The Customer Is the Collateral

01The ledger

Compute credit is three markets that share one driver and do not share a rating method. The Compute Credit Tracker at CCIR, as of September 21, counts $238 billion of tracked debt across 239 instruments and 57 issuers, and splits it into chip-collateralized loans ($48.0 billion, 36 instruments), operator campus loans ($34.6 billion, 42 instruments) and landlord vehicles and securitizations ($155.2 billion, 25 instruments). It tracks the multi-tenant colocation securitizations, the oldest market, separately.

Building-backed bonds. The Structured Finance Association puts outstanding data center securitizations at $61 billion so far in 2026, against $4 billion in 2020, and cites a Barclays estimate of $180 billion by the end of 2028. The market is now 12 percent of esoteric asset-backed issuance, from 3 percent in 2020, with roughly $50 billion of issuance expected this year. The structures are conventional: master trusts that allow collateral substitution, a five-year anticipated repayment date and a thirty-year legal maturity, credit that rests on tenant quality and lease term. Charles River Associates counted 69 issuances worth $34.4 billion through the third quarter of 2025; senior tranches are mostly rated A or A-minus, and spreads narrowed from 2.56 percent in 2020 to 1.45 percent in 2025. Chips are described throughout as a separate, emerging sector.

Landlord vehicles. These are the deals that moved the market. A sponsor and a private credit manager form a company that owns one campus; the sponsor leases it for fifteen to twenty years and holds a minority stake; the company issues long bonds against the lease. Meta’s Hyperion vehicle with Blue Owl, in October 2025, raised about $27 billion at 6.58 percent, rated A-plus, amortizing into 2049. Meta’s El Paso vehicle with BlackRock’s infrastructure and credit arms raised $12.3 billion in July 2026, rated A-plus by S&P and AA-minus expected by Fitch, at 7.534 percent. Meridian Arc, a Fluidstack joint venture with a fifteen-year lease guaranteed by Google, sold $5.7 billion of five-year notes in April at 6.25 percent, rated double-B, on a book of roughly $20 billion. Galaxy’s second Helios phase in Texas, whose tenant is CoreWeave, priced $3.507 billion in July at 9.875 percent, due 2031. The high-yield market took data center paper at a pace PitchBook described as a binge; the Bank of England reports that AI issuers accounted for 41 percent of non-refinancing high-yield issuance in the United States this year.

Chip loans. This is the market that carries the word “compute” in its name and is the smallest of the three. The rated record is short enough to print.

Three things stand out. Every facility sits in its own special-purpose company and maps to one borrower; the tracker finds no structure that pools several operators. Every investment-grade rating names, or is reported to name, a single highly rated customer. And the first facility, in 2023, is the only one in the table priced off the hardware.

The shadow. Above the rated stack sits the debt of the sponsors themselves, and above that, their commitments. S&P Global counted $225 billion of hyperscaler bond issuance in the first half of 2026, up almost tenfold on the year, with $400 billion projected for the full year. The Bank of England notes that the five largest AI spenders were 3 percent of outstanding US investment-grade debt at the end of 2025 and more than 15 percent of this year’s issuance. KBRA puts the contractual commitments and contingent support of the seven largest AI infrastructure companies at about $3.2 trillion, from $575 billion at the end of 2024. A Nikkei study cited by Fortune puts off-balance-sheet obligations at the five largest at $1.65 trillion; Moody’s estimate is $1.2 trillion. Nvidia’s own quarterly filing for the period ended July 2026 discloses $366 billion of supply, cloud, lease, equity and capital commitments, plus $108.5 billion of land and power guarantees and $56 billion of cloud and lease commitments: about $530 billion gross, by SemiAnalysis’s reconciliation, against $33 billion of debt on the balance sheet. And on July 9, S&P cut Oracle to BBB-minus, one notch above high yield, citing a $42 billion cash deficit, leverage above four times, and OpenAI’s roughly half share of a $638 billion backlog. The agency’s adjusted debt figure includes $260 billion of lease commitments falling due between fiscal 2027 and 2029.

That downgrade matters to the chip market more than any chip deal, because the chip market has spent 2026 borrowing the credit of exactly this kind of counterparty.

