--- title: "The Contract Is the Collateral" source: https://www.second-breakfast.co/blog/the-contract-is-the-collateral site: "Second Breakfast" author: "Marc" date: "2026-10-20" tags: ["investing", "finance", "private-credit", "compute-markets", "data-centers"] reading_time_minutes: 37 word_count: 7362 --- # The Contract Is the Collateral > Compute already has a large, measured market: a GPU loan book is priced on who signs the offtake. The futures everyone talks about trade almost nothing, and the return sits with whoever packages and protects GPU credit. On 15 October 2026 the market repriced CoreWeave's own credit at distress levels. Its five-year credit default swaps traded above 1,100bp intraday for the first time, on S&P Global Market Intelligence data [reported by Reuters](https://money.usnews.com/investing/news/articles/2026-10-15/explainer-why-are-credit-default-swaps-unnerving-ai-investors), and its 9.75% notes due 2031 yielded 15.78% ([CNBC](https://www.cnbc.com/2026/10/15/coreweave-credit-default-swaps-hit-distress-levels-as-investors-sour-on-ai-debt.html)). GPU loans that fortnight priced on their contracts. Lambda borrowed USD 1.008B at 6.78% fixed on 1 October 2026 against GPUs serving two investment-grade customers ([Lambda](https://lambda.ai/blog/lambda-closes-1-billion-senior-secured-fixed-rate-financing)). Volta's unrated package, for a site where Anthropic has reportedly agreed to buy capacity for six years, needed SOFR + 650 at 95, [Bloomberg reported](https://www.businesstimes.com.sg/companies-markets/banking-finance/jpmorgan-goldman-boost-yield-volta-ai-loan-lure-investors) on 16 October 2026. The market priced the contracts, not the chips. The commodity market that's supposed to price compute had nothing to say. NYMEX's GPU1 and GPU2 futures still read "Approval Pending" at the CFTC, whose review runs to 9 November 2026 ([CFTC](https://www.cftc.gov/IndustryOversight/IndustryFilings/TradingOrganizationProducts/62544)). My thesis: through 2027, compute risk gets priced in the credit market, not on exchanges. Floating-rate GPU loans without investment-grade offtake paid **265bp** more over SOFR than investment-grade-backed loans across 2026's filed and company-released terms (my calculation, nine loans). My base case is a floating-rate gap above 200bp in every half-year through 2027, with US-listed compute futures under USD 250M of open interest at 31 December 2027 (zero today). So the return sits with whoever gets paid out of the credit market's existing cash flows: packaging loans for neoclouds without investment-grade customers, and lending into that wedge with residual protection and short amortization. Exchanges, indices and options sit near the bottom of my ranking because of time to liquidity rather than size. Two observations before the end of 2027 would change my mind. One is convergence: the half-year mean gap falls below 100bp, or a matched pair (same borrower, same half-year, matched tenor and recourse) prices within 150bp. The other is listed compute futures passing USD 250M of open interest. **Figure 1 — The GPU loan book is measured in tens of billions; listed compute futures hold nothing · Where compute risk is priced, as of 2026-10-19. CCIR is a single publisher formed in 2026, so I checked 16 of its rows against filings; commitments are not drawn balances. Kalshi's contracts are USD 1 binaries, so open interest equals maximum payout. · Sources: CCIR GPU-backed credit register (2026-10-03); CoreWeave, IREN, Nebius and Nscale filings and Lambda's release; CFTC product page for NYMEX GPU1 (2026-10-19); Kalshi (2026-10-19). Author's calculation.** - GPU-collateralized loans committed · CCIR register: USD 67.2B - **Extra spread without investment-grade offtake · 2026 floating-rate loans: 265bp** - Open interest · US-listed compute futures: 0 contracts - Payout open · Kalshi compute binaries: about USD 186,000 In my recent post, [Paid to Wait](https://www.second-breakfast.co/blog/paid-to-wait), I made the case for a private credit fund that lends behind the build, and wrote that the contract is the filter, not the return. One layer down, at the chips, the contract is also the collateral. Where I sit: I lead capital formation at PL Capital, which invests in decentralized compute and crypto rails, and I'm Demo Day Lead at Protocol Labs. PL Capital [backed](https://www.linkedin.com/posts/protocolvc_activity-7480695429689430016-zzOT) Prime Intellect's founders from the start, and Protocol Labs [led Compute Labs' 2024 pre-seed](https://www.theblock.co/news/deals/2024-07-03-solana-based-gpu-tokenization-protocol-compute-labs-funding-token-valuation-303398); Compute Labs launched a financing platform on 15 October 2026. So I have reasons to want compute marketplaces, onchain credit and packaging to matter. Packaging still ranks first on my scoring, and I'd rather say so here than let the ranking look neutral. Onchain credit ranks sixth of thirteen. I've kept every company claim to public, dated sources. Three claims carry the piece: 1. **Compute is priced on the offtaker.** Across nine 2026 floating-rate GPU loans, whether the offtaker is investment-grade accounts for about **~94%** of the spread variance; the signing date, under 1% within 2026 (across vintages it matters more). 2. **Lenders buy the contract and leave the chip.** CoreWeave's newest GPU loan advances **70%** of cost against six-year book value (its schedule isn't public), with no appraisal. Resale risk stays with the equity; lenders share renewal risk once a loan outlasts its contracts. 3. **The commodity market has no second side.** The proposed futures settle on on-demand indices near **USD 2.8** an hour, while lenders hold multi-year contracts that no series of realized prices tracks. ## One borrower, one day, two prices GPU loans are sized and priced on who signs the offtake contract, and the cleanest proof is two years old. CoreWeave's delayed-draw term loan DDTL 2.0, dated 16 May 2024, set two prices for one borrower on one day: SOFR + 6.00% to 6.50% against investment-grade contracts and SOFR + 13.00% against the rest ([8-K](https://www.sec.gov/Archives/edgar/data/1769628/000119312525227562/d910811d8k.htm)). Its 2025 DDTL 2.1 tranche let CoreWeave borrow 90% of capex against GPUs serving Microsoft but only 75% against GPUs serving OpenAI ([amendment](https://www.sec.gov/Archives/edgar/data/1769628/000119312525227562/d910811dex101.htm)). **Figure 2 — In 2026, GPU loans backed by investment-grade offtake priced about 265bp tighter than loans without it · Margin over SOFR at signing, in %, for GPU-collateralized floating-rate loans. Plain rows are 2026 signings on filed or company-released terms; italic rows are the 2024 vintage and Volta's reported price. Fees, discounts, tenor and recourse are not normalized. · Sources: CoreWeave 8-Ks (2025-10-02, 2026-03-31, 2026-05-18, 2026-08-10); IREN 8-K (2026-06-01); Nebius 6-K (2026-07-17); Nscale S-1 (2026-09-18); Lambda release (2026-08-27, priced 2026-08-12); Bloomberg via The Business Times (Volta, 2026-10-16). Author's calculation.