Markets for Compute

There is game theory in this. My best move depends on what I expect you to do, and you are making the same calculation about me. Every defended reservation and disputed queue position in this book has been a move in that kind of game. A price can make the cost of holding capacity visible. It does not remove bargaining power, unequal budgets, or the need for rules about what the organization must protect.

How to read this chapter

The chapter looks outward, then turns the same reasoning on the enterprise’s shared pools.

1. A market appears — the 2026 compute futures, their status, and what a hedge does to a bill.
2. Comparing negotiated agreements — what an outside price adds to a private one.
3. Why exchange trading has been difficult — the contract and index questions that come first.
4. What is available and what is announced — the instruments and what to ask before hedging.
5. If we all want the same GPUs — allocation inside the enterprise.
6. Where this conversation goes — what standardization leaves alone.
7. The question the curve answers — a traded price against a commitment, and a comparison to run now.

The question to carry: would a price improve a particular decision, and can the proposed market supply one we could trust?

A market appears

On August 11, 2026, CME Group and Silicon Data announced futures on the hourly rental price of Nvidia H100 and B200 GPUs, with trading to start October 5. NYMEX, the CME exchange that would list them, asked the CFTC to approve the contracts rather than certifying them itself, and on September 21 the CFTC’s Division of Market Oversight extended its review to November 9, citing “novel or complex issues.”¹²

The other announcements are no further along. ICE has published specifications for contracts on Ornn’s Compute Price Index, but its September 29 notice said October 5 would bring only reference data and hypothetical settlement prices; listing awaits the end of the CFTC’s comment period on October 20, and no date has been set.³ Nodal Exchange and Compute Desk say their contracts are scheduled for this year, and Architect’s planned exchange is pending regulatory review.⁴⁵ What did trade on October 4 were H100 perpetual futures on crypto venues, none of them a U.S.-registered futures exchange: about $1.5 million of volume in 24 hours and under $100,000 of open interest across seven venues, less than a tenth of a month’s rent on 512 H100s.⁶

The CFTC’s request for comment describes the market we already buy in. The Commission “preliminarily understands that compute markets are fragmented and price formation primarily occurs in opaque bilateral transactions” and suggests that compute “may not yet exhibit” the fungibility, standardization, and liquidity that derivatives markets usually rest on.⁷ A capacity team meets that problem at every renewal: we rarely know what anyone else pays for the same thing, and the same thing is hard to define.

Doubters already act through company shares, among them the contrarian Chapter 19 suggested studying: Michael Burry’s Scion Asset Management reported puts on one million Nvidia and five million Palantir shares at September 30, 2025. Form 13F values an option at its underlying shares, which is how the Palantir position became a widely reported $912 million bet; Burry later wrote that those puts cost $9.2 million.⁸ An equity put also mixes rental prices with everything else that moves a share price, which is part of the case for a contract on the rental price itself.

A futures contract is a standardized obligation cleared through an exchange. The proposed NYMEX H100 contract covers 730 GPU-hours and settles in cash against the month’s average of Silicon Data’s daily U.S. on-demand H100 index.⁹ If rental prices rise, the position gains about as much as our bill rises, to the extent our price moves with the index. The part that doesn’t is basis risk: the gap between what we pay for the hardware, region, terms, and dates we need and what the index measures.¹⁰

Open full-size diagram ↗

A hypothetical shows what the hedge would do. Suppose we expect to rent 512 H100s next April at market rates from neoclouds, the specialist GPU rental providers. Our rate has run about $0.40 per GPU-hour above the index, and an April contract trades at $2.80. Covering April’s 368,640 GPU-hours (512 × 720) takes 505 contracts, and if the basis holds, the hedge fixes our cost near $3.20 per GPU-hour, about $1.18 million:

The wider basis costs about $111,000 (368,640 × $0.30) that the hedge does not return. The falling case lands on plan, but its futures loss of about $184,000 (505 × 730 × $0.50) is paid in daily variation margin as the futures price falls, before any cheaper invoice arrives. A buyer renting from a hyperscaler starts with a wider basis; Silicon Data publishes a separate hyperscaler reading because that capacity “typically prices above neo-cloud and marketplace supply.”¹¹

Comparing negotiated agreements

What cloud capacity has lacked is a shared signal of what it costs. My claim is narrower than Adam Smith’s invisible hand, which in The Wealth of Nations concerns how people employ capital.¹²

An enterprise agreement can give a buyer a rate, a supplier a revenue commitment, and both some certainty. The trouble is comparison. Discounts, availability, and terms are negotiated customer by customer and kept confidential, so the same provider is a different seller to every buyer, and two prices for the same machine can include different support, cancellation, or payment terms.