02The price of a promise

Lenders price the customer, not the chip, and the spread ladder says so with unusual precision. The same borrower with the same hardware pays 2.25 points over the benchmark rate when Meta stands behind the contract, 4.50 points when two sub-investment-grade customers do, and 5.50 points when the loan runs two years past its contracts. In 2023, with hardware and assigned contracts as the only support, the rate was 11 to 14 percent all-in. CCIR’s summary of the rating agencies’ published reasoning is that the counterparty opens the grade and the structure sets the notch.

The best-rated deals do not manage resale risk; they design it out. Revenue is a fixed payment for availability, not usage. The loan amortizes inside the contract. Rating assumptions reported by CCIR include 70 percent utilization, debt service covered 1.15 to 1.40 times, and reserves of three months against a six-month norm. If the customer pays, the chip’s resale value never enters the waterfall. If the customer does not pay, the chip is what is left, and no rated deal has yet been tested on that question.

The landlord market prices the same promise at longer tenor. Meridian Arc is the cleanest case: a double-B issuer, a joint venture of a neocloud and a developer, borrowed $5.7 billion at 6.25 percent, 233 basis points over Treasuries, because Google guarantees fifteen years of rent. The guarantee is the credit. The Fluidstack lease, the campus and the chips it will hold are what the guarantee is attached to. Google has now guaranteed about $4.5 billion of Fluidstack lease obligations at TeraWulf and Cipher sites, plus the fifteen-year Meridian Arc lease, taking warrants in the landlords in exchange.

Meta’s two vehicles are the market’s cleanest repricing gauge. Hyperion, October 2025: A-plus, 6.58 percent. Sopaipilla, July 2026: A-plus, 7.534 percent. Same sponsor, same 80/20 structure, same maturity profile, same tenant. The 95 basis points between them is what nine months of supply did to the price of Meta’s promise, and the second deal still cleared. Whether that gap widens on the next campus is the single most legible number in the market.

A finance professor’s reading of the same ledger draws the line that matters. Sascha Steffen separates hyperscaler take-or-pay contracts, which are credit substitution because an investment-grade entity pays whether or not it uses the capacity, from AI-lab take-or-pay contracts, which are credit deferral because the lab must keep raising capital to honor them. He puts OpenAI’s cloud commitments near $590 billion against annual recurring revenue near $25 billion and a 2026 cash burn near $27 billion. A rating that leans on the first kind of contract describes something the contract can support. A rating that leans on the second, or a loan whose tenor outruns the contract, describes something the collateral cannot.

The people holding the paper are, increasingly, insurers. PIMCO is reported to hold about $18 billion of Hyperion. Apollo’s $3.5 billion of debt to the xAI chip vehicle formed in January, Valor Compute Infrastructure, with Nvidia as a $1.9 billion anchor equity investor and 100,000 GB200 chips leased to xAI, is reported to sit with Apollo’s insurance affiliate Athene. Three of the six managers on Nvidia’s financing platforms own or control insurers likely to hold the resulting notes. Steffen’s point is that a downgrade channel exists: rating cuts trigger capital charges under US risk-based capital and Solvency II, forcing sales of assets with no secondary market. The January 2027 Solvency II revision cuts spread-risk charges on securitizations roughly in half, from about 46 to about 22 percent, which will pull European insurance demand toward exactly this paper.

03The model in the drawer

The instrument the whole market is walking toward is a pool of chip-backed loans from many operators, sliced by seniority. Nobody has issued one. Two Federal Reserve economists have modeled it. Seung Jung Lee and Sriram Nagaraj’s working paper, Pricing, Hedging, and Securitizing AI Compute, dated August 23, 2026, treats a chip-hour as a non-storable commodity, like electricity or an airline seat, and builds four layers of market on that fact: a benchmark price, futures and clearing, privately negotiated swaps, caps and floors, and securitization. It states the authors’ views, not the Federal Reserve’s.

