← Latent Minds Institute Economics strand · originally published under MO3 Research
Working Paper · No. 2026-08Latent Minds Institute

A Field Guide to GPU Credit Risk

A $7 trillion debt market is forming around an asset most lenders cannot price. This paper builds the mental map: the five risk layers, the ratios that matter, and the nuances that kill loans. By the end you should be able to size, price, and stress a GPU loan yourself.

Companion to Paper 2026-07Draft · Tinker FreelyAuthor: Muhammad Zane Abdullah
01The problem: a trillion-dollar market with one deal template 02The five layers of GPU credit risk 03The two numbers: DSCR & LTV. Size a loan 04Collateral that melts: residual value 05Reading a Neocloud's book: the rating engine 06Building the spread: price the loan 07What actually kills the loan: stress tornado 08Twelve nuances underwriters miss
01

The problem: a trillion-dollar market with one deal template

By 2029, AI debt outstanding will approach $7T, second only to US mortgages. Yet nearly every dollar lent today fits a single template, and the template is running out of road.

The template: A Neocloud signs a 5-year take-or-pay compute contract with an investment grade hyperscaler. Lenders don't underwrite the GPUs, the tokens, or the operator. They underwrite Microsoft's promise to pay, add roughly 90bps for execution risk, and fund at 70-80% loan-to-value. This is why CoreWeave borrowed at 5.9% against a Meta backstop while its own unsecured bonds yield around 10%. That 410bps gap is the market saying: we can price the hyperscaler; we cannot price you.

Three forces that break the template:

ForceWhy it breaks the template
Backstops are finiteHyperscaler balance sheets cannot guarantee trillions. Nvidia stepping in as backstopper of last resort (the "central bank of AI") is a bridge, not a destination, and it takes roughly an 18% revenue share for the privilege.
Demand is short, funding is longInference providers won't sign beyond 1 year; startups want bursts of capacity between funding rounds. A 5-year-offtake-only market rations out exactly the customers growing fastest.
Lenders lack toolsNo public price index history, no residual value curve, no operator quality standard, no demand model. Banks "hide behind the shield" of the IG offtake because they literally have nothing else to look at.

The opportunity, and the subject of this paper, is the discipline that fills that gap: underwriting Neoclouds as merchant businesses, the way banks  learned to underwrite merchant power plants, memory fabs, and ship charters. Every one of those markets started exactly here: long dated assets, volatile spot prices, lenders demanding a utility grade offtake, until someone built the risk framework that let credit flow without one.

Mental anchor for everything that follows: a GPU loan is a project finance deal wearing an equipment finance costume. The collateral matters far less than the cashflows, because the collateral loses 20-30% of its value per year no matter what. Creditors are lending against a revenue stream, secured by a melting ice cube.
02

The five layers of GPU credit risk

Every GPU loan decomposes into five stacked risks. The \ structuring is deciding which layers the deal transfers away (via offtake, backstop, insurance, hedges) and which the lender actually holds. Toggle the deal structures below and watch who holds what.

Risk transfer map · who holds each layer?MODEL

LayerThe questionPrimary evidence
1 · Offtaker creditWill the buyer of the compute pay?Credit rating, financials, contract enforceability, parent guarantees
2 · Market / priceWhat will re-let rates be when contracts roll?Rental price indices, forward curve, supply pipeline, token demand growth
3 · Execution / operatorCan this team stand up and run the cluster?Track record, uptime SLAs, network architecture, ops bench, ClusterMAX-style ratings
4 · Asset / residualWhat is the hardware worth if we repossess?SKU depreciation curves, secondary market prints, next-gen launch cadence
5 · StructuralDoes the paper actually protect us?SPV isolation, cash sweeps, cross-default, security over contracts not just chips, datacenter lease terms

The key insight most newcomers miss: layers 1 and 2 are substitutes. A strong offtake makes market risk irrelevant for its tenor; no offtake means you are underwriting the market itself. Today's lenders can only handle layer 1. The \game is learning to price layer 2, which is exactly what the forward market in Paper 2026-07 exists to make tractable.

03

The two numbers: DSCR & LTV. Size a loan

Everything in GPU credit reduces to two ratios. DSCR (debt service coverage ratio) is cash available for debt service divided by debt payments each period. The covenant floor is 1.30×, and it is tested in the downside scenario, not the base case. LTV (loan-to-value) is loan divided by cluster cost. Market standard is 70-80%, but the binding constraint is almost always DSCR, not LTV. Size a deal yourself:

Loan sizer · 1,024 × GB300 · $60M cluster costMODEL
Max loan @ 1.30× worst-yr
...
Implied LTV
...
Worst-year DSCR
...
Binding year
...

