The Contract Has a Date. It Needs a Dealer.

A futures contract is a machine for moving a specific risk from someone who has it and does not want it to someone who will carry it for a price. Before asking whether the Silicon Data or Ornn contracts are well designed, the prior question is whether the risk they reference — the month-ahead on-demand rental rate of a specific GPU SKU — sits on anyone’s balance sheet in a form that a cash-settled monthly average can offset. The exchange press releases answer this with a list: “traders, financial institutions, AI builders and cloud-service providers” (CME, May); “AI developers and hyperscalers” (CME, August); “the institutional buyers and operators who need it most” (Ornn); “operators and consumers of compute” and “energy market participants” (NATIVX). Lists are not evidence. What follows is our own map of who holds the exposure, which direction it runs, and how it is priced today.

Direction is stated from the perspective of the on-demand rate: a participant who loses when the rate falls is “short” in hedging terms (would sell futures). Sources for pricing practice: CoreWeave Q2 2026 call; American Compute neo-cloud model note (Mar 2026); Peony on neo-cloud capital raises (Aug 2026); Friedman on DDTL 4.0 (Aug 2026). Full citations in Sources.

The finding that organises everything else

Every commercial participant in that table who has a real exposure prices it fixed. The neo-cloud sells two-to-five-year take-or-pay at a fixed rate; CoreWeave’s CEO describes the business as “building scale through these long-term take or pay contracts,” and revenue backlog was $104 billion as of June 30, 2026, against which term debt is sized. The lender lends against that contracted revenue and explicitly excludes merchant revenue from the borrowing base. The buyer signs the same contract from the other side. There is, as far as the public record shows, no index-linked physical compute contract anywhere — no lease that resets to Silicon Data, no loan covenant that references Ornn, no ABS trigger keyed to an OCPI print. We searched for one across operators, lenders, rating-agency reports and exchange filings and did not find it.

This matters because it inverts the sequence every successful commodity benchmark followed. In oil, the physical market moved to formula pricing off the benchmark first — PEMEX in 1986, most exporters by 1988, over 60% of traded oil tied to spot-linked prices by 1987 — and the floating exposure that created is what producers, refiners and eventually banks needed futures to offset. In gas, FERC Orders 436 and 636 unbundled the pipelines and created a spot market at Henry Hub (1988) before NYMEX listed the contract (1990). Compute is being asked to run the film backwards: list the futures against an index that nothing physical references, and hope the physical market reprices to it. That can happen — it is roughly what the prediction markets are doing with Ornn’s print — but it is the harder direction, and it is why the end-user list in the press release is a list of hopes.

On August 11 NYMEX filed submission 26-370 with the CFTC for two contracts: Silicon Data H100 Rental Index Futures and Silicon Data B200 Rental Index Futures, rulebook chapters 1045 and 1047, Globex codes GPU1 and GPU2. The same day CME issued SER-9785 with the commercial terms. Two things about the filing are worth more than the specs themselves. First, it is a Rule 40.3(a) voluntary request for product approval, not a 40.2 self-certification — CME chose to ask the Commission rather than certify and list, which is why every release since May has carried “pending regulatory review.” Second, the two exhibits that would let anyone outside the exchange evaluate the cash market — Exhibit B (position limits, accountability and reportable levels) and Exhibit E (the supplemental market information about the underlying cash market) — are filed under separate cover, Exhibit E with confidential treatment requested. The public filing’s statement that the index is “supported by a robust level of trading in the underlying cash market” is boilerplate with no number behind it that the public can see.

ICE’s two complexes are less documented still. The May 19 announcement with Ornn names the index (OCPI), the SKUs (H100, H200, B200, RTX 5090) and the settlement style (USD, cash), and nothing else: no ICE listing entity, no contract size, no listing date, and no product certification we can find. The July 1 NATIVX announcement is the same shape but does give a timeframe — contracts “expected to be launched later this year, subject to completion of relevant regulatory processes” — and names Synova Global as the licensor behind the COIL sub-indices. Neither release names the exchange or clearing house that would list the contracts. Ornn’s own design note from December 2025 sketches an Asian-style contract on a 720-hour unit, which suggests the two complexes will look alike at the term-sheet level and differ only on the thing that matters: whose print settles them.

