Interactive companion: AI Infrastructure Financing report (Agent Mode exports and light/dark Display on the live page). Essay companion: Is the AI Infrastructure Buildout a Bubble?.

This paper is a citation-aware research framework for AI physical infrastructure credit. Every major claim should be read against the interactive report's provenance badges. It is not investment advice.

Executive Summary

The asset class is maturing, but the evidence is structure-specific.

CoreWeave's financing history is the clearest public record of GPU credit moving from expensive private lending toward a segmented institutional market. Its 2023 DDTL 1.0 still carried a 15% effective interest rate as of March 31, 2026, per the Q1 2026 10-Q. DDTL 3.0 closed at SOFR + 4.00% in July 2025 to support an OpenAI-linked contract. DDTL 4.0 closed March 30, 2026: an $8.5B facility at SOFR + 2.25% floating, maturing March 31, 2032, secured by substantially all assets of CoreWeave Compute Acquisition Co. VIII, LLC, with only a limited "bad acts" parent guarantee. Press coverage reported A3 / A(low) ratings and anchoring by Blackstone Credit & Insurance; the 8-K itself names neither a rating nor the customer. Seven weeks later, DDTL 5.0 broadened distribution as the first publicly syndicated HPC-backed delayed-draw term loan, and the market drew a line: $3.1B at SOFR + 4.50%, rated Ba2 / BB+, backed by contracts with two large non-investment-grade customers. Same physical collateral, different contracts and counterparties, a different credit regime.

The closest historical rhyme is not "subprime again." Current AI infrastructure financing has productive assets, identifiable corporate counterparties, amortization, reserve mechanics, and significant long-duration capital. It does not yet show the scale, synthetic multiplication, or pervasive overnight funding that made mortgage credit systemically explosive. The stronger analogy is a hybrid of telecom overbuild, project finance, private credit, and equipment finance. The danger appears if financing availability begins to drive construction, contracts prove less durable than their headline value, and mark-sensitive intermediaries or levered funds hold more exposure than public disclosures reveal.

Bottom line: DDTL 4.0 demonstrates that one tightly structured, contract-backed GPU financing can reach investment grade. It does not establish that GPUs as a collateral class are investment grade. DDTL 5.0's below-investment-grade ratings, wider spread, and explicitly non-investment-grade customers make that distinction a matter of public record, not analyst opinion.

1. Market Size Without False Equivalence

The bubble map gives the headline size comparison: area is proportional to dollar size, and hovering or clicking any bubble shows its basis of measurement. The paired bar charts then separate stocks and commitments from annual issuance, because those are not the same kind of number. Bar widths are computed from the values; the logarithmic default keeps small categories visible, and the linear toggle shows the honest disproportion.

The size of things: six markets, one scale not apples-to-apples

Click a bubble to pin its details here.

These categories mix an outstanding stock (subprime), cumulative multi-year capex (telecom), annual flows (AI debt issuance, data-center securitization), and commitment stocks (CoreWeave capacity, one Meta project). Comparing a stock with a flow understates the eventual AI number if issuance persists at current rates, and overstates it if issuance stalls. That is why the bar charts below keep the two kinds of number apart.

Switch to linear scale Currently: logarithmic display.

Stocks and commitment capacity not directly comparable

Peak U.S. subprime mortgagesOutstanding stock, 2007 unverified

$1.3T

Telecom / fiber buildoutCumulative capex + debt, 1996 to 2002, $0.5T to $1T range shown at midpoint approx

~$750B

CoreWeave DDTL 1.0 to 5.0Identified commitment capacity components verified

$27.1B

CoreWeave total debtNet of discount and issuance costs, Mar. 31, 2026, 10-Q verified

$24.9B

S&P-rated U.S. DC ABSOutstanding, Mar. 31, 2026, per S&P paywalled

$23B

Commitment capacity includes undrawn amounts and is not balance-sheet debt: the verified 10-Q net total of $24.9B also includes ~$5.0B of OEM and software-license financing and predates full DDTL 5.0 drawdown. Peak subprime is an economy-wide stock, shown for scale context only.

