The Reserved Contract, Unbundled

When a neocloud signs a twelve-month reserved contract with an AI lab, the document does three things at once. It fixes a price for a year of GPU-hours. It extends credit: the seller delivers first and invoices monthly, so at any moment it is owed most of a year’s revenue by a company it cannot margin. And it locks volume: the buyer takes or pays, whether the racks are busy or not. In exchange the buyer prepays a slice up front and accepts a rate well below what the same capacity would cost a month at a time.

Compute Desk’s own research, written before either announcement, describes the credit part plainly. The upfront payment, typically around a fifth of the contract, is “partly working capital that neocloud suppliers use to finance the hardware they’re renting back to you, and partly a credit substitute for buyers who aren’t investment-grade.”^[12] The same note records what the shortage did to tenor: minimum commitments now start at six months, the sweet spot has moved to one to two years, and reserved H100 rates rose roughly 40% between late 2025 and spring 2026 with no instrument to manage the move.^[12] A seller that wants price certainty has to sell a year; a buyer that wants capacity certainty has to buy one. Nobody gets to pick just the risk they wanted.

That is what a market with only a reserved contract looks like: one price for three risks, and no way to see the split. The two announcements this summer are an attempt to write the split down. Nodal’s future takes the price risk out of the document and clears it. Architect’s exchange-for-physical takes a cleared price and puts the delivery back. What is left in the bilateral contract, once those two are done, is the part a clearing house cannot hold — and the ledger in section 06 is an attempt to put a number on each piece.

I have written about the pieces separately: the EFP mechanism itself, read through the crude, gold and metals precedents; Nodal’s margin engine and what it does when compute sits beside power; and the question of who holds a floating GPU-hour exposure at all. This piece is the seller’s view of all three together.

The word “IOSCO” appears twice in this stack and means two different things, which is worth being precise about because both matter to a seller. The first is the number. Compute Desk’s Hopper, Blackwell and H100 benchmarks launched earlier this year; on August 13 General Index announced that they are now calculated and published under GX Benchmarks Ltd, its wholly owned, FCA-regulated subsidiary, built to the IOSCO Principles for Financial Benchmarks and the UK and EU Benchmark Regulation.^[11] Compute Desk still supplies the transactions and distributes the series; an administrator with no position in compute now owns the calculation. That is the arrangement I described in Five Indices, One Price as buying independence institutionally while moving the incentive question down a layer onto contributor concentration, and nothing since has changed that read.

The second is the plumbing. Nodal Clear, which will clear door one, publishes a disclosure against the CPMI-IOSCO Principles for Financial Market Infrastructures — the standards most people in the business still call CPSS-IOSCO, after the committee that wrote them in 2012. It registered with the CFTC as a clearing house in September 2015 and was the first to elect the Commission’s Subpart C regime at registration, which binds it to the international standard rather than the domestic minimum; ESMA has recognised it as a Tier 1 third-country clearing house since March 2022.^[6] A benchmark built to one IOSCO standard settling into a clearing house disclosed against the other is, as far as I can find, a first in compute. It is not a first anywhere else; it is simply what a commodity contract looks like when it is finished.

Then the two doors. Door one is the Nodal future: on September 3 Nodal Exchange and Compute Desk announced GPU futures on Hopper, Blackwell “and beyond,” financially settled against Compute Desk’s daily blended benchmarks and cleared by Nodal Clear, “scheduled to be available this year, subject to regulatory compliance.”^[1] Paul Cusenza, Nodal’s chairman and chief executive, framed it around the power book: “Nodal has the majority of the North American power futures open interest, and power is a significant operating cost for data centers.”^[1] Compute Desk’s founder Andrawes Bahou framed it around who hedges: “For the first time, hyperscalers, AI infrastructure companies, AI labs and their lenders will be able to manage their financial exposures to compute.”^[1]

Door two is older and less regulated so far. On July 8 Architect Financial Technologies and Compute Desk announced ComputeConnect, an exchange-for-physical network in which capacity providers answer delivery requests over an open protocol, standard basis tables are published for different SKUs, memory configurations and locations, and the futures legs are booked to Architect’s American Innovation Exchange — a designated contract market Architect acquired in May and which remains pending regulatory review.^[3][5] The physical leg settles through Compute Clear, which Compute Desk describes as its settlement infrastructure and which the network says guarantees delivery.^[4]