The end users, in their own words

We read CoreWeave’s Q2 2026 transcript looking for the words futures, hedge, index and spot. None appear. The CEO’s characterisation of demand is the opposite of a floating exposure: long-term take-or-pay, with the one concession that “enterprise tends to want to enter into contracts that are not five years in length.” The CoreWeave DDTL 4.0 deck — $8.5 billion, rated A3, SOFR+225 — contains no reference to rental price risk or to any index; as Dave Friedman put it, “the contract has deliberately engineered most merchant price risk out of the senior tranche.” Macquarie’s Fluidstack facility is backstopped by Google, not by a swap, because — in Friedman’s February phrasing — there is “no insurance or swap market where Macquarie can offload the risk.” The only operator quoted in any exchange release is Hyperbolic, a GPU marketplace, and it speaks of hedging tools “becoming essential” for neo-clouds and labs in general, not of any intention to hedge itself. Architect’s Brett Harrison reports, on X, inbound interest from neo-clouds, data-center operators, energy companies and FCMs; that is a vendor’s account of its pipeline and we treat it as such.

Silicon Data’s CEO has said the product “has to be a functional, usable product, not a casino.” Sourcery’s May report listed the question among its open unknowns: “Who will be the first large hedgers and market-makers on these exchanges? No exchange has disclosed committed participants.” Three months and one CFTC filing later, no exchange has.

Set the listing aside and ask the design question on its own terms: of the risks that live on the balance sheets in Section 01, which could a standardised, cash-settled, monthly-average instrument plausibly transfer, and what would it leave behind? The honest answer is that the instrument maps cleanly onto exactly one exposure — the merchant tail of a US neo-cloud running the same SKU that the index samples — and degrades from there in a way that is measurable and, in places, already measured.

Basis, in the order it will bite

Tier basis is the one we have written about before and the one Bandi & Su quantify: H100 hyperscaler $7.43 vs neo-cloud $2.50 in their sample, and — the part that should stop any enterprise treasurer — some cross-provider index pairs are negatively correlated, so that “a futures contract that references the Silicon Data’s H100 NEO index is not an ideal hedging instrument for a user in the market segment covered by a negatively-correlated Ornn’s index.” A hedge that is noisy is a nuisance; a hedge that may have the wrong sign is a new position.

Utilisation basis is the blind spot we flagged in the Silicon Data audit and it has not closed: no administrator publishes an observed utilisation or realised-revenue series. A neo-cloud short GPU1 against merchant capacity is hedged against the posted rate falling. It is not hedged against the posted rate holding at $2.50 while the fleet rents 55% of its hours, which is the failure mode that actually impairs operators and their lenders. The index cannot see it, so the contract cannot carry it.

Obsolescence basis has no analogue in oil. A barrel of WTI in 2029 is a barrel of WTI. An H100-hour in 2029 is rent on hardware three generations old, priced by the scarcity of whatever replaced it. The 36-month strip therefore gets less meaningful as it lengthens, which is exactly the part of the curve a renewal hedge needs. Friedman’s taxonomy lists six bases for compute — liquidity, locational, quality, calendar, settlement, obsolescence — and notes that the last “has no such anchor.” The SKU-specific design is the right call for settlement integrity and the wrong call for tenor; a chip-agnostic, performance-normalised index (NATIVX’s energy-normalised COIL is one attempt) would trade the first for the second.

Settlement basis is the one the prediction markets have already exposed. Allium’s analysis of Kalshi’s Ornn-settled contracts found that two days before closure the price “typically deviated from the actual closing price by around 10%.” A 10% miss at T-2 on a monthly-average settlement is not a hedging instrument; it is a referendum on the index. Part of that is thin flow. Part is that an index revised by provider additions and methodology changes of mid-single digits carries a settlement risk a hedger cannot diversify.