Annual issuance, 2025, plus one project for scale flows vs one stock

Broad AI-related debt issuance2025, "low hundreds of billions" shown at ~$200B approx

~$200B

Hyperscaler public bonds2025 issuance, per Reuters unverified

~$120B

Data-center securitization2025 new issuance, KBRA verified

$27B

Beignet / Meta project notesSingle project's secured debt, a stock shown for scale only unverified

$27.3B

S&P-rated DC ABS issuance2025, narrower rated universe paywalled

$9.25B

JPMorgan / Bloomberg coverage discussed data-center securitization reaching $30B to $40B per year in 2026 to 2027. The gap between KBRA's $27B and S&P's $9.25B shows how "market size" moves with the measurement perimeter. One project (Beignet, purple) now equals the entire 2025 securitization market: single names are reaching systemic-attention size before the asset class does.

What the numbers actually say. AI infrastructure credit is already large enough to matter to investment-grade, high-yield, private-credit, project-finance, and structured-finance investors, and is compounding. It remains far smaller than the mortgage complex that entered 2008. But individual projects and issuer facilities are now large enough to create concentration, syndication, and refinancing problems on their own, and the outstanding stock of the whole complex is hard to measure with confidence, which is itself a risk flag.

2. The CoreWeave Financing Ladder

The repricing is real, but the facilities are not fungible. Read this as market segmentation by contract and counterparty quality, not as a universal decline in the risk premium on GPUs.

DDTL 1.0 $2.3B ~15% eff.

  • July 2023
  • 15% effective rate confirmed in Q1 2026 10-Q; $1.44B still outstanding
  • Guaranteed by CoreWeave
  • High-cost private credit

rate verified

DDTL 2.0 $7.6B S + 6.0-13.0%

  • May 2024
  • Pricing tiered by customer credit quality per the 2025 10-K
  • $4.4B carrying, 11% effective (10-Q)
  • Contract quality drove the loan price

balance verified terms unverified

DDTL 2.1 $3.0B S + 4.25%

  • September 2025
  • Incremental tranche
  • $3.0B carrying, 9% effective (10-Q)
  • Same secured equipment-finance architecture

balance verified margin unverified

DDTL 3.0 $2.6B S + 4.00%

  • July 2025
  • Supports OpenAI-linked contract
  • $1.7B carrying, 9% effective (10-Q)
  • Parent guaranteed

research base balance verified

DDTL 4.0 $8.5B S + 2.25%

  • March 2026, matures 2032
  • Non-recourse with "bad acts" carve-outs; DSCR covenant 1.15x
  • Customer not named in 8-K
  • A3 / A(low) and Blackstone anchor per press coverage

8-K verified

DDTL 5.0 $3.1B S + 4.50%

  • May 2026, ~5.5yr
  • Ba2 / BB+
  • Two large non-investment-grade customers
  • First publicly syndicated HPC-backed DDTL; tightened 50 bps in syndication

press release verified

ObservedContract credit changes the loan DDTL 2.0 reportedly priced from SOFR + 6.0% for specified investment-grade customers to SOFR + 13.0% for non-investment-grade contracts. The financing documents themselves reject the idea that the GPU alone sets the price. The 10-Q's verified effective rates (15%, 11%, 9%, 9%) fall monotonically with vintage and structure.

ObservedInvestment grade is structure-specific DDTL 4.0 is non-recourse and contract-backed and reached IG. DDTL 5.0 broadened distribution and gained tradability but stayed below IG, with the press release itself citing two non-investment-grade customers. Tradability and rating quality did not arrive together.

ImportantCapacity is not exposure The $27.1B sum is maximum identified DDTL commitment capacity. Verified 10-Q total debt was $24.9B net at March 31, 2026, including ~$5.0B of non-DDTL OEM financing, and predates full DDTL 5.0 drawdown. Neither number alone describes lender risk.

3. Recovery and Required-Spread Framework

Interactive recovery and required-spread model

A transparent stress framework, not a pricing engine. Residual value is decomposed into disclosed accounting life and an explicit market discount, amortization runs before the stress date, and the verdict maps your assumptions onto real observed prints.