The structural fact a seller has to hold onto is that these are two exchanges, not one. An exchange-for-physical is a privately negotiated futures trade in a particular exchange’s contract — one party buys, the other sells — paired with an opposite physical transaction, and it is submitted to and cleared by that exchange. A short on Nodal therefore cannot be unwound through a network that books its futures leg to the American Innovation Exchange. Nodal’s rulebook has provided for EFPs for years — Rule 4.7 in the current edition — and defines a Nodal contract as one settled “either financially or physically,” so nothing prevents a Nodal-side physical leg.^[9] But none has been announced, and until one is, the seller chooses a door by what it wants: a cleared cash hedge beside the power book, or a path that ends in racks. Figure 1 draws the stack as it stands.

Take a concrete seller. A neocloud has 1,024 B200s energising in December, uncontracted. It has a lender who advanced against the racks on the strength of contracts it does not yet have, a sales team, and a spreadsheet that says what the capacity has to earn. Today it has one real move and one bad one. It can sign a reserved contract, or it can run the racks on demand and hope. Both doors add a third and a fourth. Here is what each looks like from the seller’s chair, step by step.

Three things stand out from the seller’s side. First, route B hedges price and nothing else. The seller still has to find a buyer every month. If the racks sit at 60% sold, the strip pays out on hours nobody bought — which is a gain if the benchmark fell and a loss if it rose, and either way not what the seller wanted. That is why the strip is sized to expected sales, and why the ledger asks for a utilisation number before it will compare the routes. The reserved contract’s volume promise is a real thing, and the futures do not replace it.

Second, the cash flows reverse. In route A the seller receives money on day one. In route B it posts money on day one and may post more every day the benchmark rises before delivery starts. Compute Desk’s research anticipates exactly this: under futures, bilateral credit risk “gives way to daily mark-to-market and initial margin set by the clearinghouse,” and the burden falls harder on the participant without an investment-grade balance sheet.^[12] For a neocloud, whose cash is famously thin and whose lender sized the loan on contracted revenue, a margin call is not a rounding error. The ledger prints the size of it.

Third, route C is the only one that ends where the seller lives, in a delivery obligation. The exchange-for-physical is not exotic; oil, grain and metals convert paper to physical this way every day, and Nodal’s own rule defines it as “a privately negotiated and simultaneous exchange of a futures position in a Nodal Contract for a corresponding cash position.”^[9] I have written the precedents up separately. What is new is that the delivery counterparty is found through a network rather than a sales call, the basis for location and configuration is published rather than negotiated from scratch, and the price the delivery contract inherits is a cleared price. What is not new is that from the swap onward the seller is back in a bilateral contract, this time with a platform guarantee sitting behind it whose mechanics nobody has published.

A seller’s strip needs a buyer, and the question of who that is has been the weak joint in every compute-futures argument so far. I laid out in The Contract Has a Date that no commercial balance sheet currently carries a floating GPU-hour exposure: the physical market prices fixed, lenders lend against contracted revenue, and nobody is named as a committed hedger in any exchange filing. That is still true. What the two doors change is not the existence of the longs but the form they can take. A lab can now be long on paper before it knows which cluster it wants, and convert when it does. A lender can be long the merchant tail of its own collateral. A power trader can be long compute for reasons that have nothing to do with compute. The table sorts them.

Two of these deserve a sentence more. The lender is the buyer I keep coming back to, because it is the one that already holds the risk. When a bank advances against a neocloud’s racks it lends on the contracted revenue and excludes the merchant tail — the uncontracted months, the residual value at the end. That excluded tail is a floating GPU-hour exposure the bank has chosen not to price rather than one it does not have. A long strip against it is the most natural hedge in the complex, and the banks on Nodal Clear’s membership list — Goldman Sachs, JPMorgan, Morgan Stanley, Citi, BofA, Wells Fargo and others^[7] — are the same institutions that lend to the build-out. Last November the FT reported that Deutsche Bank was exploring ways to hedge its data-centre exposure, including synthetic risk transfer.^[17] A cleared future is the cheaper instrument for the same job, once it exists.