The folk history of WTI is that NYMEX listed it in March 1983 and the oil industry came. The documentary history is slower and more useful. The contract traded about 3,000 lots in its first month and took a year to clear 100,000 in a month; daily volume averaged under a thousand contracts at launch and reached roughly 18,000 by the end of 1987. The established oil companies’ reaction was, in Yergin’s words, one of “skepticism and outright hostility” — a senior executive at one of the majors dismissed oil futures as “a way for dentists to lose money” — and the early floor, in Goodman’s account of the exchange, was populated by locals who had never traded a physical barrel. The end users on the exchange’s own list did not show up for years.

Three things happened in the meantime, and they are the adoption mechanism. First, the physical market repriced to the screen: the pre-existing Cushing wet-barrel trade gave the contract a real deliverable, Platts had been assessing WTI since 1981, PEMEX adopted formula pricing in 1986, Saudi netback pricing followed, and by 1988 market-related pricing was the main method in international crude trade. Once the physical barrel floated off the benchmark, every refiner, producer and trader had an exposure the futures could offset. Second, Wall Street built the dealer layer: Morgan Stanley’s commodities group from 1985, J. Aron under Goldman, then Phibro and the trading houses, writing bilateral swaps — for producers, for airlines, for Mexico’s 1990–91 sovereign hedge at $17 — and offsetting the residual on NYMEX. Third, the regulator accommodated the dealer: in 1991 CFTC staff granted the first hedge exemption to a swap dealer to hold futures against a book of commodity-index swaps, and the 1993 Part 35 exemption and the 2000 CFMA kept the bilateral layer outside exchange rules. Within a few years, Yergin notes, most of the major oil companies and some exporting countries were participating — the price risk had become too large to stay out.

The dealer share, measured

Our working prior — that a significant portion of WTI exposure was banks hedging bilateral activity with producers and consumers rather than producers hedging directly — is correct for every period the CFTC has data for, and unmeasurable before that. The Commission’s internal swap-dealer classification begins in 2000; public disaggregated data begins June 2006. What it shows is not subtle.

Henry Hub is the cleaner cash-market-first case and the faster adoption. FERC’s special marketing programs created a spot market in the early 1980s (2–2.5 tcf a year against 17–18 tcf of consumption, per a 1986 Energy Law Journal survey); Order 436 opened pipeline access in 1985; the hub itself was formed in 1988; NYMEX listed in April 1990 and the contract had, per Parsons (citing Brinkmann and Rabinovitch), “the fastest takeoff of a contract in the history of NYMEX”; Order 636 unbundled sales from transport in 1992; marketers’ share of supply went from 20% in 1987 to 49% in 1995; and traded futures went from 0.42 tcf in 1991 to 80 tcf in 1995. The marketers — Enron, NGC, the names that would later define the bilateral layer — were the gas analogue of the oil swap dealers, warehousing basis and term risk for utilities and producers and laying off the flat price at the Hub. Note what the regulator did in both cases: it did not approve a futures contract into a bundled, fixed-price physical market and wait. It unbundled the physical market first, and the benchmark was chosen because the spot liquidity already existed there.

The graveyard, sorted by cause of death

Most contracts fail. The literature summarising Silber (1981) and Carlton (1984) puts the share that never attract a sustainable level of trading at roughly two-thirds to three-quarters, with most failures arriving soon after launch; the CFTC’s own review of that work says simply that “most new derivatives fail, usually soon after their launch.” The compute contracts do not need to beat that rate to be worth listing; they need to avoid the specific diseases.

The academic criteria distilled from this record — Black (1986), Silber (1981), Carlton (1984), and the later case-study literature — come down to a short list: a large and volatile cash market; homogeneity or a grading standard; two-sided hedger demand with commercial basis small relative to price volatility; no close substitute that already does the job; and a pool of speculators and market makers. Brorsen and Fofana found that an active cash market “perfectly predicts” whether a commodity gets a futures market. The CFTC’s request for comment is, read plainly, the staff asking whether the first and second criteria are met. Section 01 is our answer to the third and fourth.

