Paper: arXiv 2611.20242

Authors: Dr Gabriel Borrageiro

Abstract

Capacity to run large language models is rented by the accelerator-hour, and their output is sold by the million tokens. Both goods now have published price indices, an over-the-counter forward has traded on one, a perpetual future on one is listed offshore, and CME and ICE have specified futures on rental indices that await the United States regulator. We ask what in these two markets can be hedged, by whom, and at what price. Three findings stand out. First, the price that exchange contracts settle on is not the price that drives operators’ revenue. Every contract references the published short-dated rental rate, while most revenue, and with it credit quality, comes from private multi-year contracts whose prices are largely undisclosed. Second, the token price and the rental rate are linked by an identity through throughput and margin, but they are different exposures, so a hedge on the rental rate does not hedge the cost of inference; at the measured capability of GPT-5 the cheapest list price fell by a factor of 17 in twelve months, and by far less at lower capability. Third, the demand for hedging already sits in credit agreements, as the years by which operators’ debt outlives their contracted revenue. Filings cannot measure that gap, because debt is scheduled year by year while contracted revenue rarely is, and the one loan that prices it settles the risk by cash prepayment against a model only the parties see. We design a reverse convertible note through which leveraged operators could buy protection against falling rental rates. Its coupon depends steeply on the annualised volatility of the rental rate, which the two public indices put between about 5% and 89% according to how each is built and the horizon measured; across that range the fair one-year coupon runs from 4.1% to 31.5%.

Complexity vs Empirical Score

  • Math Complexity: 6.0/10
  • Empirical Rigor: 5.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper introduces a novel financial instrument (reverse convertible note) for a nascent market (compute/inference for LLMs), demonstrating strong conceptual and mathematical development. While it identifies data limitations for full empirical estimation, it rigorously analyzes market structure and hedging needs. The overall contribution is significant for understanding risk transfer in this emerging asset class.

Research Flowchart

  flowchart TD
    A["Research Goal: What in these two markets can be hedged, by whom, and at what price?"] --> B{"Methodology: Analyze market structure, pricing links, and hedging demand"};
    B --> C["Data/Inputs: Published price indices, OTC forwards, perpetual futures, CME/ICE specs, GPT-5 capabilities, Credit agreements, Loan filings"];
    C --> D{"Computational Processes: Throughput/margin identity, Volatility calculation for rental rates, Reverse convertible note design"};
    D --> E["Finding 1: Exchange contracts settle on short-dated rental rates, not private multi-year contract prices driving revenue"];
    E --> F["Finding 2: Token price and rental rate are different exposures; rental rate hedge doesn't hedge inference cost (GPT-5 price fell 17x in 12 months)"];
    F --> G["Finding 3: Demand for hedging in credit agreements (debt outlives contracted revenue); Reverse convertible note designed (coupon 4.1%-31.5% based on 5%-89% volatility)"];

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper designs a reverse convertible note that lets leveraged operators buy protection against falling rental rates.

Source sentence
We design a reverse convertible note through which leveraged operators could buy protection against falling rental rates.

Assumes 5

  • The rental rate follows a stochastic process with volatility that can be described by a log‑normal model for a single maturity. model
    Source sentence
    Without jumps, the log rate at any one date is normally distributed under Equation (2), so Black’s formula holds with the volatility appropriate to that date, which is how Section 9 uses it.
  • No jump risk can be priced because there are no option markets for compute derivatives. model
    Source sentence
    And because no position in the futures reproduces a jump of unknown size, the market is incomplete [Tsay, 2010, sec. 6.9.1]: no trading strategy in the futures replicates an option payoff, and the premium for jump risk must be estimated from option prices, of which there are none.
  • Supply of accelerator capacity is effectively fixed over the horizon of a derivative because grid‑connection pace limits new capacity. market
    Source sentence
    Energised capacity sets the ceiling: it moves at the pace of grid connection rather than of manufacturing (Section 3), so over any horizon a derivative trades, supply is close to fixed.
  • Power cost is the floor of the rental rate, passing wholesale electricity prices through to rental contracts. market
    Source sentence
    Power cost sets the floor: electricity is the largest variable input to operating an accelerator, so wholesale power prices pass into rental rates, and a rental contract is in part a power contract.
Show 1 more
  • Committed multi‑year rental contracts are bilateral and undisclosed, so the visible short‑dated rental index does not reflect operators’ revenue exposure. data
    Source sentence
    Every contract references the published short-dated rental rate, while most revenue, and with it credit quality, comes from private multi-year contracts whose prices are largely undisclosed. Only the first is transparently priced, and this is the central data problem of the market.