** - CoreWeave DDTL 4.0 · Meta contract, non-recourse: 2.25 - IREN Hardware 3 · Microsoft contract: 2.25 - Nscale Ward County · rated Baa1: 2.375 - Nebius · customer investment-grade per the company: 2.50 - Lambda term loan B · investment-grade offtaker, unnamed: 3.00 - CoreWeave DDTL 5.0 · two non-investment-grade customers: 4.50 - Nscale GPU Financing Facility · offtaker undisclosed: 5.00 - **CoreWeave DDTL 5.5 · contracts averaging three years: 5.50** - Nscale Iceland · customer redacted: 5.50 - *Volta · unrated, single anonymous-sourced report: 6.50* - *CoreWeave DDTL 2.0 · investment-grade contracts, 2024, midpoint of 6.00 to 6.50: 6.25* - *CoreWeave DDTL 2.0 · non-investment-grade contracts, 2024: 13.00* My calculation from the filings: across those nine loans, investment-grade offtake alone explains about 94% of the variance in spread. Loans with it average 2.48% over SOFR and loans without it 5.13%, a gap of 265bp. Drop Nebius and Lambda, whose investment-grade customers no filing names, and it widens to about 283bp; drop the two whose offtakers are merely undisclosed and it's about 252bp. Volta's reported price would take it to 292bp, but it rests on one anonymous-sourced report, so I leave it out. Margins aren't all-in yields, but disclosed fees and discounts tilt the same way. Nscale's undrawn fees and the Iceland facility's repayment multiple sit on the non-investment-grade side, while Lambda's 99.5 issue price adds roughly 10bp to 20bp a year on the other (my estimate). Normalizing would more likely widen the gap. Fixed-rate deals point the same way at a smaller gap. On CCIR's figures, investment-grade-backed GPU debt priced 1.61 to 2.00 points over Treasuries, against 2.96 to 3.72 for Valor's 2026 financings of GPUs leased to xAI, about 150bp apart (my calculation). Valor's first such financing priced at 6.28 points in November 2025; after SpaceX acquired xAI in February 2026, per CCIR, the spread fell with the lessee's credit ([CCIR](https://ccir.io/credit)). One publisher and a different benchmark keep these out of the 265bp. The awkward detail points the right way: the cheapest CoreWeave loan has the least CoreWeave in it. DDTL 4.0 lent USD 8.5B to a subsidiary holding a Meta contract, with no recourse to the parent beyond a bad-acts guarantee, at SOFR + 2.25% on its floating tranche, rated A3 by Moody's and A (low) by Morningstar DBRS ([8-K](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000129/crwv-20260330.htm)). DDTL 5.0 and 5.5 carry full parent guarantees and pay 4.50% and 5.50% ([DDTL 5.5 8-K](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000357/crwv-20260807.htm)). Lenders priced Meta's signature above CoreWeave's, though [Fitch keeps](https://www.fitchratings.com/research/structured-finance/fitch-rates-coreweave-compute-acquisition-co-viii-loans-a-sf-outlook-stable-01-04-2026) DDTL 4.0 within two rating categories of CoreWeave. Two same-issuer pairs exist inside one half-year: CoreWeave's DDTL 4.0 and 5.0, 225bp apart, and Nscale's Ward County and Iceland facilities, 312.5bp apart ([S-1](https://www.sec.gov/Archives/edgar/data/2110365/000119312526395475/ck0002110365-20260918.htm)). Neither is matched on recourse. Every 2026 loan without an investment-grade customer carries a parent guarantee, so the filings can't separate offtaker credit from recourse. The loans with more parent support pay more. S&P says it outright: its preliminary BB on IS AN03, a Crusoe-sponsored GPU loan of up to USD 509M in Iceland, notes that coverage alone would support bbb+, then caps the rating at bb for the offtaker's and the data-center operator's credit ([S&P, 11 September 2026](https://www.spglobal.com/ratings/en/regulatory/article/-/view/type/HTML/id/3626094)). Four notches for the signatures. The cross-section has a weakness I'd rather name: across 2023 to 2026 the signing date explains more spread variance than credit support does (R-squared about 0.49 to 0.58 against about 0.35, on 13 to 14 loan legs), because investment-grade spreads at CoreWeave fell about 400bp from 2024 to 2026. The sort is the structure; the level is the cycle, and October showed it can reverse. ## The lender buys the contract and amortizes the chip CoreWeave's DDTL 5.5, signed 7 August 2026, is the most detailed filed example, and its [credit agreement](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000357/ex101creditagreement.htm) underwrites the contract at every turn. | DDTL 5.5 clause | What it says | |---|---| | Advance | Up to 70% of funding-date capex; capex tied to uncontracted GPU capacity doesn't count | | Collateral value | Six-year straight-line book value (the schedule is omitted from the filed exhibit); no appraisal, market-value test or loan-to-value covenant | | Cash trap | Projected contract value below 2.40 times the loans, at least USD 1.68 per USD 1 of capex at funding (my calculation, if the redacted definition is gross contracted revenue) | | Renewal | Prepay if an expiring contract isn't renewed or replaced at levels that hold 1.35x coverage and the 2.40x contract ratio | | Hedging | Interest rates on at least 95% of projected principal; a GPU rental hedge would need the Required Lenders' consent, and none is required | The investment-grade version goes further: DDTL 4.0 sizes the loan at 1.20 times coverage of Meta's contract cash flow, which CoreWeave says implies up to about 102% of cost ([investor deck](https://s205.q4cdn.com/133937190/files/doc_presentations/2026/Mar/2026-03-DDTL-4-Overview.pdf)). A loan that can exceed the hardware's cost isn't underwritten on the hardware. [Fitch's A-sf case](https://www.fitchratings.com/research/structured-finance/fitch-rates-coreweave-compute-acquisition-co-viii-loans-a-sf-outlook-stable-01-04-2026) assumes 99% of contracted revenue is realized, and its named sensitivities are a 7% revenue haircut (to BBBsf) and a counterparty downgrade, not the value of the chips. S&P, rating IS AN03, says matching the debt's tenor to the compute agreement "removes the project's reliance on residual asset values". Maturities line up with the contracts: IREN's Microsoft-backed debt matures when Microsoft's final fee is paid ([8-K](https://www.sec.gov/Archives/edgar/data/1878848/000114036126023427/ef20075181_8k.htm)), and [Lambda's term loan](https://lambda.ai/blog/lambda-closes-926-million-senior-secured-term-loan-b-facility) amortizes fully by 31 December 2030, "aligned with the contracted cash flows and useful life of the underlying GPU infrastructure". A reserve-based oil and gas lender would notice what's missing: a hedge forced on the collateral's output, like Talos's 50% of projected producing oil for four quarters ([Talos](https://www.sec.gov/Archives/edgar/data/1724965/000119312526290589/d14154dex101.htm)). GPU lenders hedge the rate and lean on the offtaker. The first cracks are in the newest paper. DDTL 5.5 runs about five years against customer contracts that average about three ([CoreWeave release](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000357/ex991pr.htm)), and Fitch's rating case assumes CoreWeave renews or replaces them "at favorable GPU lease rates" ([Fitch](https://www.fitchratings.com/research/corporate-finance/fitch-assigns-coreweave-new-ddtl-bb-rating-affirms-idr-at-bb-outlook-positive-16-07-2026)). Multicoin, a venture investor, reads DDTL 5.5 as lenders "explicitly underwriting the residual value" ([Multicoin](https://multicoin.capital/2026/10/08/compute-capital-markets/)). I read it, as the commentator Dave Friedman [does](https://davefriedman.substack.com/p/what-makes-gpu-revenue-bankable), as lenders charging for two years of renewal exposure behind a parent guarantee, 1.35x minimum cover and a prepayment trigger. Follow one server through its life and the split is plain ([Figure 3](#fig-3)). One