Trading AI Compute: The Next Trillion Dollar Market?

The 60 Second Read

"I actually believe a new asset class will be buying futures of compute." That is BlackRock CEO Larry Fink, and his case is shortage: the United States is short power, compute, chips, and memory. The capital agrees. AI capital spending passed oil and gas in 2026, $765B projected against an estimated $681B, and the 2031 estimates widen the gap to $1,636B against $831B. This Deep Dive asks whether compute becomes the world's largest commodity, what unit trades, and who wins once the price is public. The obstacle is fungibility. Across 13 providers, the identical Nvidia H100 rents for $1.79 to $10 an hour, 2025 published rates, and one study found performance gaps of as much as 34.5% chip to chip. A barrel of WTI crude collapses to two numbers, density and sulfur; no two measurements do that for a GPU-hour. Silicon Data and Ornn publish GPU indices on Bloomberg, and the market forming around them points to electricity's regional, cash-settled shape. The first contracts, on Silicon Data's indices, are dated: October 5, 2026, pending regulatory review. Neoclouds, the specialist clouds that buy GPUs to rent out, carry the debt and need this market most. From launch day, volume decides.

Key Takeaways

• Takeaway 01

  AI capex passed oil and gas

  AI capital spending is projected at $765B for 2026 against an estimated $681B for oil and gas, with the 2031 estimates at $1,636B against $831B. Compute is the biggest slice, $494B growing to $1,127B. The figures come from different houses, Goldman Sachs for AI and Mordor Intelligence for oil and gas, so the comparison is approximate.

• Takeaway 02

  One GPU-hour is not one product

  The same Nvidia H100 rents for $1.79 to $10 an hour across 13 providers, published 2025 rates, and a study running the same workload across 3,500 GPUs at 11 providers found H100 performance varying by as much as 34.5%. Network setup, location, power reliability, and platform support all price in, and none of it shows up in the label.

• Takeaway 03

  Compute behaves like electricity

  An idle GPU-hour disappears the moment it passes, latency limits how far work can travel, and location sets cost, the three constraints that forced electricity into regional pricing. The US hosts roughly three-quarters of the world's AI compute, and the export controls in place since October 2022 have split the market into US-aligned and China-aligned systems.

• Takeaway 04

  The GPU-hour trades first

  Four layers could carry the trade: GPU, GPU-hour, FLOP, and token. The contracts launching first are written on the GPU-hour: Silicon Data's H100 and B200 rental indices. A FLOP, one floating-point operation, is the same math on any chip, but its conversion formula between generations is unwritten; with no unit winning on paper, volume will likely pick the standard.

• Takeaway 05

  The financial layer is assembling now

  CME and Silicon Data launch compute futures on October 5, 2026, pending regulatory review; ICE's rival contracts, on Ornn's and NATIVX's indices, await the same sign-off, and Kalshi announced forward curves in July 2026. Sovereign and pension funds committed about $120B to AI infrastructure in 2025 and 2026, and AI labs are hiring for capital-markets roles.

• Takeaway 06

  Neoclouds need it most

  CoreWeave carries $8.5B in GPU-backed debt, and hyperscalers charge roughly three times neocloud rates for the same chip, a premium a public index would size. Michael Burry, who shorted the 2008 housing market, puts a GPU's realistic working life at two to three years; Atreides founder Gavin Baker argues it could stretch to ten or fifteen. Only real trading volume settles it.

Why This Matters

Commodity markets have formed where scarce supply met surging demand, and every failed contract, from onion futures to DRAM, traced to concentrated supply or a product that could not be standardized. That is the screen to hold compute against. Standardization is the open half: the first contracts are written on the GPU-hour, and a published conversion formula that translates FLOPs across chip generations is still missing. Volume is the other: the October 5 launch, still pending review, is the first test, and a liquid curve would settle the dispute that matters most, Burry's two-to-three-year GPU against Baker's ten-to-fifteen. Until then, the cold start problem sets the bar: sellers will not list without buyers, buyers will not show without sellers, and traders will not quote a price without volume.

Featured Charts & Graphs