BlackRock paper, co-written with ETF product executives, floats exchange-traded compute futures

Back out the growth rate and you can see the spend a compute forward curve would sit on. Sell-side analysts put AWS, Microsoft's Intelligent Cloud and Google Cloud at about $1.1 trillion of combined revenue by 2030, rising 29% a year from 2025 ^[5]. Divide 1.1 trillion by 1.29 to the fifth power and the 2025 base is roughly $308 billion ^[6]. Hedging demand on a book that size is a real business, and BlackRock says the demand would come from both providers and consumers of capacity ^[1].

The obstacles the paper lists decide whether the contract can exist at all. Chip generations differ in productivity, energy costs differ by region, and there are no workable standards yet for cash settlement or for delivering contracted capacity ^[7]. BlackRock calls these "important but ultimately resolvable design considerations" ^[8]. A cleared contract needs a grade and a delivery point: a chip-hour on last generation's silicon where power is cheap is not the same unit as one on this generation's where power is not, and nothing settles until somebody writes that definition down.

The collateral leg is the half that could work without a futures exchange at all. The paper says claims on compute capacity could be shown, moved, pledged as collateral and settled on a blockchain, and that this might bring more institutional investors in ^[2]. An investor with no use for a GPU can hold the price risk. An operator who has already contracted for capacity can borrow against the claim.

On the payment side the case rests on the rail's existing scale. Adjusted stablecoin volume passed $11 trillion in 2025, on par with Visa and Mastercard, after growing about 80% a year from 2020 against roughly 8.5% for ACH ^[9][10], and circulating stablecoin market cap topped $300 billion in September 2026 ^[11]. BlackRock's argument against the incumbent rails is narrow and specific: card networks and ACH carry onboarding requirements and settlement economics poorly suited to always-on, very low-value payments ^[4]. In the paper's sketch, an agent queries marketplace APIs for capacity on price, latency, location and hardware, then pays over a rail such as x402 ^[18].

The measured version of that flow is small. TRM Labs estimates AI agents accounted for 0.6% to 7.5% of $25.62 million in screened x402 payments since May 2025 ^[12], which is between about $154,000 and $1.92 million ^[13]. The top of that range is one part in 5.7 million of last year's stablecoin volume ^[14]. "AI represents machine-native intelligence, while digital assets represent machine-native money," the paper says ^[15].

Who wrote it matters to how you read it. Two of the four listed authors run product: Jay Jacobs is U.S. head of equity ETFs and William Helm heads U.S. iShares product innovation, alongside digital assets head Robert Mitchnick and digital assets research head Will Su ^[16]. BlackRock is also one of the founding validators of Circle's Arc blockchain, which launched its mainnet this month and was built partly for AI agents transacting on their own ^[17].

The dull outcome is that hyperscalers keep selling capacity bilaterally on multi-year contracts, never publish the prints an index needs, and the futures line stays a paragraph in a research paper. The cheap outcome, and in my view the likely one, is a financially settled contract on a proxy index that never touches physical delivery. The third is that agents compound from a small base and the stablecoin rail argument carries whatever happens to the contract. Evidence against the first two: a listed compute contract whose open interest stays near zero a year after launch, or a TRM update that keeps the agent share of x402 under 1% as the screened base grows.