The paper’s finding at the securitization layer is the one this market needs to read. It uses the standard large-pool credit model: many borrowers, each defaulting when a shared economic factor plus its own luck turns bad enough. Its twist is that the shared factor is the long-run compute price itself. A weak compute market therefore does two things at once: it raises defaults, and it cuts what seized chips fetch. Recovery equals the depreciated chip value, scaled by the state of the compute market, divided by the loan-to-value at the start, with a salvage floor.

Value the slices of such a pool with a single average recovery rate, the way a newcomer from hardware or cloud might, and the error is not evenly spread. It concentrates in the piece sold as safest.

How far an average-recovery valuation understates expected loss, by slice. +200% means the true expected loss is three times the valuation. Lee and Nagaraj (2026), simulation 8.

The second finding is about collateral. Because credit losses and chip depreciation multiply, a chip loan needs far more collateral at longer maturities than a loan against an asset that holds its value: 1.3 times face value at zero maturity, and at three years 2.4, 4.3 or 7.9 times for yearly decay rates of 0.2, 0.4 or 0.6. The paper’s own estimate from resale prices is a decay rate of 0.46 to 0.53 a year, a loss of roughly 40 percent of value each year, with two-year-old H100s keeping 30 to 40 percent of peak new price.

Three further results bear on structure. Contracts with a straight payoff, futures and swaps, have one fair price and spread risk; contracts with a bend in the payoff, caps, floors and sold protection, have a price range and concentrate it, and the range narrows with more strike prices only for capped payoffs, never for uncapped ones. The tail of a protection seller’s loss grows without limit as price spikes become more violent, and the spike parameter is the one input that cannot be measured from public data. And a token price cannot be hedged with chip-hour futures: the best hedge explains 3 percent of its variation, because software efficiency moves token prices in ways the rental rate never sees.

The paper should be read with its own caveats. It is tested against itself: ten simulations confirm the algebra, none tests whether real rents behave this way. The threefold figure uses a yardstick that mature credit desks abandoned after 2008, and the authors say their worry is that newcomers rebuild the shortcut. The pool it models is not today’s deal, which rests on one customer, not many similar loans; the authors note that dependence on one buyer is itself the dangerous link between default and recovery. Of the roughly thirty inputs the framework needs, public data supports four: the long-run price level, its volatility, its drift, and the chip depreciation rate. Everything else, including the entire credit layer, is assumed.

What survives the caveats is direction. When recovery falls as defaults rise, the senior slice is the most understated, and today’s market has the same link by a different route. A pool of single-customer loans to Meta, Microsoft and Oracle is diversified by name and concentrated by cause. So is a pool of small operators serving small buyers. The paper’s warning applies to the second generation of deals. The first generation has moved the same risk into a handful of customers whose fortunes share one driver.

04Long paper, short silicon

Every structure in section 01 borrows long against an asset that earns short. The mismatch is the market’s central design problem, and the market’s answers to it are the intermediate forms in section 05.

The Bank of England put it plainly in July: frontier data centers can quickly become outdated if they cannot support the latest chips, and issuing debt with maturities aligned to chip life would imply shorter tenors than typical credit investors prefer. Hyperion and Sopaipilla amortize into 2048 and 2049 against leases of fifteen to twenty years and chips that KBRA’s research treats as a re-leasing question, not a durability question: whether customers keep using the capacity at an attractive price, not whether the chip still works. Six-year accounting lives sit against a competitive life that Quinn Emanuel’s litigation alert puts at two to four years and against H100 rental rates down 70 to 90 percent since 2023.

The resale record is thin and getting thicker. American Compute’s June report, built on 76,775 completed secondary-market transactions, refuses to publish a single depreciation curve and publishes bands instead: an H100 at 45 to 74 percent of list price in 2026, 34 to 59 percent in 2027, 24 to 51 percent in 2028, with anything above the band labeled dangerous. The Federal Reserve paper’s decay estimate sits inside those bands. Both say the same thing: a lender who books a residual at the top of the range is booking a hope.