The weakness shows in a disruption. Each agreement was written against one customer’s view of supply at signing and does not move when supply does. What a shortage costs is settled through renewals, expedited purchases, secondary markets, and reallocation, out of sight of other buyers. A public price would show that exposure earlier, and sometimes let buyers hedge part of it, without removing delivery risk or making private contracts irrational.

Why exchange trading has been difficult

Electricity offers a comparison, with limits. Its “missing money” problem asks whether energy-market revenue will pay for enough capacity to keep the system reliable when price caps and other design choices suppress scarcity revenue.¹³ A rate card is not a regulatory cap, but in a shortage a stable posted price can coexist with no available capacity, leaving the customer to delay work, find another supplier, or hold capacity defensively. The missing money shows up as missing work, a cost Volume I, Chapter 1 asked us to describe even when we cannot price it.

Open full-size diagram ↗

The second obstacle is the index. Silicon Data standardizes more than 10,000 rental records a day from over 40 providers with a proprietary model it recalibrates weekly; it restated its H100 history after a December 2025 methodology change and changed the H100 baseline again in September 2026.¹⁵ Ornn says it uses only printed transactions, and Compute Desk blends private transactions.¹⁶¹⁷ Each design has a weakness: providers can move the rates they post, a transaction index is only as good as the trades reported to it, and a modeled index can be revised. The CFTC asked what would stop a provider from moving a settlement index by adjusting a posted rate it administers.¹⁸ Before asking whether an index tracks our bill, I would ask what it is made of and who could move it.

The third obstacle is change. A rental index on one GPU type does not track the cost of an accepted answer after a model, serving system, or accelerator changes. Falling prices give sellers a reason to hedge, but they change which contracts attract trading.

Nor is provider spot an auction: AWS replaced bidding on EC2 Spot in 2017 with provider-set prices that change gradually.¹⁹ An independent venue could add a reference the provider does not set and a place to transfer exposure. Neither has yet been demonstrated.

What is available and what is announced

AWS Capacity Blocks, described in Volume I, Chapter 10, are the nearest thing to a forward market that a large provider runs. Two terms beyond those matter to Finance: Savings Plans and Reserved Instance discounts don’t apply, and the prepayment lands in the bill for the month of purchase, not of use.²⁷ A block is a delivery commitment with a price attached, not a hedge.

Compute Exchange offers forward contracts, delivery months ahead at a price fixed now on bilateral terms, under the heading “Hedge GPU exposure with futures.”²⁸ A bilateral forward can deliver machines but depends on the seller’s ability to deliver, with no clearinghouse behind a default; a cleared future delivers nothing, settles daily, and has a clearinghouse behind every trade.

The OneChronos–Auctionomics proposal, a market in bilateral forwards announced in July 2025 and still marked “coming soon” on October 4, 2026, asks whether an auction can represent the combinations a buyer needs: hardware, interconnect, and a usable window.²⁹ Its precedent is the 2017 broadcast incentive auction Paul Milgrom helped design, which paid television broadcasters $10.05 billion to give up or move channels and raised $19.8 billion selling the freed spectrum.³⁰ That shows technical constraints honored during a reallocation, not that a GPU pool needs such complexity.

Options make a different trade. A call on the index pays any rise above its strike in exchange for a premium, which need not be small. Buying a $3.00 call and selling a $3.60 call costs less but pays at most $0.60 per GPU-hour: if the index averages $4.20, everything above $3.60 is ours again.³² A plan rarely needs protection against every price, only against the increase that breaks it, provided someone has decided in advance what we do above the upper strike: defer work, move it, or pay. As of October 4, 2026, with the compute futures above still awaiting listing, there were no exchange-traded options on them either.