Data

  • Public compute rental rate indices (rate‑card based and transaction‑based) · public sources market-prices public
    Source sentence
    Both goods now have published price indices, an over-the-counter forward has traded on one, a perpetual future on one is listed offshore, and CME and ICE have specified futures on rental indices that await the United States regulator. Fourth, we specify a process for the rental rate and show why it cannot yet be estimated: the two rental indices freely available to us, one built from rate cards and one from
  • Public token price indices for four laboratories · Ornn market-prices public
    23 August 2026 to 11 November 2026 · Anthropic, DeepSeek, OpenAI, Google
    Source sentence
    Ornn publishes both legs, a rental index for four accelerators and a token price index for four laboratories, though the token index is available to us only from 23 August 2026. Between 23 August 2026 and 11 November 2026 the H100 rate moved +12%, while token prices moved −43% for Anthropic, −60% for DeepSeek, +18% for OpenAI and +65% for Google, each measured between the means of its first and last seven
  • Silicon Data daily index for non‑10 hyperscaler on‑demand capacity · Silicon Data market-prices public
    Source sentence
    Each covers 730 accelerator-hours, one month of a single accelerator, and settles in cash against Silicon Data’s daily index for non-10 hyperscaler on-demand capacity, with monthly listings out to 36 months and a tick value of $7.30 [CME Group, 2026].
  • Epoch AI Capabilities Index · Epoch AI benchmark public
    Source sentence
    Epoch AI’s Capabilities Index combines scores from more than fifty benchmarks into one scale, anchored so that Claude 3.5 Sonnet scores 130 and GPT-5 scores 150 [Epoch AI, 2026b].
  • Operator credit agreements and debt schedules · public filings fundamentals public
    Applied Digital, CoreWeave, Nebius, others
    Source sentence
    nnot be measured from published information, because operators schedule debt year by year and, with few exceptions, report contracted revenue in coarse bands or as a single total; one credit agreement settles it by cash prepayment against a model that is not public. 8.3 The one mismatch the filings prove Applied Digital is the only operator in our sample whose debt schedule and contracted revenue schedule share a