The rental rate is what the residual is ultimately worth, and it is now published daily. Silicon Data’s H100 index, the reference for the futures set to list on October 5, prints $2.53 per chip-hour. That number is the key to reading Nvidia’s revenue floors. Under the AI Compute Partnership announced in July, Nvidia agreed to rent unused capacity from a small operator at a base rate reported near $2.35 an hour for a GB300, against $4.50 to $4.60 for a five-year contract, for six years, taking half the revenue above the base. The reported base sits near the published H100 spot print and well under the contract rate for the newer part, though the two are different chip generations, so it reads as an order of magnitude rather than a like-for-like comparison. Either way the floor is set to repay the lender, not to make the operator money. That is what a lender wants from a floor, and it is why the first two facilities under the program, Firmus at a reported $25 to $30 billion and SharonAI at about $4.9 billion, could be advanced at 70 to 80 percent against it. On August 27 the Wall Street Journal reported that Nvidia had paused the program over antitrust concerns raised internally, after requiring operators to vet customers through Nvidia and spread capacity across smaller buyers; Nvidia said the model “is still in place and continues to evolve.” The commitments remain in the filing.

What no deal in the record does is carry the depreciation curve in its documents. No public chip loan tests loan-to-value against a published index. None requires the borrower to hedge. CoreWeave’s August facility runs about five years against contracts averaging three, with renewal or re-leasing covering the gap, at 5.50 points over the benchmark. That is renewal risk and resale risk coming back into the market at a price, which is better than coming back without one, and it is also a rated chip loan whose repayment depends in part on a rental rate nobody has yet contracted.

05Where it goes

The market moves from a few large loans, each resting on one customer, toward pools of many smaller operators serving many smaller buyers. The shift from training to inference drives it; McKinsey expects inference to grow 35 percent a year to 2030 against 22 percent for training and to pass half of all AI compute by then. Inference is thousands of buyers on shorter, usage-based contracts, served by more and smaller operators, on older chips, closer to users. It suits pooling, and it turns revenue into a usage and price risk that no customer contract fixes.

The evidence for the middle stage arrived in the last nine months, from every direction at once.

From the vendor: the revenue floors above, and on August 10 Nvidia’s non-binding platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize over $500 billion of third-party capital, with Nvidia’s chips described as “fungible and transferable across customers and operators.” The release names no backstop; SemiAnalysis reports a residual-value guarantee of up to 25 percent as the program’s design. If that is the term, it is the first vendor residual guarantee in the market, and its trigger, the price at which Nvidia’s guarantee pays, becomes the most important undisclosed number in compute credit.

From the hyperscalers: Google’s lease guarantees, which turn a bitcoin miner’s building into a double-B bond and a neocloud’s lease into fifteen years of investment-grade rent. Meta’s residual-value support for Hyperion, reported at $28 billion and carried in footnotes rather than as a liability, is the same instrument from the tenant’s side.

From the insurers: a Munich Re-backed warranty on the gap between predicted and realized resale price on defaulted chip loans, launched in March by the on-chain lender USD.AI and the insurer Barker. USD.AI is the one lender in the record writing the shape of loan the Federal Reserve paper prescribes: 70 to 80 percent loan-to-value, three-year amortization with principal and interest every thirty days, a three-month reserve, depreciation reinsurance, monitoring on the hardware and a resale channel named in advance. It is small, it is on-chain, and it is the one place where several chip loans already sit behind a single liability.

From the banks: Nebius’s $775 million facility in July was underwritten and syndicated by a group of ten banks with MUFG as structuring agent, at 2.50 over, and the company expects to raise more capital at similar terms against $40 billion of contracted revenue from Microsoft and Meta. A repeatable bank template is what a warehouse line needs before a pool can be built.

From the borrowers: CoreWeave, which invented the chip loan, priced $3.7 billion of convertible notes on September 18 at 2.875 percent, due 2033, its second convertible of the year. The largest chip borrower is moving up its own capital structure, toward paper that references its equity rather than its hardware, while its facilities carry about $25 billion of debt and $536 million of quarterly interest expense at the first-quarter count.