Project finance lens ABS lens

Bull: strong contract Base case Bear: weak contract 2008-style funding stress Reset

Accounting useful life

CoreWeave discloses six years for technology equipment; Amazon five to six years for servers (per issuer disclosures; Amazon figure not independently checked).

Stress date

Years after origination, straight-line book depreciation assumed.

Market discount to book

Generation obsolescence, utilization, removal cost, secondary-market depth. This slider is the residual-value debate.

Initial advance rate

GPU lending commonly cited at 50% to 70% of FMV; stronger structures push higher.

Annual principal amortization

Share of beginning balance repaid each year before stress. Real DDTLs amortize monthly.

Cumulative default probability

User assumption over the stress horizon, not a rating-implied figure.

Contract recovery credit

Recoverable share of outstanding debt from assigned contract cash flows after default. High for hard take-or-pay with a solvent counterparty, near zero for usage-based revenue.

Liquidation cost

Removal, transport, remarketing, downtime, legal, enforcement.

Liquidity premium

Spikes in funding stress even when collateral is unchanged. The 2008 lesson.

Par anchor (reference print)

DDTL 4.0 cohort, S + 225 (IG, insurer-anchored) DDTL 3.0 cohort, S + 400 (concentrated counterparty) DDTL 5.0 cohort, S + 450 (public syndication, BB)

Which observed print marks at par. Structural premium is fixed at 125 bps and spread duration at 2.5 to keep the model from becoming a tautology dial.

Opening loan

per $100 original cost

Stress-date balance

after amortization

Gross book value

straight-line basis

Net collateral value

after market + liquidation haircuts

Recovery rate

collateral + contract credit

Annual expected loss

horizon average

Required spread

EL + liquidity + 125 structural

Indicative price

vs selected par anchor

Breakeven GPU value

min gross resale to cover debt

Stress balance = Opening loan × (1 − annual amortization)^stress years Book value = Original cost × max(0, 1 − stress years ÷ useful life) Net collateral = Book value × (1 − market discount) × (1 − liquidation cost) Recovery = min(stress balance, net collateral + contract credit) ÷ stress balance Annual EL ≈ cumulative PD × loss given default ÷ stress years Required spread ≈ annual EL + liquidity premium + 125 bps structural premium Indicative price ≈ 100 − 2.5 × (required spread − reference spread) ÷ 100

Limitations, stated plainly: no waterfall, tax, swap, reserve account, covenant, draw schedule, construction risk, cure period, or customer-default correlation is modeled. Contract recovery is a user assumption, not a legal conclusion. Accounting life is not market value. Structural premium (125 bps) and spread duration (2.5) are fixed by design. Use this to find which assumptions dominate outcomes, not to estimate a tradable fair spread. All outputs are illustrative analysis.

4. Contract Quality Is a Waterfall

"Take-or-pay," "reservation," and "backlog" are starting points. Lenders need to know whether cash survives delivery failures, amendments, disputes, setoff, bankruptcy, and assignment. The market already prices the difference.

StrongestContract-backed project credit

  • Solvent, preferably rated counterparty

  • Firm minimum payment with narrow termination rights

  • Direct assignment to lenders and enforceable step-in rights

  • Delivery milestones satisfied or tightly funded

  • Restricted amendment, setoff, and netting rights

  • Debt amortizes inside the contracted term

MiddleCapacity reservation

  • Payments may hinge on availability or acceptance tests

  • Renewal cliffs and volume step-downs create tail risk

  • Service-level credits erode cash flow

  • Assignment may need customer consent

  • Construction and interconnection risk remain material

WeakestUsage-based or forecast backlog

  • Revenue depends on actual consumption

  • Customer can optimize, migrate, or internalize workloads

  • Low utilization exposes lenders directly to asset value

  • Headline backlog can exceed legally unavoidable payments

  • Residual-value and refinancing assumptions become primary

The evidence for the hierarchy

verified The DDTL 5.0 press release itself distinguishes its two non-investment-grade customers; DDTL 4.0's 8-K conditions events of default on "certain material contracts." The 2025 10-K reportedly discloses tiered DDTL 2.0 pricing by customer credit (S + 6.0% to 13.0%, not independently checked). These are direct observations, not analogies.