If the contracts are going to be used for something other than speculation, the history offers two routes, and they are not exclusive. Path A is the WTI route: a dealer writes bilateral, structured exposure to commercial counterparties and uses the futures to manage the residual. Path B is the Henry Hub route: the physical market reprices to the index, creating floating exposure that end users hedge directly. We score the compute contracts against the preconditions of each.

Path A: the dealer is already in the room, wearing a lender’s badge

The WTI analogy breaks in one place and it is the interesting place. In 1985 the banks came to oil as outsiders with a trading idea. In 2026 the banks are already the largest creditors in the compute complex. Blackstone, Magnetar, Macquarie, Carlyle, JPMorgan, Morgan Stanley, Goldman, MUFG and PIMCO are the names in the DDTL syndicates and ABS books; they have sized every facility against contracted revenue and written the merchant tail and the residual value out of the borrowing base precisely because they cannot hedge them. That is a bilateral exposure in search of a dealer — and the dealer and the exposed party are, in several cases, the same institution.

This gives Path A a plausible first trade that oil never had: the lender requires, or prices, a hedge on the uncontracted capacity and residual value of the collateral, writes it as a structured product (a fixed-for-floating on the index, or the residual-value swap Ornn is already marketing), and lays the flat-price residual off in GPU1/GPU2. Friedman has proposed exactly this — lenders “could require borrowers to hedge a portion of uncontracted capacity” — as a proposal, not a practice. It is the one version of the end-user story in which the natural short (the operator), the natural risk-warehouse (the lender) and the natural futures user (the dealer desk inside the same bank) line up on the first day.

Four things stand in the way, in rising order of difficulty. Liquidity: a dealer cannot lay off a 12,000-contract residual into a market whose largest predecessor had done $4.4 million of notional by late July; the first dealer has to be willing to warehouse unhedged for a while, which is what Morgan Stanley and J. Aron did in 1985–87 and what the bank withdrawal from coal shows can reverse. Position limits: Exhibit B is not public, and a hedger that cannot see the limit cannot plan a program. The exemption regime: the risk-management exemption that let swap dealers hold futures against a book from 1991 was eliminated by the 2020 Part 150 rule; what remains is the pass-through swap offset, available only where the counterparty’s side of the swap “qualifies as a bona fide hedge.” A neo-cloud hedging merchant capacity probably qualifies; a lender hedging residual value on collateral is a harder argument, and an insurer writing residual-value cover is harder still. The regulatory channel that made WTI’s dealer layer possible is narrower now than it was then. The index: a dealer pricing a residual-value or merchant-revenue swap needs the floating leg to track the counterparty’s actual economics, and a posted-rate, neo-cloud-tier, US-scoped, periodically-restated series is a weak floating leg for a lender whose collateral earns realised revenue in a fleet that may be 40% idle. The dealer can price that basis. It will price it wide.

Path B: the physical market is starting to float on its own

Nothing physical references either index today. But the precondition for Path B is not that contracts reference the index; it is that the physical market has floating exposure to be referenced. Three forces are creating it. Tenor compression: CoreWeave’s own admission that enterprise does not want five-year contracts means more of the book rolls more often, and every roll is a floating exposure on the renewal date. Spot deflation: H100 on-demand from roughly $8 at the 2023–24 peak to $2.50–3.50 by late 2025 has made every fixed contract written at the top a visible loss for the buyer and a visible gain for the seller — which is exactly the experience that taught oil buyers in 1986 to demand formula pricing. The prediction-market print: Kalshi, Polymarket and Robinhood settling weekly and monthly contracts on Ornn’s index, with a forward curve out to a year, is a daily public reference that did not exist in March. It is small and it is noisy, but it is the first thing a buyer can point to in a renewal negotiation and say “the market says $2.52.”