Holding capacity has incentives of its own, and I have responded to them myself. When I was a customer of a capacity team at a previous employer, I named my holding capacity in a cluster “emergency-response” — a name carefully chosen to make the allocation sound mandatory rather than convenient — and I made sure to run periodic load tests on that capacity so it would never show up on the unused-capacity report. I am not proud of it; I was rationally defending work I believed in against a reclamation process that could not distinguish idle capacity from insurance. An untouchable-sounding name works a little like one of Schelling’s focal points, a salient choice that coordinates expectations, though choosing it deliberately is also persuasion.³³ A price for holding capacity might change that decision; so might a credible promise that it could be recovered when needed. Charging for a reserve without fixing the return path leaves the reason to hold it. The same holds between companies: a buyer that doubts it can get GPUs back keeps paying for idle ones.

If we all want the same GPUs

Now turn the reasoning inward. Several divisions want the same accelerator pool, and a weekly meeting can no longer settle every collision. The organization needs an allocation mechanism, whether queues, executive decisions, service priorities, prices, or a combination, and calling all of them a market hides the choice.

Open full-size diagram ↗

One failure resembles Garrett Hardin’s tragedy of the commons: whoever holds the resource gets the benefit while others share the cost.³⁴ But a managed GPU pool is not an open-access commons, and Elinor Ostrom showed that users can govern shared resources with monitoring and enforceable rules of their own, not only through privatization or central control.³⁵ Chapter 18’s reserves and Chapter 20’s return conditions already address why a team fears releasing capacity; a market proposal should improve them, not pretend they are unnecessary.

The design space is wider than “post a price,” and most of it is familiar from this book:

These are alternatives, not a ladder. A division’s budget is not a measure of its value to the business, and critical work may need a protected allocation before any surplus goes to auction. I would start with an estimate of the next unit’s opportunity cost, what another use would gain or replacement supply would cost, informed by a matched external quote. In a linear model the shadow price is the local value of relaxing a constraint; an observed rental price is not automatically that value, and when decisions come in whole servers, a solver’s dual values need care before anyone treats them as prices.³⁸ The estimate can guide the people allocating capacity long before it is published to the teams asking for it.

Even a simple comparison can reveal an avoidable purchase. If one division has compatible committed capacity idle while another pays an incremental rate for the same work, the enterprise is paying twice for one need, provided the agreement and the account, region, and configuration restrictions would have allowed sharing.

An internal pricing proposal has to answer ordinary objections. Teams need budget certainty, which may favor a stable rate or a protected allocation over a price that moves hourly. Teams will shape requests to fit any rule, so keep the relationships of Chapter 17. I would start with one constrained pool, run the proposed allocation beside the current method, and examine which work changes and why before letting the mechanism control access.

The result I want is a change in demand: someone finds their work can move to a quieter interval, use a different resource, or release a reservation without sacrificing what it protects. The capacity team gains something too. When it allocates a shortage by judgment, it plays good cop and bad cop at once, and the division that loses hears a verdict on itself. A credible price does not soften bad news, but it depersonalizes it: the conversation moves from who deserves the GPUs to what to do about a price everyone can see.

Where this conversation goes

Standard contracts may make prices easier to compare, and portable deployment may make a competing offer easier to act on, but these are separate achievements. An H100 index does not make a workload portable, move its data, preserve its security controls, or reproduce its performance elsewhere, and a liquid future does not make a private service commitment assignable.

For a capacity team, this adds index, basis, margin, forward, and option to Chapter 20’s vocabulary of commitments: enough to explain an exposure to procurement, Finance, and Treasury, and to recognize a hedge that leaves it where it was. I am interested in this development because a credible shared reference could improve conversations that now rely on separate queues and privately negotiated rates, and it would give us one more piece of evidence to challenge.

The question the curve answers

If these markets mature, several practices in this book gain a comparison. Chapter 20’s negotiation can check a quoted rate against a traded one once hardware, location, service, and period match. Chapter 18’s reserve gains a financial question, whether the price risk can be hedged while the operating arrangement secures supply, and Chapter 22’s monthly review can follow both on separate lines.

