What makes a compute benchmark hedgeable

Abstract

Compute is being financialised on the assumption that a price index is enough to support a derivatives market. It is not. A benchmark becomes hedgeable through specific structural properties, each of which has been learned expensively by an older commodity complex. This note extracts those properties, applies them to compute, and is explicit about which ones compute currently fails, including my own index.

The short version: the tradeable object in every mature complex is an average over a delivery period rather than a spot fix; the estimator must have a non-degenerate delta vector; the panel must be dense enough that the estimator is locally smooth; and the settlement value must become final. Compute today satisfies the first, is being built to satisfy the second, and largely fails the third and fourth.

---

1. The state of play

Both listed contracts settle on a monthly aggregation of a daily $/GPU-hour index. That convention is not an accident, and §3 explains why it is the single most important design decision in the whole complex.

---

2. Five design rules, and where they come from

2.1 The delta vector must be non-degenerate

For a weighted median, ∂I/∂pₖ = 1 for exactly one constituent and 0 for all others. Three consequences follow: providers who are not the median have no index beta, so the natural short has no reason to sell it; the response is a jump function; and replication requires switching 100% of notional at the crossing point, which is unwarehouseable.

This is why SOFR works and a six-name median does not. SOFR is a volume-weighted median over thousands of daily repo transactions, so its constituent is an observation rather than an institution. Density, not the choice of statistic, is what makes a median safe.

Compute verdict: achievable, but only by aggregating over offers rather than providers, which in turn requires a unit definition that admits marketplace inventory (see Research Note 2026-01).

2.2 Composition changes must not move the level

Every mature index separates the price effect from the composition effect with a divisor or multiplier reset. The Baltic Exchange dropped Handysize from the BDI on 1 March 2018 and changed the multiplier from 0.10907849 to 0.1 to hold the level continuous. The C5TC(180)→C5TC(182) transition in January 2026 used a fixed differential, a new multiplier of 0.11026, and legacy publication through 24 December 2026. Platts announced WTI Midland's entry into Dated Brent twelve months in advance.

The pattern is consistent: fixed cardinality, pre-announced reconstitution, a published divisor, and parallel publication of old and new definitions.

Compute verdict: necessary, but with a caveat most treatments miss. A Laspeyres link with a same-day divisor reset contributes exactly zero return on entry and exit. Where panel churn is high and correlated with price moves, a venue leaving because it cut price, the link ratchets: rises made in-panel are permanent, falls taken while absent are laundered out. I measured this at roughly +7.8% per cycle for a constituent at a 25% cap. Chain-linking is the right tool for scheduled reconstitution and the wrong tool for daily churn.

2.3 Assessed prices carry a bias that widens exactly when you need the hedge

Catalog rates are step functions with an unobservable, time-varying discount to transacted levels. The bias widens in a glut and collapses in a squeeze, so tracking error against the real market is largest precisely in the tail being hedged. That is what makes list-price indices unhedgeable rather than merely imprecise, and it is the structural lesson of LIBOR.

Compute verdict: the binding constraint. Hyperscaler catalog rates for EU H100 have not moved in a month while marketplace prices moved 12.3%. Ornn markets OCPI to ICE explicitly on being built only from printed transactions. Any index without a transaction feed, mine included, must say so on its face rather than implying otherwise.

2.4 Settlement must be final

Append-only revision is excellent audit practice and incompatible with clearing. Platts corrects only for technical error or an upstream source correction, and explicitly never for information arriving after publication. SOFR revises same-day only, at 14:30 ET, and only for changes above one basis point.

The cost of breaking finality is not theoretical. After the LME cancelled nickel trades in March 2022, open interest fell from roughly 236,000 to 126,000 lots, and volumes fell across unrelated metals. Finality is a property of the whole venue's credibility, not of one contract.

Compute verdict: unsolved, and mostly unaddressed. Note that finality is a publication property rather than a storage property. An immutable database and a final settlement value are complementary, not in tension.

2.5 A gap is an undefined settlement

A minimum-source gate that publishes nothing is correct for a research series and fatal for a cleared one, because a CCP needs a number on every settlement date. Real contracts define a fallback waterfall. CFTC Part 38 Appendix C names low participant counts as a manipulation-susceptibility factor in terms that are uncomfortably direct: "situations susceptible to manipulation include those in which the volume of cash market transactions and/or the number of participants... are very low."

Compute verdict: the honest answer for a thin panel is to keep gapping and not claim settlement suitability. A fallback waterfall that readmits the hyperscaler tier a methodology has just excluded, at a 32–43% level difference, is worse than a gap. It is a gap wearing a number.

---

3. The one rule that matters most: settle on an average, never a fix

Across every mature complex, the object a contract settles against is an arithmetic mean over the delivery period:

• EEX and Nasdaq power futures settle on the mean of hourly prices across the delivery period (~2,880 market time units in a month), baseload and peakload.
• ICE JKM settles on a half-month assessment window.
• CME's own compute futures represent "one month's worth of rental costs."