The hyperscaler is the buyer everyone assumes and nobody has yet seen. It rents from neoclouds in size and it sells capacity in size, so it is a two-way counterparty by construction; but its own contracts are the ones priced at the top of the tier ladder, and the benchmark is built from the transactions Compute Desk sees, which are not obviously those. Whether a hyperscaler’s treasury would hedge a bill against an index it does not print into is the question I raised about the CME contract in The Contract Got a Date, and it applies here with the tier unknown rather than known.

Start with the size of the thing being replaced. CoreWeave reported a revenue backlog of about $104 billion at June 30, 2026, and $9.7 billion of deferred revenue against $35 billion of debt.^[15] Every dollar of that backlog is a customer’s promise to pay; every dollar of deferred revenue is CoreWeave’s promise to deliver. Both are unsecured credit between AI companies, and the debt is sized on the first while the second sits ahead of it in the customer’s mind. One neocloud, one balance sheet. The whole market is built this way, because until this summer there was no other way to build it.

Now the replacement. When the seller in section 03 sells a strip on Nodal, the clearing house steps between it and every buyer by novation. The seller owes Nodal Clear; Nodal Clear owes the buyer; neither firm has a contract with the other. The price of that is margin. Nodal Clear sets initial margin by expected shortfall — the average loss in the worst tail of a distribution built from one to three years of returns plus chosen stress scenarios, at 99% for a single portfolio and 99.5% across the house — and collects variation margin twice every business day, with a one-day close-out assumption for customer accounts.^[6][8] If a member fails, the losses run down a waterfall the disclosure sets out in order.

The seller’s balance sheet changes in a specific way. It stops lending to its customer — the 80% receivable disappears, and with it the credit allowance and the covenant conversation about customer concentration — and it starts posting collateral to a broker. The prepayment goes too: nobody prepays a future. So the working-capital function that Compute Desk’s research identified inside the reserved contract has to be funded some other way, and the obvious way is the same bank that was lending against the receivable. This is the sense in which the lender is the natural first dealer: it already carries the name, it already sizes the advance on the contract, and a dozen of the clearing members that would carry the seller’s futures account are its own affiliates. The hedge and the loan can sit on one balance sheet, priced by one risk desk, which is roughly how the oil market financed itself for forty years.

What the clearing house cannot do is hold the delivery. In route C the seller is cleared until the swap and bilateral after it. The exchange-for-physical does not remove the delivery credit; it shortens the window — price risk is cleared from trade date to swap date, and delivery risk runs from swap date to the end of the contract under a Compute Clear guarantee. Compute Desk’s trading product lists escrow among its functions,^[11] which is the natural shape for a delivery guarantee on capacity, but nothing published says whether Compute Clear works that way, who backstops it, or what happens if a provider fails in month seven. That is the piece of the credit that has not been intermediated, and it is the piece a buyer of delivery cares about most.

Then there is the price of the intermediation, which is what section 06 is for. Compute Desk’s read of its own B200 prints puts a one-year reserved contract about 34% below its estimate of the price for delivery a year out, and attributes most of that gap to the value of committing — volume, prepayment, credit and the financing value of the longer commitment to the provider — rather than to an expected fall in price; deal size alone, it says, moves the rate 25 to 30%.^[14] Part of the gap, then, is the value of the prepayment to the seller. Part of it is the credit the seller extends. Part of it is the volume promise. And part of it, possibly, is simple disagreement about where the price is going. A cleared future strips the first two out by construction — it carries no prepayment and no counterparty credit — and it carries no volume promise either, which makes the futures price, once one trades, the price of delivery with the commitment discount removed. That is the number the seller has never had. The gap between the reserved rate and the strip, once the ledger prices the prepayment and the allowance, is the market’s bid for volume certainty — or its view of the forward. Either answer is worth knowing, and today neither is visible.

One seller, three routes, the same year of capacity. Route A is the reserved contract as it exists. Route B is the Nodal strip with the physical sold at market. Route C is the Architect strip exchanged for a delivery contract. Every input is movable and every default is labelled; the reserved multiple and the forward strip are read from Compute Desk’s published B200 curve,^[14] the prepayment and the tenor from their research, the margin model from Nodal Clear’s disclosure with the level assumed, since no compute margin rate is published. The dollars are illustrative. The shape is the point.