What KBRA's lease research confirms

verified Power, not space, drives economics; most hyperscale leases are net leases whose cost-allocation details set margin stability; absolute triple-net structures shift life-cycle capex to tenants, aiding near-term cash flow but increasing residual risk; and expanded termination, contraction, and assignment rights reduce cash-flow visibility, especially in single-tenant assets.

Do not overstate the named counterparty

verified The DDTL 4.0 8-K describes a customer contract but names no customer, no rating, and no anchor investor. "Meta-backed" and "A3 / A(low), Blackstone-anchored" derive from press coverage and the research base. A careful reader should hold those attributions one notch looser than the filing facts.

Hidden correlation. The same event can hit the customer, the borrower, and the collateral simultaneously. A sharp improvement in accelerator efficiency or a capex slowdown could compress compute prices, weaken a neocloud customer's economics, trigger contract disputes or non-renewals, and depress used-GPU values at once. Treating counterparty default and collateral loss as independent, as simple models do, understates tail risk. In the model above this appears as moving PD, market discount, and contract credit adversely together, which is exactly what the 2008-style preset does.

5. Funding Transmission

Public disclosures identify arrangers, anchors, ratings, and some structures. They do not provide a holder-level map. The honest framework is functional: who originates, who warehouses, who finances the buyers, and who can hold through a mark. No percentages are offered because none can be verified.

Asset and contract SPV

GPU servers, data-center equipment, customer contracts, reserve accounts, pledged equity. DDTL 4.0's borrower structure is disclosed in the 8-K.

disclosed

➞

Transit balance sheets

Arranging banks (MUFG and Morgan Stanley on DDTL 4.0 and 5.0, disclosed), bridge lenders, warehouse facilities, dealer inventory, fund-finance providers.

partly opaque

➞

End holders

Insurers, asset managers, private-credit funds, loan investors where eligible, pensions, and other institutional accounts.

shares unknown

Run risk Highest where long assets meet short or mark-to-market liabilities: warehouse lines, repo-like arrangements, NAV loans, marginable fund leverage. These matter more than the identity of the ultimate pension or insurer beneficiary.

Valuation risk Private marks can delay recognition, but delayed recognition is not loss absorption. Covenant tests, borrowing bases, ratings, and refinancing can force an economic mark even while accounting stays smooth.

Distribution risk DDTL 5.0 created a publicly syndicated, tradable HPC-backed loan. That improves price discovery and breadth, and it also transmits any repricing faster than a private buy-and-hold facility would.

Why this matters. Forced selling needs three ingredients: assets marked optimistically, funding that can be withdrawn quickly, and holders who cannot wait. 2008 had all three at enormous scale because repo and SIV funding rolled nightly against mismarked collateral. Today's chain has less of ingredients two and three at the end-holder level, which is real structural progress. But every dollar in a warehouse, on a dealer pad, or in a levered fund can be told to sell exactly when GPU marks are falling.

The missing dataset is itself the finding. There is no public consolidated view of AI-infrastructure exposure by bank warehouse, fund leverage, insurer account, loan fund, or dealer inventory. This report states the gap rather than filling it with illustrative percentages that screenshots would turn into facts. Distributed ignorance about who holds what, at what mark, with what leverage, is precisely how small markets produce outsized damage.

6. Crisis Comparison by Transmission Mechanism

The relevant question is not whether AI infrastructure "looks like 2008." It is which crisis ingredients are present, how strong they are, and where confidence is low.

analyst judgment The scores below are qualitative judgments on the assembled evidence, not measurements. Reasonable analysts will disagree by a notch either way.

Underlying asset overbuild Medium

Collateral obsolescence High

Obligor opacity Low

Customer concentration High

Runnable short-term funding Low-Med, unmeasured

Structured-finance complexity Medium

Synthetic multiplication Low

System-wide scale Low-Med

DimensionSubprime / 2008Telecom / fiberAI infrastructure, July 2026

Primary errorUnderwriting and correlation were mispriced.Demand and pricing were extrapolated into overbuild.Contract durability, utilization, build timing, and hardware economics may be extrapolated faster than they are tested.