Until trading starts, a planned contract supplies a specification, not a curve. Once there is a curve, it can show when our assumptions are out of line with traded prices, but its slope alone cannot tell us that shortages will persist or that efficiency will win: futures prices carry compensation for bearing risk as well as expectations, and a thinly traded distant month is weak evidence.³⁹ Before reading a quoted month as a forecast, I would ask who could transact at it, in what quantity, and how well that exposure matches our plan.

My manager’s description of the weekly Kobayashi Maru has stayed with me because it recognizes the job’s contradiction without suggesting we can administer it away. Customers need us to help make things possible, while resources, money, and time remain finite. A bill, a server count, a utilization chart, and now a futures curve are all production, in Rick Rubin’s sense: arrangements laid over the systems that drive consumption, and judged by the capability they are meant to buy. Strip them back and the thing is still the system. Finding a way through takes technical curiosity and mathematical care, but it also takes the willingness to keep listening after the first answer proves incomplete, including when it was our answer. That is what scale without surprise means in the end: understanding the system well enough that the important outcomes stop being surprising.

Notes and sources

1. CME Group, “CME Group and Silicon Data to Launch Compute Futures on October 5 to Unlock Hedging for the AI Economy”, August 11, 2026. Company announcement.↩︎

2. CFTC, product record 62544, “Approval Pending (90),” checked October 4, 2026; Division of Market Oversight, letter to CME Group, September 21, 2026, extending review of NYMEX Submission No. 26-370 to November 9, 2026, for “novel or complex issues.” An extension is not a disapproval.↩︎

3. Intercontinental Exchange, “ICE and Ornn to Launch GPU Compute Futures Contracts”, May 19, 2026: futures on Ornn’s Compute Price Index, which the companies describe as built “only from printed transactions”; launch is “subject to regulatory approval.” ICE Futures U.S., “New Compute Index Reference Data to Be Made Available on Monday, October 5, 2026”, exchange notice, September 29, 2026, checked October 4, 2026: listing “pending completion of the request for comment period” at the CFTC, with “no listing date or timeline” set; index data with hypothetical daily settlement prices in the interim; contract specifications attached. A statement of intent, not a listing.↩︎

4. Nodal Exchange, “Nodal Exchange and Compute Desk to launch Compute Futures Contracts”, September 3, 2026: contracts to settle against “Compute Desk’s daily, blended GPU price indexes,” described as based on “exclusive coverage of private compute transactions” (vendor description), “scheduled to be available this year, subject to regulatory compliance.” Checked October 4, 2026.↩︎

5. Architect Financial Technologies, “Architect Financial Technologies Partners with Compute Desk to Launch Exchange-for-Physical Market for GPU Compute”, July 8, 2026: the American Innovation Exchange is “forthcoming” and “pending regulatory review”; ComputeConnect would publish “standard basis tables for different SKUs, memory configurations, and locations.” Company announcement, checked October 4, 2026.↩︎

6. Loris Tools, “H100 Perpetual Futures”, checked October 4, 2026, 21:41 UTC: seven venues, $1.51 million of 24-hour volume, $80,800 of open interest, a $2.647 reference price, and 84.0% of volume on one venue. A third-party, unaudited aggregation of crypto-venue data; an order of magnitude only.↩︎

7. CFTC, Release 9286-26, August 19, 2026, and “Request for Comment on the Listing of Compute Derivatives Contracts”, 91 FR 54259, August 21, 2026, comments due October 20, 2026; quotations from its discussion of challenges; the posted-rate question is 2(b). The views are preliminary, and the request approves nothing.↩︎

8. Scion Asset Management, Form 13F information table for September 30, 2025: Nvidia puts on 1,000,000 shares, $186,580,000; Palantir puts on 5,000,000 shares, $912,100,000. SEC, Form 13F FAQ, question 44: option entries report the underlying security. The $9.2 million (50,000 contracts at $1.84, each covering 100 shares) is from Burry’s X post of November 13, 2025, quoted in Business Insider, November 13, 2025. Neither shows when or why the positions were opened.↩︎