GPU server, five stages: the lender holds credit risk through the contract and shares renewal risk once a five-year loan outlasts three-year contracts, fenced by the 1.35x and 2.40x prepayment trigger; the sponsor's equity holds resale and re-contracting price, where a used H100 sold for a median USD 11,700 against USD 26,880 asked. ## Nobody marks the tail No GPU loan I found runs a market-value test, because there is no market price a lender could defend in a workout. How long a GPU lasts depends on who's counting. | Who's counting | The answer | |---|---| | CoreWeave's accounts | Six years, up from five in 2023 ([10-K](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm)) | | Meta's accounts | 5.5 years for servers from 2025, cutting that year's depreciation by USD 2.92B ([10-K](https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm)) | | Amazon's accounts | Five years for a subset, down from six ([10-K](https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm)) | | Bank underwriters | Usually three to four years, an Impax portfolio manager told Reuters ([Reuters Japan, 2026-10-02](https://www.reuters.com/jp/markets/japan/KGPXPEEMAJOUDAPZ6SLQ3X3IMM-2026-10-02/)) | | Posted neocloud rents | About 21% lower for each year of chip age, across five generations, on 2026-09-28 ([CCIR](https://ccir.io/research/gpu-age-curve)) | | H100 prices in 2026 | Up: Ornn's H100 value went from USD 1.76 on 2026-03-31 to USD 2.97 on 2026-08-31, and SemiAnalysis's one-year contract index from USD 1.70 in October 2025 to USD 2.35 in March 2026 ([SemiAnalysis](https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity)) | A lender who wanted a number would get a range, and the used market can't settle it. CCIR's record of executed resales from July 2023 to September 2026 adds up to USD 29.4M across 13,555 units ([CCIR hardware](https://ccir.io/hardware)). Over the 90 days to 12 October 2026, a used H100 SXM5 sold for a median of USD 11,700 on seven sales, against a median ask of USD 26,880. The only filed fleet sale I found, WhiteFiber's 126 H200 units (almost certainly eight-GPU servers), fetched about USD 26.1M against USD 24.3M of book value ([10-Q](https://www.sec.gov/Archives/edgar/data/2042022/000121390026056212/ea0288695-10q_whitefiber.htm)). A gain, on one trade. Old chips keep earning: in Bandi and Su's data for 18 April 2026, an A100 earned USD 4.54 per H100-equivalent hour against USD 2.51 for an H100 ([arXiv](https://arxiv.org/abs/2607.12156)). So the loans amortize inside the offtake. Ankur Patel, a partner in Ares's credit group, told the [Fixed + Floating podcast](https://www.fixedandfloating.com/p/who-pays-for-the-gpus) that one of a lender's main goals is "to make sure you have as little residual value at the end". Auto residuals became bankable on a tape and a record: [Manheim's index](https://site.manheim.com/en/services/consulting/used-vehicle-value-index.html) rests on more than five million used-vehicle transactions a year, and [Fitch's criteria](https://assets.fitchratings.com/downloadFile?reportType=report&sfReport=false&slug=structured-finance%2Fconsumer-abs-rating-criteria-residual-value-addendum-11-10-2024) start from an originator's residual-setting record. GPUs have neither, so the first securitizations take them as small slices: Bloomberg reported GPU loans at about 15% and 20% of two equipment asset-backed deals, per people familiar with them ([Equipment Finance News](https://equipmentfinancenews.com/news/lender-operations/gpu-financing-filters-into-abs-with-stonebriar-wingspire-deals/)). ## NVIDIA buys the hours, not the chips As buyer of last resort, NVIDIA writes renewal and utilization backstops. It doesn't write GPU residual values, and at scale nobody else does either. | Backstop | What it covers | What it leaves out | |---|---|---| | NVIDIA's guarantees, USD 108.5B of maximum exposure ([10-Q](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm)) | USD 105.0B at SB Energy's PORTS site, made up of data-center, power and transmission costs ([guaranty](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda2027q2ex101.htm)), which OpenAI has agreed to reimburse ([8-K](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000069/nvda-20260817.htm)); USD 3.5B of land, power and shell for AI clouds | GPUs | | NVIDIA's AI cloud agreements, about USD 36B of purchases, typically over six years | Rental demand; the clouds may resell the capacity at better rates and share the revenue | Default: its USD 6.3B CoreWeave order can be terminated if either side sits in insolvency proceedings for more than 90 days ([8-K](https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm)) | | NVIDIA's residual-value support, up to 25% of an opportunity ([NVIDIA blog](https://blogs.nvidia.com/blog/nvidia-ai-factory-compute/)) | A slice of the residual, project by project | The rest; its USD 500B financing platforms "remain subject to execution of the final agreements" ([NVIDIA](https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Partners-With-Apollo-BlackRock-Blackstone-Brookfield-Goldman-Sachs-and-KKR-to-Establish-AI-Compute-Infrastructure-Financing-Platforms-to-Mobilize-Over-500-Billion-of-Third-Party-Capital/default.aspx)) | | Google behind Fluidstack's leases, about USD 1.4B at Cipher | Lease payments, paid for in penny warrants Cipher booked at USD 544.5M, about 39% of the backstop ([Cipher 10-Q](https://www.sec.gov/Archives/edgar/data/1819989/000181998925000112/cifr-20250930.htm)) | GPUs; at TeraWulf, a lease ended for construction delay ([10-K](https://www.sec.gov/Archives/edgar/data/1083301/000108330126000031/wulf-20251231.htm)) | | Insurers: Barkr's valuation schedule; American Compute (launched 2026-10-17, no bound policy or limit disclosed), on A- rated capacity from Incline | The shortfall below a forward schedule, or below an insured floor: for example, insure 20%, sell at 5%, collect 15% ([American Compute](https://www.amcompute.com/blog/gpu-residual-value-insurance-incline-accelerant)) | Breadth: Munich Re calls its role a guarantee on Barkr's model, narrower than residual-value cover ([The Insurer](https://www.theinsurer.com/ti/analysis/gpu-residual-value-insurance-paper-used-to-tackle-ai-financing-risk-2026-08-26/)) | Lloyd's insured the computer lessor Itel against early lease returns from 1975; cheaper computers gutted its lease values, and it filed for bankruptcy in 1981 with USD 1.3B of debt ([UPI](https://www.upi.com/Archives/1981/01/19/Itel-Corp-a-computer-leasing-firm-troubled-by-financial/7749348728400/)). Travelers' Gulf unit wrote vehicle residual-value insurance from 1996 to 2001 for USD 157M of premiums, then took USD 239M of charges in the first quarter of 2003. Every policy on one generation is the same put. So the gap stays open in two places: the hardware tail, and the renewal tail for neoclouds NVIDIA doesn't back. That's the adverse-selection tranche, the risk the cheapest capital declined. It pays only with rated capacity and short tenors, and only to someone holding data nobody else has. ## Plenty of indices, no second side The derivatives market is announced or tiny, and the contracts on file hedge the wrong exposure. | Instrument | Status, 2026-10-19 | Settles on | Measured activity | |---|---|---|---| | NYMEX GPU1 (H100), GPU2 (B200) | Pending; CFTC review to 2026-11-09 | Silicon Data on-demand index, monthly average | None | | ICE HPR (H100), BKL (B200) | Announced; listing