The contract designs help Path B in two ways and hurt it in three. They help by settling on a monthly average (which is how rent is billed) and by listing 36 months (which is how renewals are dated). They hurt because the settlement series is quote-based and tier-specific — the physical market will not adopt a reference that does not reflect the transactions it actually does; because there is no disclosed governance for adding or retiring SKUs, which is the Brent lesson; and because the two complexes settle on different indices that Bandi & Su find “tend to be separated” and are sometimes negatively correlated, so that the liquidity the physical market would need to see is split before it exists. The Architect/Compute Desk EFP is the most important Path B design element on the board, because it is the only mechanism that forces the paper price to meet a physical delivery; it is also pending, on an exchange that is itself pending.

The question this piece set out to answer is whether the market is sufficiently developed to support index instruments with significant application beyond speculation. Our answer is no, not yet, and the reason is structural rather than a matter of time: the physical market prices fixed, the settlement index prices posted quotes, and the only participants with floating exposure are lenders who have engineered it out of their senior exposure rather than hedged it. The contracts as filed will trade on listing day between market makers who are also index-provider shareholders and a retail-scale flow migrating from Kalshi. That is what WTI looked like in 1983 and what Case-Shiller looked like in 2006, and the difference between those two outcomes was not the contract — it was whether the physical market and a dealer layer arrived.

The WTI analogy holds as a mechanism and fails as a forecast. It holds because the channel through which commercial risk reached the exchange — a bank warehousing bilateral exposure and laying off the residual — is the channel most likely to carry compute risk, and the banks are already the counterparties. It fails as a forecast because the 1991 exemption that opened that channel is gone, because the physical market has not floated, and because the index that would have to serve as the dealer’s floating leg is not yet one a dealer would trust at size. The design elements that most improve the odds are not on the term sheet: a disclosed settlement series with a transaction-only data hierarchy, published position limits, a SKU-succession rule, an EFP that works, and — the one that costs nothing — a single sentence from one lender saying a hedge requirement is in a covenant. Watch for that sentence. It is the PEMEX moment.

Contract documents

• NYMEX Submission 26-370 (Aug 11, 2026) — 40.3(a) voluntary approval; chapters 1045/1047; floating price definition (Rule 1045101); Geography: United States, the sole parameter in Rule 1045102.C “Index Technical Configuration”; Exhibits B and E under separate cover, E confidential.
• CME SER-9785 (Aug 11, 2026) — 730 GPU-hour unit, $0.01 tick = $7.30, 36 months, block minimum 5, fees.
• CME press release, Aug 11, 2026 and May 12, 2026 — user lists; Keavey, Li, Duffy and Wilson quotes.
• ICE × Ornn, May 19, 2026; ICE × NATIVX, Jul 1, 2026; Ornn contract design note, Dec 17, 2025; Ornn GPU Value Protection (residual value swap), Jan 2026.
• CFTC PR 9286-26, Aug 19, 2026 and 91 FR 54259 (Aug 21, 2026) — Request for Comment on Compute Derivatives, RIN 3038-AF77; comments due Oct 20, 2026.
• Silicon Data H100 index page (neo-cloud SDH100RT $2.53; separate hyperscaler series); index announcements (Dec 3, 2025 restatement; Mar 25 and Jun 30, 2026 provider additions); Series A, Aug 11, 2026 — a $30.5M initial closing led by the Valor Atreides AI Fund, with CME Ventures (named CME Group in the Business Wire version), DRW, Jump, Wintermute, Tectonic, F-Prime, Samsung Next, VanEck and Further among the other investors.

End-user record

• CoreWeave Q2 2026 call transcript, Aug 11, 2026 — take-or-pay, $104B backlog, “not five years in length”; no mention of futures, hedging, index or spot. DDTL 4.0 overview.
• Friedman, “The market is pricing payment certainty” (Aug 10, 2026); Fluidstack and GPU financing (Feb 23, 2026); Friedman basis-risk primer (Jul 2, 2026).
• Peony on neo-cloud capital raises; American Compute on the neo-cloud model.
• Semafor, May 26, 2026 (Wilson on hyperscaler opacity; Li “not a casino”); Sourcery, May 5, 2026; Fortune, Aug 12, 2026 ($4.4M Kalshi notional; Allium ~10% T-2 deviation; Duffie); Octus / Cahill on a hypothetical compute swap, Aug 14, 2026.
• Bandi & Su, “(Early) AI Compute Asset Pricing,” arXiv 2607.12156 — tier levels, index separation, negative cross-correlations, positive risk premium.