From the regulators, a change of tone rather than rule. The Federal Reserve’s May Financial Stability Report records that half of its surveyed contacts named AI as a risk, up from about 30 percent in the previous survey, and that software has become the largest sector in private credit portfolios as redemptions from semi-liquid vehicles accelerated. The Bank of England cites an OECD estimate that private credit’s share of AI investment financing rose from 9 percent in 2024 to 34 percent in 2025, and JPMorgan’s estimate that AI chips alone may need more than $2 trillion of funding over five years. The IMF’s April report names circular financing along the AI value chain as a channel through which a slowdown would propagate. The Chicago Fed counts $450 billion of large-bank commitments to AI software, energy and semiconductor firms, $150 billion drawn, and flags the unmeasured indirect exposure through the private lenders banks finance. The Dallas Fed estimates that $300 billion of AI-related investment-grade bonds in 2026 adds about $360 billion of ten-year-equivalent duration to the market, roughly an eighth of annual Treasury supply, and that the swaps operators use to fix their private floating-rate loans are already visible in the relationship between the Treasury curve and swap spreads. Staff at the Securities and Exchange Commission said on August 7 that building-backed data center deals fall outside the asset-backed securities rules; counsel warn the reasoning may not extend to chip deals. Bondholders sued Oracle in January over the disclosure of its borrowing, and shareholders sued CoreWeave over its description of demand.

From the reference price: CME’s H100 and B200 rental index futures are scheduled to take effect October 5, a date CME’s own product page still marks as pending regulatory review. Nodal Exchange and Compute Desk have announced Hopper and Blackwell futures for later in the year. Kalshi’s event curves, the only forward prices live through the summer, ran on about $4.4 million of volume and were missing by about 10 percent two days before settlement. A listed curve is what turns a rental-rate floor into a hedgeable exposure and a loan-to-value test into a number a trustee can read. It is also, the paper notes, what makes pooling and slicing possible, and if deals are marked to a thin curve the mark is least reliable in exactly the stressed conditions that test the deal.

What the pool will need that a single loan does not. Credit files on many unrated names. An independent record of hours actually rented, rather than the borrower’s count. Standard, assignable contracts and a named backup operator who can run the chips if one fails. A rent hedge on the index the deal is marked to. Eligibility rules, concentration limits by operator, buyer, chip family, site and power region, and cash-trap triggers that fire when usage, rents and resale values fall together. And loan-level performance and recovery data, the paper’s top request, of which the public record today contains none: no default, no workout, no sale price.

What diversification fixes, and what it does not. Many small names remove single-customer risk. They do not remove the shared cause. In an AI slowdown, small buyers cancel together, usage falls across small operators, rents fall, and used-chip prices fall. That is the exact condition under which the senior slice is most understated, and it is the condition the first pool will be built into.

06Watchpoints

The signs the market is getting there, in the order they are likely to arrive.

Open interest beyond twelve months once the listed futures begin trading. A curve reaching loan maturities is the precondition for every other item on this list.

The first chip loan with a hedge requirement, or a loan-to-value test against a published index. Oil and gas lenders advance more against hedged production; the first compute lender to do the same will set the template for the warehouse lines that precede a pool.

The first public recovery print. One defaulted chip loan, worked out in public, with the sale price disclosed, would do more for the pricing of every senior slice than any model.

The terms of Nvidia’s residual guarantee, when the platform agreements are signed. The trigger price, the share covered and the tenor decide whether the vendor has absorbed resale risk or merely deferred it.

Whether the revenue-floor program restarts, and in what form. The Journal’s report and Nvidia’s response are consistent with a restructuring rather than a withdrawal. The customer-approval terms that raised the antitrust concern were also the diversification terms a lender wanted.

Meta’s third campus, and its spread to Sopaipilla. The Hyperion-to-Sopaipilla gap of 95 basis points is the only clean repricing series in the market. The next print says whether the bid for landlord paper is deepening or tiring.

Oracle’s rating, and any other BBB-minus counterparty behind investment-grade compute paper. The chip market’s grades are borrowed. A downgrade to high yield at a major offtaker would be the first test of what a rating leaning on a take-or-pay contract is worth when the payer’s own paper is junk.

The first true multi-borrower pool, and whether its documents link recovery to defaults. The paper’s testable prediction is that quoted ranges on capped products narrow as strike ladders deepen while uncapped ones stay wide. The first pool’s senior tranche is where that prediction gets its market test.