after the comment period, no date | Ornn's Compute Price Index (OCPI), monthly average | None | | Architect; Nodal with Compute Desk | Announced; contracts pending review | Compute Desk indexes | None | | Kalshi compute series | Live; USD 1 binaries since 2026-03-02 | Ornn | 5.56M contracts lifetime; 185,954 open | | Lighter H100 perpetual | Live, offshore, closed to US persons | Ornn, as oracle | USD 41.52 in 24 hours | | OTC forwards | Live, bilateral; Wintermute's [first H100 forward](https://www.linkedin.com/posts/compute-desk_congratulations-to-compute-desk-partner-pluto-activity-7480732064108183552-3Wa5) on 2026-07-08 | Bespoke | Not disclosed | Kalshi's contracts are [USD 1 binaries on an index level](https://assets.kalshi.com/regulatory/product-certifications/INDEXVALUE.pdf), so 5.56M contracts is about USD 5.56M of maximum payout. Lighter's only liquid perpetual, the "DRAM" market, with about USD 2.17M of volume in the 24 hours to 19 October 2026, [tracks a memory-stock ETF](https://docs.lighter.xyz/trading/real-world-assets-rwas/market-specifications). Strip those out and measured activity rounds to zero. **Figure 4 — The futures would settle on the on-demand hour; lenders carry multi-year contract value, and no transaction series prices it · H100 price per GPU-hour by source, 2026-10-18 and 2026-10-19. The filed futures settle on on-demand indices; lenders and operators carry multi-year contract value. Silicon Data also publishes a model-built term curve (not shown), and no series of realized term-lease prices exists. CCIR's two-year rate is one publisher's modeled estimate. · Sources: CCIR (modeled contract rate, 2026-10-18); Vast.ai marketplace median of asks; Ornn OCPI settle (2026-10-18); Silicon Data SDH100RT and hyperscaler readings (2026-10-18); Computable via Financial Datasets (08:20 UTC, 2026-10-19); Lambda and CoreWeave on-demand list prices (2026-10-19).** - *CCIR · two-year contract rate, modeled: 1.40* - Vast.ai marketplace · median of asks: 1.73 - **Ornn OCPI · executed on-demand, settles ICE: 2.81** - Silicon Data neocloud · settles NYMEX: 2.83 - Computable · posted rate cards: 3.81 - Lambda · on-demand list: 3.99 - CoreWeave · on-demand list: 6.16 - Silicon Data · hyperscaler reading: 7.15 The indices disagree by method. Computable averages posted rate cards, while Ornn and Silicon Data use executed or standardized rentals and agreed within USD 0.02 on 18 October 2026. My correlation of Ornn's and Computable's daily changes over the 28 days to 18 October 2026 is 0.03. A benchmark regulator would ask about conflicts. Silicon Data, whose index would settle NYMEX's contracts, shares a chief executive, Carmen Li, with the Compute Exchange marketplace, per the companies and press reports. DRW helped found both ([CME](https://www.cmegroup.com/media-room/press-releases/2026/5/12/cme_group_and_silicondatapartnertolaunchfirstcomputefutures.html)). Ornn's [methodology](https://data.ornn.com/methodology) says its affiliates trade physical capacity. The bigger gap is structural. Both filed US contracts settle on on-demand prices by rule, while lenders and neoclouds hold multi-year contract value. Silicon Data does publish a [daily term curve](https://www.silicondata.com/products/forward-curve) from one to 36 months, which CME has [distributed](https://www.cmegroup.com/notices/electronic-trading/2026/10/20261012.html) as non-tradable prices since 15 October 2026. But it blends posted rates with transactions, keeps its history on an enterprise tier and covers only neoclouds for the H100. The term readings disagree even on the sign. CCIR models a two-year H100 contract at USD 1.40 an hour, half the on-demand print, while SemiAnalysis's one-year index read USD 2.35 in March 2026, when Ornn's on-demand value was USD 1.76. No series of realized lease prices exists to check them against. Brett Harrison, who runs Architect, [argues](https://x.com/BrettHarrison/status/2072327852498797048) that imperfect cross-hedges are how hedging works in many major US markets. Even at a 0.7 correlation, a minimum-variance hedge leaves 71% of the standard deviation ([Ederington's definition](https://www.dallasfed.org/~/media/documents/research/papers/1980/wp8001.pdf)), and nobody has measured even that for compute. The regulator agrees. The CFTC's [request for comment](https://www.federalregister.gov/documents/2026/08/21/2026-17163/request-for-comment-on-the-listing-of-compute-derivatives-contracts) says it preliminarily believes compute "may not yet exhibit" the fungibility, standardization and liquidity behind a commodity market. A day before comments closed, the [docket](https://www.regulations.gov/docket/CFTC-2026-1850/comments) held 39 comments, none from CME, ICE, Silicon Data, Ornn, Kalshi, NVIDIA, a hyperscaler or a large neocloud. Yet the legal default runs toward listing: under [17 CFR 40.3](https://www.ecfr.gov/current/title-17/chapter-I/part-40/section-40.3), GPU1 and GPU2 are deemed approved at the end of 9 November 2026 unless the CFTC issues a reasoned notice that it won't approve them, NYMEX agrees to a further extension, or an amendment restarts the clock. A listing is likely. Liquidity is another question. Compute in 2026 looks like electricity in 1996. NYMEX listed power futures before the market had a transparent spot price. Palo Verde open interest stood at 8,353 contracts on 24 March 1998 ([CFTC data](https://publicreporting.cftc.gov/d/6dca-aqww)), and NYMEX delisted all six of its power contracts in February 2002 ([NGI](https://naturalgasintel.com/news/nymex-ready-to-begin-power-contracts-again-starting-with-pjm/)), which [Moulton's study](https://www.sciencedirect.com/science/article/abs/pii/S0140988304000970) blames on erratic hedge ratios. Power futures worked once system operators produced locational prices; ICE's PJM Western Hub real-time peak mini had 395,915 contracts open on 13 October 2026. Compute has neither a market operator nor a transaction-based benchmark, and [bandwidth in 2000](https://www.wired.com/2001/11/enron-a-bandwidth-bloodbath/) and CME's cash-settled [uranium future](https://www.reuters.com/business/energy/cme-seeks-revive-uranium-trading-with-physical-futures-launch-sources-say-2026-05-07/) show the other two ways a contract fails: a heterogeneous product, and buyers who don't need the exchange. The base rate is harsh: 68% of exchange-traded derivatives launched worldwide this century had failed by 2022 ([Applied Finance Letters](https://doi.org/10.24135/afl.v11i1.520)). In [Bekkerman and Tejeda's study](https://scholarworks.montana.edu/bitstreams/9831f9ba-54e7-4283-8e12-a3762cc0c225/download), the best predictors of survival are a large cash market and no alternative hedge. Compute has the first; it fails the second, because buyers already hold a hedge. CoreWeave earned 98% of its revenue in the second quarter of 2026 from committed take-or-pay contracts ([10-Q](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm)), and the multi-year fixed-price contract is that hedge. That also thins the second side. The natural longs of the rental price are many: neoclouds on uncontracted capacity, lenders through renewal, NVIDIA through its purchase commitments. The natural shorts are inference clouds selling fixed per-token prices on floating GPU costs, as Wafer's Eugene Ye [points out](https://wafer.substack.com/p/lets-talk-about-trading-compute), and labs buying spot. A market with one natural side