Bars show DSCR per year at the max loan. The red line is the 1.30× covenant. Notice the shape: early years are fat (high rents, low amortization progress), late years pinch as prices decay against fixed debt service. Now push decay to 30%: the max loan collapses even though year-1 economics look identical. The lender's real exposure is the decay assumption, and nobody agrees on it. This single slider is why the market needs a forward curve.

Practitioner shortcut: a quick sanity mark for any deal is payback period = cluster cost ÷ year-1 net cash. Under about 2.5 years, the deal survives almost any decay assumption; over 4 years, you are making a bet on terminal demand. Most disagreements between bulls and bears are really disagreements about which side of 3 years the payback sits.
04

Collateral that melts: residual value

In equipment finance, repossession is the backstop. In GPU finance, the collateral is deflating at silicon speed: each new generation makes the old fleet worth less per FLOP, and a wave of off-lease supply can crater secondary prices exactly when defaults cluster. This is the same wrong-way risk that burned aircraft lenders in 2009 and auto lenders in every recession.

Collateral coverage · loan balance vs. GPU value over timeMODEL
Underwater period
...
Max shortfall
...
Recovery if default yr 2
...

Orange: GPU fleet value. Blue: loan principal outstanding. The shaded gap is the lender's unsecured exposure at each point. Flip to a 40% balloon and watch the underwater window widen. Balloons are how deals "pencil," and they are the single largest hidden risk in the market. Any repayment you rely on after the collateral crosses the loan line is a bet on cashflow, not collateral.

Two conclusions. First, residual value is a cushion, never the thesis: treat any recovery assumption above 40% of original cost past year 3 with suspicion. Second, the depreciation slope is itself uncertain, and the 36-month forward from Paper 2026-07 is the only market cleared estimate of it that could exist. Until then, underwriters triangulate from secondary market prints, rental price indices, and next-generation launch cadence.

05

Reading a Neocloud's book: the rating engine

When there is no hyperscaler shield, the underwriter's real object of study is the Neocloud's customer book, the portfolio of rental contracts. Four dimensions matter: who (credit quality of renters), how long (tenor mix vs. the loan), how concentrated (top-customer share), and how well run (operator quality: uptime, networking, support; the thing ClusterMAX-style ratings measure). Below is a toy scoring engine. Build a book and see how a rating desk would grade it:

Neocloud rating engine (toy scale: AA to CCC)SIGNATURE MODEL
...
Implied standalone grade

The counterintuitive one: longer average tenor is not always better. A book of 5-year contracts to two thinly funded AI startups is worse than a churning 6-month book across forty funded inference companies. The long book has locked in its counterparty risk, while the short book reprices and rediversifies constantly. Tenor is only valuable multiplied by counterparty quality. This is the nuance rating agencies took a decade to learn in shipping.
06

Building the spread: price the loan

Spread over the risk-free rate is built layer by layer. Each risk you hold demands compensation; each risk transferred away removes a block. The two market anchors: hyperscaler backstopped paper at roughly SOFR+225 (about 5.9% all-in) and top-tier Neocloud unsecured at about 10%. Every real deal prices somewhere between. 

Spread builder · all-in yield waterfallMODEL
All-in yield
...
vs. hyperscaler template (5.9%)
...
vs. unsecured ceiling (10%)
...
PBT margin impact*
...

*The borrower's side of the same number: at 70-80% LTV, moving all-in cost from 5.6% to 10% crushes a Neocloud's pre-tax margin from about 15% to about 5%. Financing cost IS the business model, which is why every Neocloud contorts itself into whatever structure prices tightest, and why the underwriter who can justify a tighter spread on a merchant book wins the deal flow.

07

What actually kills the loan: stress tornado

Underwriting is asking "what breaks first?" in order. The tornado below shocks each variable one at a time by a realistic adverse amount and ranks the damage to worst-year DSCR (using your 3 deal). This ordering, not the base case, is the mental map of the market:

Single-factor stress · impact on worst-year DSCRUSES section 3 DEAL

Bars measure DSCR destruction from each standalone shock. The consistent lesson across parameterizations: price/decay risk dominates everything, utilization is second, opex (power) is the sleeper, and rate risk, the thing traditional lenders instinctively model, is usually near the bottom. GPU credit risk is a compute-price problem wearing a credit costume. Adjust the section 3 sliders and watch the ranking hold.