Solvency II in January 2027. Halving the capital charge on securitizations for European insurers will pull demand toward chip-backed paper at precisely the moment US supervisors are asking what banks’ private-credit counterparties hold.

The sign it is not getting there. More loans running past their contracts at tighter spreads. That is resale and renewal risk coming back in without a price, and it is the one pattern the record already contains.

Sources

The paper: Seung Jung Lee and Sriram Nagaraj, Pricing, Hedging, and Securitizing AI Compute: A Non-Storable Commodity Framework for Infrastructure Risk, Federal Reserve Board, August 2026, SSRN working paper 7342241. A plain-English reading of the paper, Securitizing AI compute: the framework, today’s market, and what must be built, is forthcoming.

Market record, checked September 23, 2026:

• CCIR, Compute Credit Tracker, as of September 21, 2026, and CCIR Research, the rating case goes investment grade
• Structured Finance Association, how data center bonds fit the financing ecosystem, July 23, 2026; Crane Frontier, data center ABS by the numbers, July 28, 2026; Charles River Associates, data center ABS risks, yields and ratings, December 2025
• Meta and Blue Owl, Hyperion joint venture, October 2025; Global Data Center Hub, the $12.3 billion El Paso bond, July 2026; Capacity, Meta’s El Paso bond and AI debt repricing
• Yahoo Finance / PitchBook, Meridian Arc’s $5.7 billion notes, April 17, 2026; PitchBook, data center frenzy stoked by high-yield bond binge, April 22, 2026; Galaxy Digital, $3.507 billion senior secured notes, July 23, 2026; TeraWulf and Fluidstack joint venture notes, December 2025
• CoreWeave, $8.5 billion facility, March 2026, $3.1 billion facility, May 2026, $2.6 billion facility, August 2026, and $3.7 billion convertible notes, September 18, 2026; Global Data Center Hub, CoreWeave’s $8.5 billion quarter, August 5, 2026; Data Gravity, how the GPU became collateral
• IREN, $3.65 billion financing, June 1, 2026; Nebius, $775 million secured facility, July 17, 2026; Lambda, $926 million term loan, August 2026; Nscale, about $3 billion of facilities, August 31, 2026
• Nvidia, financing platforms with six managers, August 10, 2026; Fortune, Nvidia’s spending commitments, August 28, 2026; SemiAnalysis, Nvidia’s backstop universe, September 11, 2026 and Nvidia’s GPU debt backstop, July 6, 2026; Yahoo Finance, on the Wall Street Journal’s report of the program pause, August 28, 2026
• Sascha Steffen, inside Nvidia’s $500 billion AI debt machine, August 14, 2026; Futuriom, the xAI chip vehicle, June 2, 2026; FourWeekMBA, Fluidstack and the Google guarantees, September 19, 2026
• PPC Land, S&P cuts Oracle to BBB-minus, July 9, 2026; Fortune, hyperscaler bond issuance and hidden borrowing, July 31, 2026; KBRA, AI infrastructure commitments and the AI-7, August 26, 2026 and evaluating re-leasing risk, June 22, 2026; Quinn Emanuel, litigation risks in financing AI data centers, March 13, 2026
• American Compute, GPU residual value report, June 15, 2026; Silicon Data, H100 rental price index; CME Group, clearing advisory 26-274, compute futures effective October 5, 2026; Fortune, the race to build compute futures, August 12, 2026
• Reinsurance News, Munich Re-backed chip-loan warranty, March 2026; USD.AI, how it works; McKinsey, the future of AI workloads, February 2026
• Bank of England, Financial Stability Report, July 2026; Federal Reserve, Financial Stability Report, May 2026; IMF, Global Financial Stability Report, April 2026, chapter 1; Chicago Fed, tail risk for banks from generative AI, February 2026; Dallas Fed, AI debt financing and duration supply, February 10, 2026; Akin, SEC staff guidance on data center securitizations, August 2026

Figures marked as reported are press-reported and not confirmed against a primary filing. Stage dates in section 05 are Kinetic Alpha estimates. No misconduct, manipulation or impropriety is alleged or implied against any firm named in this piece. Research and education. Not investment advice, and not an offer of any product or service requiring registration.