needs a dealer. That's how the first firms are entering: Jump hired Itô's founding team to run a compute desk that starts with mid-market bilateral contracts, and DRW reportedly plans to trade through three teams ([Crypto Briefing](https://cryptobriefing.com/jump-trading-drw-compute-power-contracts/)). Cash-settled and physically delivered designs alike hedge multi-year contract value with instruments built on the on-demand hour. Most of the seed writing comes from investors and founders who invest in the answer they argue for, so I've weighed each argument on documents rather than on its author. ## Thirteen ways to get paid, ranked I set four tests before scoring anyone: existing cash flows, rated capacity or balance sheet for whoever holds the tail, transaction data others lack, and no need for exchange liquidity before 2028. Start with where money already changes hands ([Figure 5](#fig-5)). The pools overlap (the wedge sits inside the register), so they can't be added. **Figure 5 — At full draw, the register would pay lenders USD 1.5B to 3.7B a year in spread; the whole 2028 derivatives fee pool tops out at USD 245M · Annual pools in USD millions, all modeled. Register spread assumes every CCIR commitment is drawn, across the range of 2026 floating-rate margins. The wedge runs from filed loans without DDTL 2.1's mixed tranche to the total with it and Volta's reported loan. Packaging fees are 1% to 2% of 2025-2026 non-investment-grade commitments, spread over two years, before the bulge brackets take the large deals. · Sources: CCIR register (2026-10-03); CoreWeave, Nscale and Nebius filings; CoreWeave DDTL 2.1 upfront fee (1.50% per draw); research dossier sizing model for 2028. Author's calculation.** - Gross spread on CCIR's register: 1500 - 3700 - Wedge paid without investment-grade offtake: 226 - 430 - **Loan-packaging fees, a year: 65 - 130** - Every index, exchange and derivatives-structuring business, 2028: 2.8 - 245 At the 3.00% median of 2026's margins, the register would pay lenders about USD 2.0B a year. The central case for the whole derivatives stack in 2028 is about USD 38M. Spread is not profit, though. At the 265bp gap, defaults of about 4.4% a year use up the wedge at 60% loss given default, and only about 3.5% at a 75% loss, the collateral shortfall Core Scientific's debtors estimated for its equipment lenders in 2023 (my calculation). Residual cover that lifts recovery to 60% moves the breakeven to 6.6%. That's what protection buys; the loss is correlated. **Figure 6 — On my scoring, packaging and protecting non-investment-grade GPU credit beats every exchange-dependent business on risk-adjusted return · My scores from the research evidence: return potential (1 to 5), probability the business's kill observation occurs by 2028, and years until real liquidity. Risk-adjusted score is return x (1 minus P(break)) x 0.85 to the power of years. Judgments, not measurements; the sensitivity is in the text. · Author's scoring from the research dossier's evidence (filed spreads, product launches, regulatory status). Author's calculation.** | Business model | Return potential (1 to 5) | P(break by 2028) | Years to liquidity | Risk-adjusted score | |---|---|---|---|---| | Deal packaging and structuring for neoclouds | 3 | 0.40 | 0 | **1.80** | | Non-investment-grade GPU credit with residual protection | 4 | 0.50 | 1 | 1.70 | | Utilization, orchestration and inference routing | 3 | 0.45 | 0 | 1.65 | | Investment-grade GPU credit | 2 | 0.20 | 0 | 1.60 | | Residual-value underwriting and insurance | 4 | 0.55 | 1 | 1.53 | | Onchain GPU credit and tokenized compute | 3 | 0.55 | 0 | 1.35 | | Used-GPU remarketing and auctions | 3 | 0.45 | 2 | 1.19 | | Principal desk or compute prime broker | 3 | 0.55 | 1 | 1.15 | | Benchmarks and indices | 3 | 0.55 | 2 | 0.98 | | Exchanges and clearing | 4 | 0.70 | 3 | 0.74 | | Delivery verification and attestation | 2 | 0.50 | 2 | 0.72 | | Lender-mandated hedging programs | 3 | 0.75 | 3 | 0.46 | | Capacity options and option-writing | 2 | 0.70 | 3 | 0.37 | The 0.85 discount takes 15% off for each year a business waits for liquidity. Packaging and protected non-investment-grade lending sit within the noise of my scoring: packaging loses first place if its chance of breaking rises above 0.43, or at a 0.95 discount. The gap below them is what holds. Say I move every break probability 0.10 against the four cash-flow businesses (packaging, both credit books, routing) and 0.10 toward the three that need a listed market (exchanges, lender-mandated hedging, options). The weakest of the first group still scores 1.35, against 0.98 for the best of the second. ### The top of the list **Deal packaging and structuring** earns fees for turning a neocloud's contracts into a financeable special-purpose vehicle, delayed-draw loan or securitization slice, on deal flow measured in filings. It needs little capital unless it underwrites, and its 0.40 break probability is the chance the credit window shuts. October showed both: JPMorgan and Goldman Sachs underwrote only the USD 1.3B lease tranche of Volta's package and left the USD 3.7B chip loan to best efforts ([Bloomberg Law](https://news.bloomberglaw.com/banking-law/jpmorgan-goldman-curb-exposure-on-5-billion-volta-ai-loan-1)). The bulge brackets own the large deals, so the opening is sub-scale operators with real contracts and no capital-markets team. That's the pitch of Compute Labs, which Protocol Labs backed (see Where I sit). I can't size that slice from public data, and no base rate exists for GPU packaging desks. It breaks if no non-investment-grade loan prices inside SOFR + 800bp at 95 or better for two straight quarters. **Non-investment-grade GPU credit with residual protection** earns net interest margin on loans priced 200bp to 400bp over investment-grade deals. Valuation schedules or residual cover cut the loss given default, and amortization finishes inside the contract. Private credit already lends here, including SharonAI's USD 356M GPU facility at 9.95% fixed. I put 0.50 on two or more such facilities defaulting or restructuring before mid-2028. Correlated losses break it, so I'd want three years of seasoning through a price downturn, with realized losses well under the wedge. No such book exists yet. This is the compute sleeve of the fund I sketched in Paid to Wait. The GPU loan book confirms how I priced it: | | Paid to Wait (2026-10-18) | GPU loans, 2026 | |---|---|---| | Credit tenants or offtakers pay | Low 200s over SOFR (data-center construction loans) | SOFR + 2.25% to 3.00% (five loans) | | Everyone else pays | 200 to 300bp more; CoreWeave-tenant debt at SOFR + 325 to 425 | **265bp more**; SOFR + 4.50% to 5.50% (four loans) | | The Fund's floor for tenant credit | SOFR + 400 | All four non-investment-grade loans clear it | | What the premium pays for | Tenant credit | Offtaker credit and, on the newest paper, renewal | | What I'd require | Data centers and compute capped at 25% of the book (10% if its tenant-credit trigger trips) | Amortization inside the contract, plus residual cover or a valuation schedule | The difference is the collateral. A data center outlives its first lease. A GPU's three-to-four-year life in bank underwriting barely outlasts a three-year contract. Third comes **utilization, orchestration and routing**, asset-light and beside the credit thesis rather than inside it: Stripe [agreed to