9. NYMEX, Submission No. 26-370, August 11, 2026, a Regulation 40.3 request for approval: 730 GPU-hours per contract, cash-settled on the average of SD-H100 on-demand settlement prices over the contract month’s business days; U.S. index geography.↩︎

10. CME Group, “Learn about Basis: Grains”, checked October 3, 2026, and Self-Study Guide to Hedging with Grain and Oilseed Futures and Options, on basis (cash minus futures), long hedges, and margin. Applying grain mechanics to compute still requires checking the actual contract.↩︎

11. Silicon Data, GPU Rental Index Methodology Overview, updated August 27, 2026; index announcements 251203-1, December 3, 2025, restating H100 history from September 1, 2024 (estimated impact −6% to −4%), and 260911-1, an H100 baseline change effective September 25, 2026 (up to +4%, no restatement); H100 index page, checked October 4, 2026, with separate neocloud and hyperscaler readings. Vendor documentation.↩︎

12. Adam Smith, An Inquiry into the Nature and Causes of the Wealth of Nations, 1776, Book IV, Chapter II; the pin factory is Book I, Chapter I.↩︎

13. Michael Hogan, “Follow the Missing Money: Ensuring Reliability at Least Cost to Consumers in the Transition to a Low-Carbon Power System”, Regulatory Assistance Project, January 19, 2017, emphasizing scarcity pricing; Paul L. Joskow, “Capacity Payments in Imperfect Electricity Markets: Need and Design”, revised December 5, 2007, arguing for forward capacity obligations. Neither addresses compute.↩︎

14. Architect Financial Technologies, “Architect Financial Technologies Partners with Compute Desk to Launch Exchange-for-Physical Market for GPU Compute”, July 8, 2026: the American Innovation Exchange is “forthcoming” and “pending regulatory review”; ComputeConnect would publish “standard basis tables for different SKUs, memory configurations, and locations.” Company announcement, checked October 4, 2026.↩︎

15. Silicon Data, GPU Rental Index Methodology Overview, updated August 27, 2026; index announcements 251203-1, December 3, 2025, restating H100 history from September 1, 2024 (estimated impact −6% to −4%), and 260911-1, an H100 baseline change effective September 25, 2026 (up to +4%, no restatement); H100 index page, checked October 4, 2026, with separate neocloud and hyperscaler readings. Vendor documentation.↩︎

16. Intercontinental Exchange, “ICE and Ornn to Launch GPU Compute Futures Contracts”, May 19, 2026: futures on Ornn’s Compute Price Index, which the companies describe as built “only from printed transactions”; launch is “subject to regulatory approval.” ICE Futures U.S., “New Compute Index Reference Data to Be Made Available on Monday, October 5, 2026”, exchange notice, September 29, 2026, checked October 4, 2026: listing “pending completion of the request for comment period” at the CFTC, with “no listing date or timeline” set; index data with hypothetical daily settlement prices in the interim; contract specifications attached. A statement of intent, not a listing.↩︎

17. Nodal Exchange, “Nodal Exchange and Compute Desk to launch Compute Futures Contracts”, September 3, 2026: contracts to settle against “Compute Desk’s daily, blended GPU price indexes,” described as based on “exclusive coverage of private compute transactions” (vendor description), “scheduled to be available this year, subject to regulatory compliance.” Checked October 4, 2026.↩︎

18. CFTC, Release 9286-26, August 19, 2026, and “Request for Comment on the Listing of Compute Derivatives Contracts”, 91 FR 54259, August 21, 2026, comments due October 20, 2026; quotations from its discussion of challenges; the posted-rate question is 2(b). The views are preliminary, and the request approves nothing.↩︎

19. AWS, “Amazon EC2 Update — Streamlined Access to Spot Capacity, Smooth Price Changes, Instance Hibernation”, November 28, 2017, on the change from bidding to smoother provider-set prices.↩︎

20. AWS, “Amazon EC2 Update — Streamlined Access to Spot Capacity, Smooth Price Changes, Instance Hibernation”, November 28, 2017, on the change from bidding to smoother provider-set prices.↩︎

21. AWS, Capacity Blocks and pricing and billing documentation, checked October 3, 2026, for the discount and billing rules described.↩︎