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Exchange contracts settle on the short‑dated rental rate, but operators’ revenue and credit quality are driven by private multi‑year contracts priced at undisclosed rates. pricing-calibration
    Source sentence
    First, the price that exchange contracts settle on is not the price that drives operators' revenue. Every contract references the published short-dated rental rate, while most revenue, and with it credit quality, comes from private multi-year contracts whose prices are largely undisclosed.
  • Token price and rental rate are linked by an identity but represent different exposures; a hedge on the rental rate does not hedge inference costs, and at GPT‑5 capability the token price fell by a factor of 17 in twelve months. theoretical
    Source sentence
    Second, the token price and the rental rate are linked by an identity through throughput and margin, but they are different exposures, so a hedge on the rental rate does not hedge the cost of inference; at the measured capability of GPT-5 the cheapest list price fell by a factor of 17 in twelve months, and by far less at lower capability. In the first case a compute future does not move, so a seller of inference who
  • The demand for hedging already exists in credit agreements, where operators’ debt outlives contracted revenue, but filings cannot measure the gap. negative
    Source sentence
    Third, the demand for hedging already sits in credit agreements, as the years by which operators' debt outlives their contracted revenue. Filings cannot measure that gap, because debt is scheduled year by year while contracted revenue rarely is, and the one loan that prices it settles the risk by cash prepayment against a model only the parties see.
  • The two public rental indices disagree on volatility by a factor of about seven to twelve, leading to a one‑year fair coupon for the reverse convertible note ranging from 4.1 % to 31.5 %. pricing-calibration
    Source sentence
    Its coupon depends steeply on the annualised volatility of the rental rate, which the two public indices put between about 5% and 89% according to how each is built and the horizon measured; across that range the fair one-year coupon runs from 4.1% to 31.5%. Fourth, we specify a process for the rental rate and show why it cannot yet be estimated: the two rental indices freely available to us, one built from rate
  • CME and ICE futures on compute have been filed but not approved; regulatory review extends into early 2027, delaying trading. other figure from elsewhere in the paper
    Source sentence
    They were filed for the regulator’s approval, rather than self-certified, so they cannot take effect by default [New York Mercantile Exchange, 2026]. The Commodity Futures Trading Commission (CFTC) extended its review to 9 November 2026, citing novel or complex issues and its open request for comment on compute derivatives [Com-modity Futures Trading Commission, Division of Market Oversight, 2026, Commodity Fu-tures

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 16 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-17 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (71)