acquire](https://stripe.com/newsroom/news/stripe-agrees-to-acquire-openrouter) the router OpenRouter on 19 August 2026. The 0.45 is the chance that hyperscalers and NVIDIA take the default slot and routing margins compress. Investment-grade GPU credit works, and that's the problem for a new entrant. Margins run 225bp to 300bp over SOFR on vehicles rated A3 to Baa2, and the edge is cost of capital. Residual-value underwriting, where most of the seed writing points, ranks fifth, because the wedge pays for a correlated, wrong-way loss. Whoever owns the remarketing tape could become the Manheim of GPUs. That's why used-GPU auctions rank above the venues. ### Further down the list **Onchain GPU credit** ranks sixth (1.35): real yield and fast distribution, but small and untested in a down cycle. LlamaRisk [recommended against onboarding](https://governance.aave.com/t/arfc-onboard-usdai-susdai-to-aave-v3-arbitrum-instance/23260/4) USD.AI's stablecoin to Aave on 15 October 2025, citing GPU collateral liquidity among other concerns. Exchanges and clearing (0.74) would be large if they worked, but the incumbents hold the licenses, the fee is about USD 1.20 a contract on CME's 2025 energy book, and one-sided demand puts real liquidity three years out. ## If the loan book is only a phase | What has to be true | Signpost | Falsified if | Check by | |---|---|---|---| | Spreads follow credit support, not date, rates or tenor | Every new filed GPU loan, reviewed half-yearly | The half-year mean gap falls below 100bp, or a matched pair (same borrower, same half-year, matched tenor and recourse) prices within 150bp | 2027-12-31 | | Lenders underwrite the contract, not a market residual | Syndicated GPU loans outlasting their contracts; any market-value test | Three or more loans in a year at least two years beyond contract tenor, or any test tied to a third-party index | 2027-12-31 | | Listed compute derivatives stay small | Open interest on NYMEX, ICE, Architect and Kalshi | Listed compute futures above USD 250M of notional open interest | 2027-12-31 | | Frontier-lab offtakers perform | Commitment cuts in neocloud, Oracle and Microsoft filings | A lab cuts at least USD 1B of a disclosed commitment, or a lab-backed SPV misses a payment | 2027-12-31 | | GPU life outlasts loan tenor | H100 12-month term rate on Silicon Data's curve; hyperscaler useful lives | The rate averages below USD 1.00 for a quarter (USD 1.20 is the warning), or a top-four hyperscaler cuts GPU lives to four years or less | 2028-06-30 | | The credit channel stays open for non-investment-grade borrowers | Spread, discount and underwriting status of each new loan | No such loan prices inside SOFR + 800bp at 95 or better for two straight quarters | 2027-06-30 | **"Credit market" may be a stage, not a structure.** Mature leased assets also run on market residuals: an American Airlines pass-through certificate filed on 27 April 2026 tests its loan-to-value against independent appraisals ([424B5](https://www.sec.gov/Archives/edgar/data/4515/000119312526179398/d56276d424b5.htm)). The GPU market assembled the same parts in 2026: residual-value cover, benchmarks from an [FCA-regulated administrator](https://www.general-index.com/post/compute-desks-indexes-are-now-regulated) and ABS slices. If by 2028 a rated GPU deal without an investment-grade offtaker closes on an appraisal, this is an adolescent asset class. The sort also has a level problem. Across vintages the date dominates, and part of the investment-grade bucket may be the vendor itself: Lambda's term-loan offtaker is [reported](https://thenextweb.com/news/lambda-917-million-gpu-leveraged-loan-nvidia-morgan-stanley) to be NVIDIA. The bucket also leans on frontier labs whose plans reportedly show large negative free cash flow (OpenAI's July 2026 presentation, relayed secondhand from the FT, projects about USD 278B of it over 2026 to 2030). One lab stumbling would turn "priced on the offtaker" into one correlated bet. **The live test doesn't discriminate.** October's stress was priced through CDS, discounts and best-efforts syndication, instruments the incumbents already run, and Goldman Sachs located the widening in neocloud and lab-exposed credit ([Yahoo Finance](https://finance.yahoo.com/markets/stocks/article/why-coreweave-nebius-and-other-neocloud-stocks-are-being-pounded-goldman-sachs-explains-125730269.html)). That counts against the risk-transfer businesses I rank second and fifth. NVIDIA's backstop for CoreWeave carries wrong-way risk, since it can end in the insolvency it exists for. And recoveries are thin: when Core Scientific filed in 2022, its equipment collateral covered about 25% of the miner-equipment lenders' claims ([disclosure statement](https://www.sec.gov/Archives/edgar/data/1839341/000119312523170749/d398991dex992.htm)). **The derivatives might surprise me.** Deemed approval is the default on 9 November 2026. The CFTC's Chairman [has said](https://www.cftc.gov/PressRoom/PressReleases/9286-26) America can't win the AI race without a compute derivatives market. A lender writing the first compute hedge into a loan would move the bottom of my ranking up. Most of that is right, and none of it touches the descriptive claim: across 2026's filings, the spread between two GPU loans is the price of the signatures behind them. It narrows the prescriptive one. The return sits in packaging and protected lending for borrowers the strategics won't back, at prices that cover a correlated loss. ## Eight signposts, fastest first 1. **The CFTC comment file, closed 20 October 2026.** A substantive letter in the final file from CME, ICE, Silicon Data, Ornn, NVIDIA, a hyperscaler or a large neocloud shortens the listed path; silence keeps it long. 2. **NYMEX GPU1 and GPU2, by 9 November 2026.** Deemed approval starts the open-interest clock; a non-approval notice strengthens the credit-first view, and an agreed extension keeps the listed path long. 3. **Listed open interest, monthly after any listing.** More than USD 250M of notional by 31 December 2027, about 120,000 NYMEX-size contracts, falsifies my view; under 5,000 six months after listing supports it. 4. **The spread gap, at each new filed GPU loan.** The thesis dies on the convergence test in the table's first row, against a base case above 200bp. A half-year with no loan in either bucket rolls into the next. 5. **CoreWeave's refinancing, at each 10-Q.** USD 6.184B of principal falls due in 2027. Refinancing inside current DDTL margins says the channel holds; a distressed exchange says it is closing. 6. **The first compute hedging covenant, through 31 December 2027.** A GPU credit agreement that requires a compute price hedge moves lender-mandated hedging toward the middle of the ranking. So does one that [assigns a different financing value](https://davefriedman.substack.com/p/what-will-actually-make-people-hedge) to hedged revenue, as Dave Friedman suggests. 7. **The first rated non-investment-grade GPU deal on residual cover or a valuation schedule, through 31 December 2027.** One supports residual-value underwriting as a business; none says lenders still prefer amortization. 8. **The H100 12-month term rate, monthly.