CollateralHomes with slow physical decay but highly leveraged prices.Long-lived networks whose economic value collapsed under excess capacity.Short-lived accelerators plus long-lived power and real estate. Different assets should not be modeled as one pool.

ObligorsMillions of households with weak documentation.Carriers and startups, many speculative.Concentrated corporate and AI-lab counterparties. Better disclosure, far higher single-name concentration.

FundingRepo, ABCP, SIVs, dealer balance sheets, bank capital.Corporate bonds, bank loans, vendor finance, equity.Private credit, project finance, secured DDTLs, public loans, ABS, project bonds, equipment finance, hyperscaler corporate debt.

AmplifierForced deleveraging and synthetic exposure.Capital-market closure and operating defaults.Construction delays, contract disputes, lower utilization, collateral markdowns, refinancing gaps, leverage on holders.

Likely first failure modeMortgage delinquencies and warehouse failures.Carrier defaults and dark-fiber repricing.Project delay or counterparty deterioration forcing a financing reset, then a wider mark across comparable GPU and data-center paper.

"Not the same as 2008" does not mean "not risky." 2008 was not caused by mortgages; it was caused by optimistic collateral marks meeting runnable funding inside structures nobody could see through. A broad AI credit correction is more likely to resemble a telecom and private-credit repricing than the opening phase of 2008. That judgment reverses if evidence emerges of large runnable warehouse exposure, concentrated bank guarantees, synthetic overlays, or widespread leverage at end holders. The correct posture is neither panic nor comfort: insist on measured residuals, holder data, and contract-level disclosure while spreads are still tight enough to demand them.

7. What to Monitor

What to monitor before spreads tell you

  1. Contract conversionRPO and backlog converted to cash, by customer and delivery cohort, rather than aggregate backlog growth.

  2. Delivery slippageEnergization, interconnection, GPU delivery, and customer acceptance milestones.

  3. Contract amendmentsTermination rights, price resets, capacity reductions, assignment restrictions, service-level credits.

  4. Debt paydown vs asset agingOutstanding principal by GPU generation and contract term, not only total debt.

  5. Secondary-market dispersionDDTL 5.0 loan prices, DC ABS spreads, project bonds, and neocloud unsecured debt moving apart or together.

  6. Funding-chain leverageWarehouse utilization, NAV and subscription lines, insurer capital charges, bank risk-weighted assets, dealer inventory.

  7. Used-equipment evidenceActual transaction prices by accelerator generation, configuration, location, warranty, removal cost.

  8. Financing-led constructionProjects launched because cheap capital is available rather than because contracted demand and power are secured. The first clear instance imports the core dotcom error.

8. Assumptions, Limitations, and Verification

Assumptions and limitations

  • The model is a single-horizon stress test; real facilities carry covenants, reserves, and cure mechanics that improve outcomes versus this framework.

  • FMV is assumed equal to cost at origination.

  • Structural premium (125 bps) and spread duration (2.5) are fixed; the par anchor is a user choice among three observed prints.

  • Chart categories still mix stocks, flows, and commitments across panels; each bar states its basis.

  • Paywalled or unfetched sources are tagged unverified throughout.

  • No claim is made about any issuer's solvency. This is market-structure analysis, not investment advice.

Verification checklist

  • All ten controls update the nine outputs and verdict immediately.

  • Presets and Reset restore expected values; active preset highlights; moving any slider clears it.

  • Lens toggle adjusts contract credit, liquidity, and par anchor, and updates the note.

  • Base case, PF lens, prints near S + 250 against the S + 225 anchor (price ~99.4).

  • Bubble areas scale with the square root of dollar values; hover shows a tooltip; click pins details.

  • Linear/log toggle recomputes every bar width from its data value.

  • Dark mode, print stylesheet, and copy-summary function work; layout holds at 375 px.

  • Page runs offline; the only network activity is following source links.

Related surfaces