08

Twelve nuances underwriters miss

The checklist of things that don't show up in the model but decide outcomes. Tagged by type: Kill Risk can zero the loan, Price Risk moves economics, Structure is fixable.

1 · Take-or-pay is not take-or-pay until you read the outs Kill risk
Many "committed" offtakes carry performance conditions (delivery deadlines, uptime SLAs, benchmark requirements) whose breach voids the commitment. The offtake is only as strong as the operator's ability to hit its conditions. Underwrite the SLA cure periods and termination triggers, not the headline TCV.
2 · The datacenter lease can outlive the revenue Kill risk
Colocation leases run 7-15 years; GPU economics run 4-6. If the SPV signs the lease, a compute-price crash leaves it paying rent on space housing worthless silicon. Check who holds the lease, whether it is coterminous with the debt, and whether there are sublet or assignment rights.
3 · Prepays create phantom equity Structure
Some Neoclouds collect 50-100% prepayment on 1-year rentals, funding the entire cluster with customer cash ("infinite IRR"). That prepay is a liability: undelivered service. If the operator stumbles, customers claim refunds pari passu with you, or ahead of you. Insist prepays sit in escrow or are recognized ratably.
4 · Power is the opex wildcard Price risk
Power is 30-50% of cash opex and increasingly volatile. A Neocloud with merchant power exposure has a hidden second commodity position. Ask: fixed PPA or floating? What tenor? A great compute hedge with unhedged power is half a hedge.
5 · Vendor concentration cuts both ways Price risk
Nvidia is simultaneously the equipment vendor, sometimes an investor, sometimes the backstop counterparty, and occasionally a customer. This alignment is stabilizing in good times and deeply correlated in bad. A backstop from the entity whose sales depend on the buildout continuing is not the same as an uncorrelated guarantee. Haircut it accordingly.
6 · Depreciation is scheduled; obsolescence is lumpy Price risk
Value doesn't decay smoothly. It steps down at each next-gen launch and each major model-efficiency breakthrough. Underwrite to the product roadmap (roughly annual cadence), and stress a mid-life step-down, not just a smooth curve (see the launch-shock slider in section 4).
7 · Utilization is not revenue quality Price risk
95% utilization sold on spot at distressed prices is worse than 80% on funded 1-year contracts. Always decompose revenue into price × volume × counterparty. High utilization can be a symptom of underpricing.
8 · Security over chips is weak; security over contracts is strong Structure
Repossessing 1,024 GPUs racked in someone else's datacenter is slow, costly, and destroys the very cashflow you wanted. The valuable collateral is the assignment of offtake contracts and receivables plus step-in rights to operate the cluster. Perfect those first.
9 · Cross-default with the parent Structure
Neoclouds finance cluster-by-cluster in SPVs. Verify true isolation: no cross-collateralization, no parent guarantee you are unknowingly junior to, no shared bank accounts. A clean SPV lets a good project survive a bad parent; a leaky one imports every other lender's problems.
10 · The 1.30× DSCR is tested on the wrong scenario if you let it be Structure
Sponsors will present DSCR on the base case. The covenant should reference a defined downside: backstop-activated pricing (if there is one) or a P90 price path off an index. Which scenario defines the test is the single most negotiated, and most consequential, line in the term sheet.
11 · Tokenomics is the demand model, and it is learnable Price risk
End demand equals tokens consumed times price per token, feeding back into GPU-hours. Model-efficiency gains cut GPU-hours per token but have historically expanded total consumption faster (Jevons). An underwriter needs a view on inference demand growth, not just training capex. That is what fills the book when contracts roll.
12 · Correlation is the portfolio killer Kill risk
Every risk in this paper (rental prices, residual values, counterparty health, secondary-market liquidity) is driven by one factor: aggregate AI demand. In a downturn they fail together, like mortgage tranches in 2008. Diversification within GPU credit is mostly illusory; sizing the whole book to survive a correlated drawdown is the discipline that separates survivors.

That is the map. Two ratios (3), five layers (2), one melting collateral curve (4), a book-quality lens (5), a spread stack (6), a stress ordering (7), and twelve traps (8). Anyone holding this frame can sit across from a Neocloud's CFO, or a credit committee, and know which questions matter. The tools to answer them at scale (the index, the curve, the rating engine) are the businesses waiting to be built.