** A quarter below USD 1.20 per GPU-hour on Silicon Data's curve is the warning, and below USD 1.00 the life assumption fails. Its current print isn't in my data, so CCIR's modeled two-year rate stands in: USD 1.40 on 18 October 2026, close enough that I treat it as a yellow light. On 15 October 2026 a single company's credit traded at distress levels while the GPU loans around it priced on the signatures underneath. In Paid to Wait the contract was the filter. Here it is also the collateral. The tail is where the next price gets set. Whoever can stand behind it with rated capacity and a tape of real resale prices will set that price, and I'd bet they set it before any exchange does. ## What I counted and how My calculations from filed or published terms: the spread gaps (265bp floating, 292bp with Volta, about 150bp fixed), the variance shares, the same-issuer pairs, the USD 1.68 per dollar of capex, the index correlation, the pools in Figure 5 and the breakeven default rates. The scorecard is my judgment (formula and sensitivity under Figure 6). CCIR supplies the register total, the fixed-rate spreads, the used-GPU record and the modeled contract rate. Paid to Wait left CoreWeave's DDTL 5.0 price unchecked; the [8-K](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000236/crwv-20260515.htm) confirms SOFR + 4.50%. Every other number is someone else's, named and dated where it appears. "Filed" means an SEC filing, "company-released" a company's own release, "reported" press, and "modeled" an estimate. The thinnest sourcing sits under Volta's price, the CDS levels (vendor data via Reuters and CNBC), Carmen Li's dual role, CCIR's modeled contract rate and Nscale Iceland's margin, which I have only from an S-1 excerpt. The CFTC status, the docket and the market prints are as captured on 19 October 2026, and all three move fast. Data freeze: 2026-10-19. The analysis files behind the calculations are available in companion data. ## References **Primary data** - CoreWeave, *DDTL 5.5 Credit Agreement, 8-K and release* (2026-08-07 to 2026-08-10). [8-K](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000357/crwv-20260807.htm), [agreement](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000357/ex101creditagreement.htm), [release](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000357/ex991pr.htm): the loan terms. - CoreWeave, *8-Ks for DDTL 4.0 (2026-03-31), DDTL 5.0 (2026-05-18), DDTL 2.0 Fifth Amendment (2025-10-02) and the NVIDIA order form (2025-09-15)*. [DDTL 4.0](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000129/crwv-20260330.htm), [DDTL 5.0](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000236/crwv-20260515.htm), [DDTL 2.0](https://www.sec.gov/Archives/edgar/data/1769628/000119312525227562/d910811d8k.htm), [DDTL 2.1 exhibit](https://www.sec.gov/Archives/edgar/data/1769628/000119312525227562/d910811dex101.htm), [NVIDIA order](https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm): margins, advance rates, recourse, the NVIDIA backstop. - CoreWeave, *Form 10-Q, quarter ended 2026-06-30* and *Form 10-K, 2025*; *DDTL 4.0 overview* (March 2026). [10-Q](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm), [10-K](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm), [deck](https://s205.q4cdn.com/133937190/files/doc_presentations/2026/Mar/2026-03-DDTL-4-Overview.pdf): debt, revenue mix, useful life and contract sizing. - IREN, *8-K on the Hardware 3 financing* (2026-06-01); Nebius, *6-K* (2026-07-17); Nscale, *S-1* (2026-09-18). [IREN](https://www.sec.gov/Archives/edgar/data/1878848/000114036126023427/ef20075181_8k.htm), [Nebius 6-K](https://www.sec.gov/Archives/edgar/data/1513845/000110465926084452/tm2620683d1_6k.htm), [Nscale](https://www.sec.gov/Archives/edgar/data/2110365/000119312526395475/ck0002110365-20260918.htm): margins and tenor. - NVIDIA, *Form 10-Q, quarter ended 2026-07-26*, *form of residual value guaranty* and *8-K on the PORTS guaranties* (2026-08-17). [10-Q](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm), [guaranty](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda2027q2ex101.htm), [8-K](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000069/nvda-20260817.htm): guarantees and cloud purchase commitments. - TeraWulf, *10-K for 2025*; Cipher Mining, *10-Q, quarter ended 2025-09-30*; WhiteFiber, *10-Q, quarter ended 2026-03-31*. [TeraWulf](https://www.sec.gov/Archives/edgar/data/1083301/000108330126000031/wulf-20251231.htm), [Cipher](https://www.sec.gov/Archives/edgar/data/1819989/000181998925000112/cifr-20250930.htm), [WhiteFiber](https://www.sec.gov/Archives/edgar/data/2042022/000121390026056212/ea0288695-10q_whitefiber.htm): Google's backstops; the H200 sale. - Meta, *10-K for 2025*; Amazon, *10-K for 2025*. [Meta](https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm), [Amazon](https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm): useful-life changes. - Talos Energy, *amended credit agreement* (2026-06-30). [Talos](https://www.sec.gov/Archives/edgar/data/1724965/000119312526290589/d14154dex101.htm): a lender-mandated hedge. - American Airlines, *424B5 pass-through certificates* (2026-04-27); Core Scientific, *disclosure statement* (2023-06-21). [American](https://www.sec.gov/Archives/edgar/data/4515/000119312526179398/d56276d424b5.htm), [Core Scientific](https://www.sec.gov/Archives/edgar/data/1839341/000119312523170749/d398991dex992.htm): appraisal-based loan-to-value; collateral coverage in a bankruptcy. - CFTC, *Request for Comment on the Listing of Compute Derivatives Contracts* (2026-08-21), the NYMEX GPU1 product page, release 9286-26, docket CFTC-2026-1850, 17 CFR 40.3 and Commitments of Traders data. [RFC](https://www.federalregister.gov/documents/2026/08/21/2026-17163/request-for-comment-on-the-listing-of-compute-derivatives-contracts), [GPU1](https://www.cftc.gov/IndustryOversight/IndustryFilings/TradingOrganizationProducts/62544), [release](https://www.cftc.gov/PressRoom/PressReleases/9286-26), [docket](https://www.regulations.gov/docket/CFTC-2026-1850/comments), [40.3](https://www.ecfr.gov/current/title-17/chapter-I/part-40/section-40.3), [COT](https://publicreporting.cftc.gov/d/6dca-aqww): status, views, comments, power open interest. - KalshiEX, *INDEXVALUE product certification* (2025-11-26); Lighter, *RWA market specifications*. [Kalshi](https://assets.kalshi.com/regulatory/product-certifications/INDEXVALUE.pdf), [Lighter](https://docs.lighter.xyz/trading/real-world-assets-rwas/market-specifications): contract design. **Analysis and market** - Fitch Ratings on CoreWeave Compute Acquisition Co. VIII (2026-04-01) and DDTL 5.5 (2026-07-16); S&P Global Ratings, *IS AN03 LLC preliminary 'BB'* (2026-09-11); Morningstar DBRS on DDTL 4.0 (2026-04-03). [Fitch A-sf](https://www.fitchratings.com/research/structured-finance/fitch-rates-coreweave-compute-acquisition-co-viii-loans-a-sf-outlook-stable-01-04-2026), [Fitch BB+](https://www.fitchratings.com/research/corporate-finance/fitch-assigns-coreweave-new-ddtl-bb-rating-affirms-idr-at-bb-outlook-positive-16-07-2026), [S&P](https://www.spglobal.com/ratings/en/regulatory/article/-/view/type/HTML/id/3626094): rating sensitivities. - Fitch Ratings, *Consumer ABS Rating Criteria: Residual Value Addendum* (2024-10-11); Manheim, *Used Vehicle Value Index*. [Fitch](https://assets.fitchratings.com/downloadFile?reportType=report&sfReport=false&slug=structured-finance%2Fconsumer-abs-rating-criteria-residual-value-addendum-11-10-2024), [Manheim](https://site.manheim.com/en/services/consulting/used-vehicle-value-index.html): how auto residuals became bankable. - CCIR, *GPU-backed credit register* (2026-10-03), *hardware and executed sales record* (2026-10-12) and *Rent and