Two Johns Hopkins financial economists have produced the first formal pricing framework for compute futures — and the first estimate of the compute risk premium: high single to double digits, paid by the fleets to whoever will carry their price risk. The framework is right. The map of the market is already out of date, and the gaps are where the trades are.
In the window between the CME × Silicon Data and ICE × Ornn futures announcements in May and the October 5 NYMEX listing date, the academy filed its first brief. Federico Bandi and Yinan Su of Johns Hopkins posted "(Early) AI Compute Asset Pricing" to arXiv on July 13, revised August 3. It is not quite the first paper to model compute derivatives — a June SSRN working paper by Amine Assody fit a reduced-form forward-curve and volatility model to the same non-storable, depreciating commodity — but it is, as far as we can establish, the first to price the risk: the first published empirical estimate of a compute risk premium. Both index administrators at the center of the listed complex, Silicon Data and Ornn, supplied their data directly to the authors.
The paper matters to this desk for a simple reason: our coverage of the compute complex to date has run through market structure — who computes the indices, what the contracts settle on, who financed the fleets, who sells the data, and who is affiliated with whom. What no one had supplied is the asset-pricing layer: what a compute future should cost, and who earns what for holding it. Bandi and Su supply exactly that scaffolding, and a first number to hang on it.
First: cost-of-carry is dead on arrival. "A GPU-hour that is not used today cannot be carried forward and delivered in the future," the authors write. "In this respect, compute is closer to electricity, freight, or other capacity-constrained service commodities than to a storable physical commodity, such as oil or metals." No storage means no cash-and-carry arbitrage, which means nothing pins the futures price to today's spot print. The economic link runs instead from the futures price to the expected spot price at expiry. Every intuition imported from oil desks fails; every intuition imported from power desks transfers.
Second: term rentals are not forwards, and the difference has a name. The closest thing to a compute forward curve today is the reserved-rental market — one-to-three-year capacity commitments. The paper's most useful conceptual contribution is to show these are contaminated proxies: a physical term rental bundles price insurance with what the authors call a real capacity-locking option — guaranteed access to scarce capacity with a particular provider, in a particular location, at a particular reliability tier. A cash-settled financial future carries the insurance and nothing else. The gap between the two is the "physical access wedge": non-negative, converging to zero as delivery approaches, and growing with horizon — because the box you locked today becomes progressively obsolete while the index it is measured against rebalances toward next-generation chips. Term-rental-implied forwards are therefore upper bounds on where true futures should trade, and the bound is loosest at the long end.
Third: once listed, futures price risk, not carry. Ft(T) = Et[ST] − λt(T), and the sign of λ is decided by hedging pressure — the Bessembinder–Lemmon logic from electricity forward markets, imported wholesale. Providers are exposed to revenue risk and sell forward; users are exposed to right-tail price spikes and buy forward. Whichever side presses harder sets the sign. The authors' bet, and their evidence, is that the fleets press harder: providers are the natural short hedgers, so futures should clear below expected spot, and the long side should be paid for carrying the risk.
Because listed futures don't yet trade, the authors build synthetic ones from Silicon Data's term-rental curves — daily, zero to 36 months on a quarter-month grid, three benchmarks — and compute what a long position held to delivery would have earned. It is the first compute futures return panel anyone has published, and the numbers are not small:
| Benchmark | Raw avg. return | Annualized |
|---|---|---|
| A100 (Silicon Data neo-cloud) | 4.2% | 7.5% |
| H100 | 13.4% | 26.2% |
| B200 | 4.3% | 11.2% |
Their own rule of thumb makes the number concrete: if you expect B200 rent at $7.00/GPU-hour a year from now, a futures print around $6.25 is not unreasonable — a ~12% discount to expectation, the insurance premium the fleet pays to lock revenue. In the rolling constant-maturity version, 11 of 15 generation-maturity strategies earn positive annualized means, with mostly positive and weakly maturity-increasing betas to the equity market. The authors are appropriately modest — a short sample, spanning an extraordinary demand wave, with H100's 26.2% largely earned around the late-2025 agentic-AI repricing — and they flag that negative short-maturity returns are an artifact of the physical access wedge running off, not evidence against the premium. First-pass, yes. But it is the first time anyone has put a number on what the long side of compute gets paid, and the number is high-single-to-double digits.
Three of the paper's findings independently corroborate positions this desk has published, and one strengthens a position beyond where we had taken it.
The hedge-mismatch warning, upgraded. Our listing-window piece showed identical H100 silicon printing 2.6× apart by counterparty tier, with the futures settling on the cheap side of the gap — so hedging a hyperscaler bill with the listed contract is a correlation bet, not a hedge. Bandi and Su's correlation network across the twelve index series goes further than our level-dispersion argument: Silicon Data and Ornn series separate into distinct clusters, and some cross-provider pairs are negatively correlated.
Negative correlation means the hedge can lose on both legs at once. "Your hedge is noisy" was our version; "your hedge may have the wrong sign" is theirs. Their own data supplies the same tier gap: the H100 hyperscaler index at $7.43 against the neo-cloud index at $2.50 — 2.97× on their sample date.
Our backwardation observation gets its theory. In early August we recorded that H100 forwards sat in mild backwardation while the circulating narrative insisted on contango. The paper's Figure 4 shows Silicon Data's forward curves downward sloping over most of the sample — and its hedging-pressure model explains why a curve can slope down with no one expecting prices to fall: leveraged fleets carrying depreciating hardware pay λ to lock revenue, exactly as short hedgers did in the grain and power markets. (The same figure shows the long end of the H100 and B200 curves turning up from late 2025 — the term structure pricing the agentic-AI demand wave.)