  1. Compute Futures: Contract Specifications (2026)
  2. Request for Comment on the Listing of Compute Derivatives Contracts (2026)
  3. CoreWeave Closes Landmark $8.5 Billion Financing Facility (2026)
  4. Regulatory Notice 10-09: Reverse Convertibles (2010)
  5. Architect Financial Technologies to Launch U.S. Futures Exchange for Compute and AI Commodities Following DCM Acquisition (2026)
  6. ICE and Ornn to Launch GPU Compute Futures Contracts (2026)
  7. AI Token Futures Market: Commoditization of Compute and Derivatives Contract Design (2026)
  8. The future of AI workloads (2026)
and 63 more
  1. Neocloud Market Forecast to Approach $400B by 2031 (2026)
  2. Meta Announces Joint Venture with Funds Managed by Blue Owl Capital to Develop Hyperion Data Center (2025)
  3. CoreWeave Closes $2.6 Billion Secured Debt Financing Facility (2025)
  4. Nebius raises $775 million in first secured debt financing (2026)
  5. Switch Raises $768 Million in Latest Data Center ABS Issuance (2026)
  6. Oracle announces Equity and Debt Financing Plan for Calendar Year 2026 (2026)
  7. Joining the dots between big AI (2026)
  8. Are credit rating agencies getting fed up with hyperscalers? (2026)
  9. How big is the open-model threat to AI hyperscalers? (2026)
  10. FalconX Executes First OTC Compute Forward Trade (2026)
  11. Architect Financial Technologies Partners with Compute Index Provider Ornn to Launch Exchange-Traded Futures on GPU and RAM Prices (2026)
  12. One Platform. Every GPU Market (2026)
  13. Token Forwards (2026)
  14. Pricing Compute Futures: Forward Curve and Volatility for a Non-Storable, Depreciating Commodity (2026)
  15. CoreWeave Secures $7.5 Billion Debt Financing Facility led by Blackstone and Magnetar (2024) — link
  16. Current Report on Form 8-K and Exhibit 10.1, Fifth Amendment to Credit Agreement dated 16 May 2024 (2025) — link
  17. CoreWeave Announces Upsize and Pricing of $2,000 Million of Senior Notes due 2030 (2025) — link
  18. CoreWeave Prices Upsized $3.7 Billion Convertible Senior Notes Offering due 2033 (2026) — link
  19. Current Report on Form 8-K: $1.59 Billion 7.000% Senior Secured Notes due 2031 issued by APLD ComputeCo 3 LLC (2026) — link
  20. Taking a position: AI debt frenzy gives birth to new CDS markets (2026) — link
  21. Volumes and Most Active Names in Credit Derivatives, March 2026 (2026) — link
  22. Markit CDX North American Investment Grade Index (2026) — link
  23. Report on Liquidity in Credit Default Swap Markets (2025) — link
  24. Intelligence per Watt: Measuring Intelligence Efficiency of Local AI (2026) — link
  25. Notice of 2026 Annual Meeting of Stockholders and Proxy Statement (2026) — link
  26. (Early) AI Compute Asset Pricing (2026) — link
  27. GPU Rental Price Indices (2026) — link
  28. OCPI Methodology (2026) — link
  29. Ornn AI Launches Token Price Index to Measure the Actual Cost of Inference Tokens from OpenAI and Anthropic (2026)
  30. Epoch Capabilities Index (2026)
  31. Updating the ECI methodology to use a Bayesian model (2026)
  32. Extension of Initial Review Period of NYMEX Submission No. 26-370 and Submission No. 26-370S for an Additional 45-Days (2026)
  33. New Compute Index Reference Data to be Made Available on Monday, October 5, 2026 (2026)
  34. Nodal Exchange and Compute Desk to Launch Compute Futures Contracts (2026)
  35. AX Perpetual Futures Exchange (2026)
  36. U.S. Regulator Slows Wall Street Push for GPU Futures (2026)
  37. Meta Announces New Strategic Venture with BlackRock to Develop Data Center in El Paso (2026) — link
  38. Financing Package: Blue Owl Capital/Beignet Investor’s US$27.3bn 23.6-year bond (2025) — link
  39. Annual Report on Form 10-K for the fiscal year ended 31 May 2026 (2026) — link
  40. Annual Report on Form 10-K for the fiscal year ended 30 June 2026 (2026) — link
  41. Annual Report on Form 10-K for the year ended 31 December 2025 (2026) — link
  42. Quarterly Report on Form 10-Q for the quarter ended 30 September 2025 (2025) — link
  43. Shareholder Letter, second quarter 2026 (2026) — link
  44. Preliminary Ratings Assigned to DataBank Series 2026-1 (2026) — link
  45. Preliminary Ratings Assigned to Flexential Issuer, Series 2026-1 to 2026-4 (2026) — link
  46. Ratings Assigned to Centersquare Series 2025-1 and 2025-2 (2025) — link
  47. Preliminary Ratings Assigned to Lohrasp Enterprise II LLC, Series 2026-1 (2026) — link
  48. Data Center e-Primer (2026) — link
  49. 17 CFR 15.03: Reporting levels (2026) — link
  50. About the Commitments of Traders Reports (2026) — link
  51. Submission No.\ 26-370: voluntary submission for product approval under CFTC Regulation 40.3(a), Silicon Data H100 and B200 Rental Index Futures (2026) — link
  52. GPU Rental Price Index (2026) — link
  53. Quarterly Report on Form 10-Q for the quarter ended 30 June 2026 (2026) — link
  54. Current Report on Form 8-K, with Exhibit 10.1 (Credit Agreement dated 7 August 2026) and Exhibit 99.1 (2026) — link
  55. Report on Form 6-K, Exhibit 99.2: Unaudited Condensed Consolidated Financial Statements for the quarter ended 31 March 2026 (2026) — link
  56. Quarterly Report on Form 10-Q for the quarter ended 30 June 2026 (2026) — link
  57. Quarterly Report on Form 10-Q for the quarter ended 30 June 2026 (2026) — link
  58. Quarterly Report on Form 10-Q for the quarter ended 30 June 2026 (2026) — link
  59. Options, Futures, and Other Derivatives (2015)
  60. Analysis of Financial Time Series (2010)
  61. Proof that properly anticipated prices fluctuate randomly (1965)
  62. Stock market prices do not follow random walks: evidence from a simple specification test (1988)
  63. Exotic Options Trading (2008)

0 of 71 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.