Benchmark history

• CME OpenMarkets on WTI’s early years; Parsons, Ann. Rev. Fin. Econ. 9 (2017); Purvin & Gertz, The Role of WTI (2010); Yergin, The Prize; Goodman, The Asylum; Insights Global on Morgan Stanley and J. Aron.
• Fattouh, An Anatomy of the Oil Pricing System (OIES 2010); OIES Insight 68 (2020); Fattouh on Dubai (2012); OIES on Brent (2022).
• Büyükşahin et al., CFTC (2008); CFTC Staff Report on Swap Dealers & Index Traders (Sept 2008); Harris testimony (2008); Disaggregated COT, Nov 24, 2009; Disaggregated COT, petroleum, Aug 18, 2026; Berkovitz testimony on the 1991 exemption; Part 150 final rule (2020).
• Energy Law Journal vol. 7 (1986) on special marketing programs; implicator.ai on tracker disagreement (Jul 29, 2026); Harrison on X.
• Juris, World Bank Note 141 (1998) on Orders 436/636 and marketer share; LBNL-41098 (1998) on NYMEX electricity; Jud & Winkler on Case-Shiller futures; Kavussanos & Nomikos on BIFFEX (JFM 2000); S&P on the coal delisting (2021).
• Brorsen & Fofana (2001) summarising Black, Silber, Carlton; Till / JPMCC case studies (2016); Cavanaugh & Penick, CFTC OCE (2014); Pirrong, manipulation survey; IOSCO PRA Principles (2012).
• Kinetic Alpha: The Contract Got a Date. · The Index Was the Wedge. The Product Is the Risk Desk. · Bandi & Su on the compute risk premium · Compute Became Collateral. NVIDIA Sold the Floor.

Claims register

• “Zero named commercial hedgers” is a statement about the public record as of Aug 22, 2026 after a search of operator calls, lender decks, rating-agency coverage, exchange releases and the CFTC docket. Absence of evidence; a private commitment would not appear.
• The settlement series for GPU1 is inferred to be the neo-cloud SDH100RT series from Polymarket’s resolution source and Bandi & Su’s data; no CME or Silicon Data document states it. We flag this as an inference in-text.
• Dealer share of early WTI open interest (1983–1999) cannot be measured; CFTC swap-dealer classification begins in 2000 and public data in 2006. The claim that dealers were a significant share of exposure is supported for 2000 onward and by contemporaneous narrative (Morgan Stanley 1985, J. Aron, the 1991 exemption) before that. We do not assert a percentage for the 1980s.
• The COT reading that swap dealers’ net short in WTI reflects laid-off producer hedges is our inference from the direction of the position; the COT does not identify the swap clients, and the producer/merchant category is itself net long on the exchange. The 1991 exemption and the 2008 Harris testimony concern index-investor swaps, the same channel in the opposite direction.
• The 1991 hedge exemption recipient is described by the CFTC and the Senate PSI as “a swap dealer”; the identification as J. Aron/Goldman appears only in secondary journalism and is not asserted here.
• Tier dispersion figures: 2.6× is our Aug 14 reading of Silicon Data’s published sub-indices ($2.73 vs $7.19); 2.97× is Bandi & Su’s sample ($2.50 vs $7.43). Both are cited as ranges.
• “Fleet may be 40% idle” is an illustrative utilisation figure, not a measured one — the point of the utilisation argument is that no measured figure exists.
• Architect AX perps are described as pending; we found no evidence they are live or any volume figure. Kalshi’s “~517K contracts open” and $4.4M notional are Fortune’s figures as of Jul 27, 2026.