At the defaults the ledger says something a neocloud treasurer will recognise. The reserved contract nets about $3.57 a GPU-hour. The Nodal strip, with the racks 85% sold at market, nets about $4.29 on every hour the seller owns — but it would net the same as the reserved contract at about 71% sold, and below that the reserved contract wins. The whole difference is one question: can this seller keep these racks seven-tenths sold, at market, for a year, without a name on a contract? Some can. Most cannot yet, which is why the reserved discount exists. The Architect route nets more than either because it keeps the volume promise and drops the prepayment, and the number it is most sensitive to is the EFP basis — the one figure in the model that nobody has printed.

Then move the stress slider to +30% and read the cash call: on the defaults, about $12 million wired to the broker before a single hour has been invoiced at the higher price. The seller is right on the trade — its racks are now worth more — and short of cash on the day. That is not an argument against the strip. It is the working-capital function of the reserved contract showing up in its new address, and it is the conversation the seller’s bank needs to have had before the first margin call rather than after.

Everything above is built on announcements, a rulebook and a disclosure. The following are the gaps, in the order a seller would hit them.

1. Which transactions settle the contract. Compute Desk’s site lists SKU-level series described as United States one-year rental rates; the GX-administered benchmarks are family-level Hopper and Blackwell plus H100; Nodal’s release says the futures settle on “daily, blended” indexes.^[1][11] If the settlement series is built from one-year reserved contracts, the Nodal future is a future on a term rate — the price at which a year of commitment trades in the settlement month — rather than on the price of an hour that month. That is a different hedge object, closer to an interest-rate future than to a crude contract, and arguably the right one for a seller whose revenue is reserved contracts. It is also not the object Compute Desk’s published forward estimate describes, which is a price for delivery at a date. The contract specification will say; nothing public yet does.
2. A physical leg on the Nodal side. Nodal’s rulebook allows EFPs and its contracts can settle physically by specification.^[9] Compute Clear exists. The obvious next announcement is the one that connects them, and until it comes the seller who wants the power book beside its hedge and a path to delivery has to open two accounts on two clearing houses with no offset between them.
3. How Compute Clear guarantees anything. Escrow, backup capacity, a balance sheet, a bond: each is a different credit, and a buyer of delivery is entitled to know which one it is buying. ComputeConnect’s page says the network “guarantees delivery through Compute Desk’s ComputeClear platform”; the July release narrows it to guaranteeing “the operational integrity of compute delivery.”^[3][4] Operational integrity is not a financial backstop, and the word “guarantee” is doing more work than any published document supports.
4. The margin rate. Nodal Clear’s method is disclosed; the level for a contract with two years of history, a generational cliff and a settlement sample nobody outside Compute Desk can see is not. The ledger’s 20% is an assumption, and the seller’s cash planning is a function of it.
5. The basis tables. ComputeConnect says it will publish standard adjustments by SKU, memory and location. Those tables are the product: they turn a bilateral negotiation into a quoted differential, and they are what a seller marks its own racks against. They are not out.
6. Who is on the other side, by name. Section 04 lists the classes. No exchange has disclosed a committed participant in any compute contract, and the CFTC’s request for comment, open until October 20, asks in effect whether the cash market is deep enough to settle against at all.^[10]
7. Compute Desk’s own three tests. Its chief technology officer wrote in July that the market turns when bilateral deals are priced to indices, when cleared volume arrives, and when debt is underwritten against hedged books — and that none of the three has happened.^[13] That remains the honest scoreboard, and it is theirs.

None of this is an argument against the design. It is a list of what the design has not yet had to prove. The reserved contract was the market’s first risk-transfer instrument, and it worked because it transferred everything at once to a counterparty you could telephone. The stack described here proposes to transfer the pieces separately, to a clearing house you cannot, and to leave the last piece — the delivery — with a platform that says it will make good. If the pieces price the way Compute Desk’s forward estimate implies, the seller that learns to use both doors keeps a large part of a discount it is giving away today. If they price the way the reserved market does, the doors are open and nobody walks through them, and we will have learned that the discount was a forward view after all. Either way the number will be on a screen, which it has never been before.

