agreement in Texas.
Those deals work. They are also bilateral, credit-heavy, and nearly impossible to exit.
The listed version fixes all three. Transferable, cleared, and marked every day.
Where it fails is basis. A buyer's real cost is a reserved multi-year contract with one provider.
The index tracks on-demand spot rates. Those two numbers can drift apart for months.
Liquidity is the second risk. The natural sellers are data centers, and the biggest buyers own their own chips.
Hyperscalers do not need this hedge. That thins the pool to mid-size labs, resellers, and speculators.
Concentration is the third. If a handful of providers set most of the input prices, the benchmark is easy to lean on.
Capital Flow Implications
Three pools open here.
The first is listing and clearing. CME and ICE collect on every contract and every margin call, indefinitely.
The second is the benchmark license. Silicon Data and NATIVX are selling the definition of the asset itself.
That is the durable rent. Index franchises outlive the products built on them.
The third is a new commodities line for banks. Compute joins power, gas, and freight on the structuring desk.
Who loses is quieter. Capacity brokers and hosting deal teams sell scarce information.
A public forward curve removes that. An enterprise negotiating cloud pricing can check a screen instead.
Cloud providers lose a little pricing power for the same reason. Opaque pricing was worth something.
Verdict
The AI industry's core input just became a listed commodity. The venues capturing it are the same two exchanges that already own energy and metals.
Nothing about that is new. Financialization always ends at the clearing house.
Watch October 5. If the date slips, the CFTC consultation is the reason.
Watch open interest more than the launch. A benchmark with no volume is a press release with a ticker.
Territory: +machines on the asset, +incumbents on the venue.
Compute becomes tradable and model-priced, while listing, clearing, and index fees land with the usual exchanges.