The settlement design converges. Their §4.1 predicts first-generation contracts cash-settled on Silicon Data's neo-cloud on-demand indices — SDH100RT, SDA100RT, SDB200RT — likely against the average daily index over the delivery month. That is what CME filed, per our verified record of the October 5 listing. Independent arrival at the same reading of the same documents is worth something.
i. The market they call unlaunched has been trading since May. The paper's empirical strategy — synthesizing futures because forward prices don't exist — is contradicted by the tape. As of its August 3 revision date, the following were live: Kalshi's compute event contracts and its Compute Forward Curves (market-implied B200/H200/A100 forwards, July 14) settling on Ornn's index; Architect's offshore perpetuals on the same index; Polymarket's terminal and touch ladders; and a dealer-intermediated OTC channel that did its first OCPI swap on May 27 and its first six-figure block in June. Kalshi, Polymarket, and prediction markets appear nowhere in the paper; perpetuals get one sentence, "in principle." This is not a gotcha about academic lag — it is a methodological opening. Market-implied curves are financial prices: no capacity-locking option, no physical access wedge, none of the contamination the authors spend a section modeling around. The cleanest version of their risk-premium test is runnable today on the curves they didn't know existed — and this desk has the data access to run it.
ii. The panel sits on an index that was restated. The synthetic return panel is built from Silicon Data curves reaching back to September 2024. Our index-methodology research recorded that Silicon Data's December 2025 revision restated the A100 series upward by roughly 35–40%. The authors caveat that early forward-curve and spot records "may not be fully harmonized" and drop the earliest start dates — the right instinct, generically applied. But a double-digit restatement in one of three benchmark series is a specific defect, not a generic one: hold-to-maturity "returns" computed across a restatement boundary are partly index administration, not market movement.
iii. No-arbitrage is not quite as dead as they say. The paper is right that nothing can store a GPU-hour, and right that cash settlement severs the physical bridge — for the CME contract. But the complex they didn't map includes an EFP: the Compute Desk × Architect facility announced July 8 converts futures positions into physical GPU capacity. An EFP cannot resurrect cost-of-carry, but it does restore delivery-date convergence discipline: a futures price that strays from the physically deliverable capacity price at expiry becomes exchangeable against it, and someone will take the exchange. Our index research identified EFP convergence as one of the two forcing functions on index quality. The paper's framework is correct about carry and incomplete about convergence — and the incompleteness matters most precisely where their physical access wedge is largest.
iv. Every object in the model inherits the utilization blind spot. The paper's headline macro number — compute service flow at 1.35% to 4.03% of U.S. GDP — is computed, by its own words, "assuming full utilization": 19.7 million H100-equivalents times the rental print times 8,760 hours. The authors are careful to label it a rental-equivalent flow, not revenue. But the assumption runs deeper than the GDP arithmetic. As our Silicon Data audit established, no administrator anywhere publishes an observed utilization or realized-revenue benchmark — so ST is a quote-based rental print, E[ST] is an expectation over quotes, and λ is estimated from returns on quotes. Rental indices can hold firm while fleets idle. The first academic treatment of compute pricing leaves the complex's largest measurement gap exactly where the industry left it: unmeasured, and now load-bearing.
Adopt the vocabulary. "Physical access wedge" and "capacity-locking option" are better names than any we had for the gap between reserved-rental curves and financial forwards, and F = E[S] − λ is the right frame for every curve-shape argument in the corpus. Cite the anchors: the 1.35–4.03% GDP band and the $700 billion hyperscaler capex figure (2.2% of GDP) are the most defensible macro-scale numbers yet published — full-utilization caveat attached.
Then run the two measurements the paper sets up but cannot perform. The wedge is observable now: Silicon Data's term curve minus Kalshi's market-implied curve, same benchmark, same tenor, same day — the first empirical print of a theoretical object the paper can only bound. And the risk premium is testable on financial prices: Kalshi's curves against realized index prints, no synthetic construction, no wedge run-off correction, from July 14 forward. Five weeks of sample is a first print, not an estimate. But that is exactly what the paper itself is — and its authors would presumably rather see their λ measured on a real curve than a synthetic one.
Paper quotations and figures extracted directly from arXiv:2607.12156v2 (HTML full text, August 17, 2026), including §3.3, §4.1, §5.2.1, §6, §7.1–7.3 and the appendix GDP calculation. The hold-to-maturity averages (7.5% / 26.2% / 11.2% annualized), the 12.61% Table 3 estimate and the $6.25-against-$7.00 rule of thumb are the authors' own numbers, quoted from a short sample they themselves label first-pass. Bessembinder & Lemmon (2002) confirmed in the paper's bibliography as the electricity-forward-premium reference. Priority checked: an earlier compute-derivatives model exists (Assody, "Pricing Compute Futures," SSRN 6926798, June 2026 — forward curve and volatility, no risk-premium estimation), and the authors themselves claim only to be "the first to study compute as an emerging asset class" to the best of their knowledge; this piece credits them with the first risk-premium estimate, not the first model. Cross-referenced Kinetic Alpha records: the 2.6× counterparty-tier dispersion and neo-cloud settlement basis (Aug 14); the SDA100RT restatement and EFP forcing-function analysis (Aug 1); the utilization negative, re-verified across administrators (Aug 16); Kalshi/Polymarket/OTC market facts (Aug 3), each verified against primary sources at the time of publication. Not established and not asserted: that the restatement contaminates the authors' post-exclusion sample; the identity of the paper's second electricity reference; any impropriety in administrator-supplied data. This piece states public facts about named firms and alleges no misconduct.