22. Compute Exchange, marketplace site, checked October 3, 2026: quotes through a “live auction across verified providers,” a services agreement signed directly with the chosen provider, and forward contracts on bilateral terms. Vendor descriptions.↩︎

23. Auctionomics and OCX Group, “Auctionomics and OneChronos Partner on First Tradable Financial Market for GPU Compute”, Business Wire, July 29, 2025: price discovery “through bi-lateral forwards”; “first” is the companies’ claim. OneChronos, “Compute Markets”, checked October 4, 2026: “Coming soon.”↩︎

24. San Francisco Compute Company, “About”, checked October 3, 2026: clusters leased “on contracts that let our customers sublease.” Vendor description.↩︎

25. CFTC, product record 62544, “Approval Pending (90),” checked October 4, 2026; Division of Market Oversight, letter to CME Group, September 21, 2026, extending review of NYMEX Submission No. 26-370 to November 9, 2026, for “novel or complex issues.” An extension is not a disapproval.↩︎

26. Loris Tools, “H100 Perpetual Futures”, checked October 4, 2026, 21:41 UTC: seven venues, $1.51 million of 24-hour volume, $80,800 of open interest, a $2.647 reference price, and 84.0% of volume on one venue. A third-party, unaudited aggregation of crypto-venue data; an order of magnitude only.↩︎

27. AWS, Capacity Blocks and pricing and billing documentation, checked October 3, 2026, for the discount and billing rules described.↩︎

28. Compute Exchange, marketplace site, checked October 3, 2026: quotes through a “live auction across verified providers,” a services agreement signed directly with the chosen provider, and forward contracts on bilateral terms. Vendor descriptions.↩︎

29. Auctionomics and OCX Group, “Auctionomics and OneChronos Partner on First Tradable Financial Market for GPU Compute”, Business Wire, July 29, 2025: price discovery “through bi-lateral forwards”; “first” is the companies’ claim. OneChronos, “Compute Markets”, checked October 4, 2026: “Coming soon.”↩︎

30. FCC, FY 2017 Performance Report: $19.8 billion in incentive-auction revenue, including $10.05 billion for 175 winning broadcast bidders. Paul Milgrom, Nobel lecture. The comparison to compute is the author’s.↩︎

31. CME Group, “Learn about Basis: Grains”, checked October 3, 2026, and Self-Study Guide to Hedging with Grain and Oilseed Futures and Options, on basis (cash minus futures), long hedges, and margin. Applying grain mechanics to compute still requires checking the actual contract.↩︎

32. Options Industry Council, “Bull Call Spread (Debit Call Spread)”, checked October 3, 2026: the payout at expiry is capped at the difference between the strikes.↩︎

33. Thomas C. Schelling, The Strategy of Conflict (Harvard University Press, 1960); Nobel committee, 2005 prize background, on focal points. Their bearing on the cluster-name story is an interpretation.↩︎

34. Garrett Hardin, “The Tragedy of the Commons,” Science 162 (December 13, 1968): 1243–1248. Applied here by analogy.↩︎

35. Elinor Ostrom, Governing the Commons (Cambridge University Press, 1990); Nobel committee, “Economic governance: the organization of cooperation”, 2009. The enterprise application is an analogy.↩︎

36. OneChronos, “Periodic Auctions”, February 2026, checked October 3, 2026, describing its U.S. equities venue. Its enterprise use here is a proposal.↩︎

37. FCC, FY 2017 Performance Report: $19.8 billion in incentive-auction revenue, including $10.05 billion for 175 winning broadcast bidders. Paul Milgrom, Nobel lecture. The comparison to compute is the author’s.↩︎

38. Gurobi, “Linear Constraint Attributes”, on the Pi (dual value, or shadow price) attribute for continuous models, and “How do I retrieve the (dual) Pi values for a MIP problem?”, on integer models.↩︎

39. Sergey V. Chernenko, Krista B. Schwarz, and Jonathan H. Wright, “The Information Content of Forward and Futures Prices: Market Expectations and the Price of Risk”, Federal Reserve International Finance Discussion Paper, June 2004: the prices studied are “heavily affected by the market price of risk.”↩︎