Two reasons. First, averaging over ~30 daily prints is dramatically harder to manipulate than a single fix. That is the lesson of the WM/Reuters 4pm scandal, after which the fixing window was widened from one minute to five, following £1.1bn in FCA and $1.4bn in CFTC penalties. Second, an average matches the economic exposure. Nobody consumes a GPU-hour at 11:00 UTC on a Tuesday; they consume a month of capacity.

This is the design rule compute has already adopted, and it is why a daily index with moderate single-day noise can still support a sound monthly contract. It also means daily precision is worth less than daily consistency, which argues against clever estimators and for boring, reproducible ones.

---

4. What compute has that oil does not: generational obsolescence

Brent in 2026 is Brent in 1996. An H100 is not.

Azure's own EEA pricing on 2026-08-15 shows H200 SXM at $13.78/GPU-hr against H100 SXM at $16.53, newer and strictly better silicon priced 17% below the older generation. Silicon Data's published statistics show the same asymmetry in volatility: H100 neocloud coefficient of variation 2.6%, against B200 at 11.4% with an 18.9% maximum daily move.

Two implications:

1. Never normalise across generations. H100→H200 is a pure memory upgrade at identical FLOPS: per BF16 PFLOP-hour the H200 is 31% dearer; per HBM-TB/s-hour it is 9% cheaper. Two defensible normalisers disagree in sign. "Effective compute" is not salvageable as a settlement basis. Both CME and ICE reached the same conclusion independently, and their contracts are per-GPU-type.
2. The headline should not be anchored to the dead end of the curve. An index whose flagship is H100 is indexing the generation with the least volatility and the least hedging demand.

---

5. Is there a correlated hedge?

A benchmark with no deliverable and no storage gets liquidity only if some other traded instrument correlates with it. For compute the obvious candidate is power.

The decomposition does not support it. Per-GPU facility load for an H100 is about 1.59 kW (700 W × 1.82 node overhead × 1.25 PUE). At European industrial rates that is roughly €0.11 per GPU-hour in the Nordics, €0.25 in Germany and €0.35 in Ireland, against a ~€2.75 price. So electricity is 4–13% of the price, and the elasticity of compute price to power is 0.04–0.13. German baseload at ~30% volatility contributes about 2.7% of compute volatility.

Power futures therefore hedge the seller's residual cost, not the index. Once capex is sunk and revenue is contracted, power is 50–65% of a neocloud's remaining controllable cost. That is a real exposure, but it is a spark-spread trade rather than an index hedge. A published "compute minus power" series would be 91–99% one leg and should not be presented as a spread.

The honest conclusion is that compute currently has no correlated hedge instrument, and liquidity will have to come from natural two-sided interest instead. There is academic support that such interest exists. Bandi and Su (2026) construct the first compute futures return panel sorted by GPU generation and maturity, and find evidence of a positive compute risk premium attributable to hedging pressure from compute providers, meaning structurally long sellers paying to lay off risk.

---

6. The regulatory position, stated carefully

Regulation (EU) 2025/914, applying from 1 January 2026, narrows the EU Benchmarks Regulation considerably: Titles II–VI now apply only to critical benchmarks, significant benchmarks, and EU Climate Transition / Paris-aligned benchmarks. Non-significant benchmarks fall outside them.

That is frequently over-read, and I over-read it myself before checking the text. New Article 2(1c) separately provides that "Article 19 applies to any commodity benchmark based on contributed input data", unless it is a regulated-data benchmark, has a majority of supervised contributors, or is a gold/silver/platinum critical benchmark. Article 2(2)(g) exempts commodity benchmarks only where total average notional referencing them stays below EUR 200 million over 12 months, a threshold a cleared contract would cross quickly.

Whether scraped public rate cards constitute "contributed input data" is an open question on which I express no view. The practical point stands regardless: exchanges and clearing members will require BMR-grade governance contractually long before the law compels it, so the sensible course is to build to Annex II shape now and take advice before any contractual use. (Not legal advice.)

---

7. Scorecard

The last row is the only one where a small, transparent, regional index has a structural advantage over well-funded incumbents, and it is the one I intend to compete on.

---

Sources

Primary: CME Group press release, 11 Aug 2026; ICE investor-relations release on OCPI; Regulation (EU) 2025/914 (EUR-Lex); CFTC Part 38 Appendix C; Baltic Exchange index change notices (2018, 2026); Azure Retail Prices API (queried 2026-08-15); Scaleway public Instance API (queried 2026-08-15); Bandi & Su, (Early) AI Compute Asset Pricing, arXiv 2607.12156.

---

TCI is a research publication. It is not investment advice and may not be used as a reference price in financial instruments. Nothing in §6 is legal advice.