This is a market-structure read of two announced products and the clearing house behind one of them, written from the seller’s chair. Compute Desk, Nodal and Architect were not contacted for it. Everything attributed to them is from their releases, product pages, rulebook and disclosures as retrieved on September 16, 2026, and from a LinkedIn post and chart Compute Desk published; where a statement is theirs rather than verified — the delivery guarantee, the investor list, the forward estimate — the text says so. The ledger is illustrative and every input is exposed. Where any of them publishes more than I found, the omission is mine and I will correct it in place.

• [1] Nodal Exchange and Compute Desk, “Nodal Exchange and Compute Desk to launch Compute Futures Contracts,” September 3, 2026. nodalexchange.com
• [2] Compute Desk, “Nodal partnership announcement,” September 3, 2026 — including the statements that the benchmarks are distributed on Bloomberg and Refinitiv and that Compute Desk’s backers include Anthropic, “one of the largest debt funds in the world,” hyperscaler CFOs and venture investors — Compute Desk’s own description. compute-desk.com
• [3] Architect Financial Technologies and Compute Desk, “Architect Financial Technologies Partners with Compute Desk to Launch Exchange-for-Physical Market for GPU Compute,” PR Newswire, July 8, 2026. prnewswire.com
• [4] Architect, ComputeConnect product page (“the only exchange-for-physical network that turns cash-settled compute futures into real, delivered GPU capacity”), retrieved September 16, 2026. architect.co
• [5] Architect Financial Technologies, “Architect Financial Technologies to Launch U.S. Futures Exchange for Compute and AI Commodities Following DCM Acquisition,” PR Newswire, May 28, 2026. prnewswire.com
• [6] Nodal Clear, LLC, “Principles for Financial Market Infrastructures Disclosure Framework,” v6.0, November 2025. nodalclear.com
• [7] Nodal Clear, clearing members list, retrieved September 16, 2026 (23 entries). nodalclear.com
• [8] Nodal Clear, “Margin Methodology,” retrieved September 16, 2026. nodalclear.com
• [9] Nodal Exchange Rulebook, 2026v3 — Rule 1.75 (Nodal Contract, “settled either financially or physically”), Rules 1.42–1.45 and 4.7 (Exchange for Related Position (EFRP); EFP defined at 4.7.1(a)). nodalexchange.com
• [10] CFTC, “CFTC Requests Comment on the Listing of Compute Derivatives Contracts,” Release 9286-26, August 19, 2026; 91 FR 54259, August 21, 2026; comments due October 20, 2026. federalregister.gov
• [11] General Index, “Compute Desk’s indexes are now regulated,” dateline August 13, 2026, posted August 17 (GX Benchmarks Ltd, FCA reg. 933348); Compute Desk product pages listing DESKH100, DESKH200, DESKB200 and DESKB300 as US one-year rental series and Compute Trader’s escrow function, retrieved September 16, 2026. general-index.com
• [12] David Lopez Mateos, “There is no GPU price hedge,” Compute Desk, May 5, 2026. computedesk.substack.com
• [13] David Lopez Mateos, “Financial Markets Will Unbundle the Neocloud,” Compute Desk, July 13, 2026. computedesk.substack.com
• [14] Compute Desk, LinkedIn post and chart “Reserved rates and the forward curve, B200,” September 2026, supplied to the author. The 0.72× and 0.54× reserved multiples, the ~34% gap to the forward, the 25–30% deal-size effect and the forward estimate behind the ledger’s strip are theirs. linkedin.com
• [15] CoreWeave, “CoreWeave Reports Strong Second Quarter 2026 Results,” August 11, 2026 — revenue $2,575M; revenue backlog approximately $104B as of June 30, 2026; deferred revenue $9,692M; total debt $35,068M. investors.coreweave.com
• [16] Jeffrey Kutler, “Can’t Get Enough ‘Compute’? Here Come the Futures,” GARP Risk Intelligence, August 28, 2026 (Don Wilson quotation). garp.org
• [17] The Fly, “Deutsche Bank exploring hedges for data center exposure, FT reports,” November 6, 2025. tipranks.com

Related on Kinetic Alpha