Age* (2026-09-29). [Credit](https://ccir.io/credit), [hardware](https://ccir.io/hardware), [age curve](https://ccir.io/research/gpu-age-curve): register, fixed-rate spreads, used-GPU prints, contract rate, rent decay. A single publisher formed in 2026. - Bandi and Su, *(Early) AI Compute Asset Pricing*, arXiv 2607.12156 (preprint, 2026-09-10). [arXiv](https://arxiv.org/abs/2607.12156): rent per unit of compute by chip generation. - SemiAnalysis, *The Great GPU Shortage* (2026-04-02). [SemiAnalysis](https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity): one-year contract index; an interested party. - Moulton, *California electricity futures: the NYMEX experience* (2005); NGI on the 2002 delisting; WIRED on Enron's bandwidth trading (2001-11-30); Dallas Fed Research Paper 8001 (1980) on Ederington. [Moulton](https://www.sciencedirect.com/science/article/abs/pii/S0140988304000970), [NGI](https://naturalgasintel.com/news/nymex-ready-to-begin-power-contracts-again-starting-with-pjm/), [WIRED](https://www.wired.com/2001/11/enron-a-bandwidth-bloodbath/), [Dallas Fed](https://www.dallasfed.org/~/media/documents/research/papers/1980/wp8001.pdf): the analogue and the hedge arithmetic. - *The Global Market for Exchange-Traded Derivatives*, Applied Finance Letters (2022); Bekkerman and Tejeda, Agricultural Economics (2017). [AFL](https://doi.org/10.24135/afl.v11i1.520), [Bekkerman and Tejeda](https://scholarworks.montana.edu/bitstreams/9831f9ba-54e7-4283-8e12-a3762cc0c225/download): base rates and what predicts survival. - Seed essays: Multicoin (2026-10-08); Wafer's Eugene Ye (2026-09-27); Brett Harrison on X (2026-07-01); Dave Friedman (2026-09-07 and 2026-10-16). [Multicoin](https://multicoin.capital/2026/10/08/compute-capital-markets/), [Wafer](https://wafer.substack.com/p/lets-talk-about-trading-compute), [Harrison](https://x.com/BrettHarrison/status/2072327852498797048), [Friedman, bankable](https://davefriedman.substack.com/p/what-makes-gpu-revenue-bankable), [Friedman, hedging](https://davefriedman.substack.com/p/what-will-actually-make-people-hedge). - Company releases and pages, dated where cited: [Lambda fixed-rate](https://lambda.ai/blog/lambda-closes-1-billion-senior-secured-fixed-rate-financing), [Lambda TLB](https://lambda.ai/blog/lambda-closes-926-million-senior-secured-term-loan-b-facility), [NVIDIA release](https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Partners-With-Apollo-BlackRock-Blackstone-Brookfield-Goldman-Sachs-and-KKR-to-Establish-AI-Compute-Infrastructure-Financing-Platforms-to-Mobilize-Over-500-Billion-of-Third-Party-Capital/default.aspx), [NVIDIA blog](https://blogs.nvidia.com/blog/nvidia-ai-factory-compute/), [American Compute](https://www.amcompute.com/blog/gpu-residual-value-insurance-incline-accelerant), [The Insurer](https://www.theinsurer.com/ti/analysis/gpu-residual-value-insurance-paper-used-to-tackle-ai-financing-risk-2026-08-26/), [CME](https://www.cmegroup.com/media-room/press-releases/2026/5/12/cme_group_and_silicondatapartnertolaunchfirstcomputefutures.html), [Ornn methodology](https://data.ornn.com/methodology), [General Index](https://www.general-index.com/post/compute-desks-indexes-are-now-regulated), [Compute Desk](https://www.linkedin.com/posts/compute-desk_congratulations-to-compute-desk-partner-pluto-activity-7480732064108183552-3Wa5), [Stripe](https://stripe.com/newsroom/news/stripe-agrees-to-acquire-openrouter), [PL Capital](https://www.linkedin.com/posts/protocolvc_activity-7480695429689430016-zzOT), [The Block](https://www.theblock.co/news/deals/2024-07-03-solana-based-gpu-tokenization-protocol-compute-labs-funding-token-valuation-303398), [Aave forum](https://governance.aave.com/t/arfc-onboard-usdai-susdai-to-aave-v3-arbitrum-instance/23260/4). - UPI on Itel's bankruptcy (1981-01-19); Travelers Property Casualty, *first-quarter 2003 earnings presentation*, Gulf residual-value slide (2003-04). [UPI](https://www.upi.com/Archives/1981/01/19/Itel-Corp-a-computer-leasing-firm-troubled-by-financial/7749348728400/); the slide sits at media.corporate-ir.net, served over http only, so it carries no link here: the base rate for residual-value insurance. - Fixed + Floating, *Who Pays for the GPUs?* with Ankur Patel of Ares (2026-09-08). [Fixed + Floating](https://www.fixedandfloating.com/p/who-pays-for-the-gpus): the lender's view of residuals. **Prices and current conditions** - Reuters via U.S. News, *Why are credit default swaps unnerving AI investors?* (2026-10-15); CNBC, *CoreWeave credit default swaps hit distress levels* (2026-10-15). [Reuters](https://money.usnews.com/investing/news/articles/2026-10-15/explainer-why-are-credit-default-swaps-unnerving-ai-investors), [CNBC](https://www.cnbc.com/2026/10/15/coreweave-credit-default-swaps-hit-distress-levels-as-investors-sour-on-ai-debt.html): CDS and bond levels. - Bloomberg via The Business Times on Volta (2026-10-16), Bloomberg Law on its underwriting (2026-10-13) and The Next Web on Lambda's offtaker (2026-08-10). [Business Times](https://www.businesstimes.com.sg/companies-markets/banking-finance/jpmorgan-goldman-boost-yield-volta-ai-loan-lure-investors), [Bloomberg Law](https://news.bloomberglaw.com/banking-law/jpmorgan-goldman-curb-exposure-on-5-billion-volta-ai-loan-1), [The Next Web](https://thenextweb.com/news/lambda-917-million-gpu-leveraged-loan-nvidia-morgan-stanley): anonymous sources. - Reuters on lenders and GPU life (2026-10-01) and on CME's uranium future (2026-05-07). [Reuters, GPU life](https://www.reuters.com/legal/transactional/nvidias-bet-that-its-chips-can-finance-ai-boom-gets-wall-street-reality-check-2026-10-01/), [Reuters Japan](https://www.reuters.com/jp/markets/japan/KGPXPEEMAJOUDAPZ6SLQ3X3IMM-2026-10-02/), [Reuters, uranium](https://www.reuters.com/business/energy/cme-seeks-revive-uranium-trading-with-physical-futures-launch-sources-say-2026-05-07/): bank underwriting lives; the uranium analogue. - Equipment Finance News on GPU loans in ABS (2026-10-06); Yahoo Finance on Goldman Sachs (2026-10-16), Crypto Briefing on Jump and DRW (2026-10-16). [EFN](https://equipmentfinancenews.com/news/lender-operations/gpu-financing-filters-into-abs-with-stonebriar-wingspire-deals/), [Yahoo Finance](https://finance.yahoo.com/markets/stocks/article/why-coreweave-nebius-and-other-neocloud-stocks-are-being-pounded-goldman-sachs-explains-125730269.html), [Crypto Briefing](https://cryptobriefing.com/jump-trading-drw-compute-power-contracts/). - Index and list prices, 2026-10-18 and 2026-10-19 (Ornn, Silicon Data, Computable via Financial Datasets, Vast.ai, Lambda, CoreWeave), and Silicon Data's term curve with CME's notice (2026-10-12). [Ornn](https://data.ornn.com/markets/h100-sxm), [Silicon Data](https://www.silicondata.com/products/silicon-index/h100), [Silicon Data term curve](https://www.silicondata.com/products/forward-curve), [CME notice](https://www.cmegroup.com/notices/electronic-trading/2026/10/20261012.html). **Prior work in this series** - [*Paid to Wait: The Case for a Private Infrastructure Credit Fund*](https://www.second-breakfast.co/blog/paid-to-wait), Second Breakfast, 2026-10-18. - [*The Conversion Layer*](https://www.second-breakfast.co/blog/the-conversion-layer), Second Breakfast, August 2026. --- *Views are my own and do not represent PL Capital or Protocol Labs. Nothing here is investment advice. Do your own diligence.* --- Source: https://www.second-breakfast.co/blog/the-contract-is-the-collateral