AI Infrastructure Financing: State of the Market

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.

Section 1

Market size without false equivalence

Comparing peak subprime stock, or a six-year fiber buildout, to a single year of AI debt is not a useful signal. The charts below put the three buildouts on multi-year totals first, then show the AI path year by year through 2027 so the cumulative becomes visible. Debt issuance is kept separate from physical capex, because they answer different questions.

Three buildouts, multi-year totals

like-for-like attempt

Click a bubble to pin its details here.

AI DC / hyperscaler capex

Cumulative 2025 + 2026e + 2027e, physical buildout approx

~$2.2T

Peak U.S. subprime

Outstanding stock at the 2007 peak, not cumulative spend unverified

~$1.3T

Telecom / fiber buildout

Cumulative capex + debt, 1996 to 2002 (~6 years) approx

~$750B

AI-related debt issuance

Cumulative 2025 + 2026e + 2027e, financing only approx

~$0.6T to $0.9T

Read the AI capex bar as the physical analogue to fiber. Read the AI debt bar as the financing that sits under part of that buildout (much of hyperscaler spend is still equity-financed cash flow). Subprime remains a peak outstanding stock: useful for systemic scale, not as a spend total. On a three-year cumulative-capex basis, the AI buildout already clears both historical comparators; on a debt-only basis it does not.

AI path through 2027: year by year, then cumulative

forward estimates

Physical buildout (left) and debt financing (right). 2025 figures are realized-or-near-realized; 2026 and 2027 are Street estimates that will move with earnings guidance. Cumulatives are running totals, not annual rates.

Hyperscaler / AI DC capex

2025

Big-5 aggregate, ~$380B approx

~$380B

2026e

Guidance cluster ~$625B to $725B, shown at $700B approx

~$700B

2027e

Goldman ~$1.1T (bull case to $1.4T) approx

~$1.1T

Cumulative through 2027e

$380B + $700B + $1.1T derived

~$2.2T

AI-related debt issuance

2025

Broad AI-related debt, low hundreds of billions approx

~$200B

2026e

Base case: issuance grows with buildout; shown at $250B scenario

~$250B

2027e

Base case continues; shown at $300B scenario

~$300B

Cumulative through 2027e

Base path ~$750B; flat path ~$600B; fast path toward $1T scenario

~$750B

Comparator	Window	What is measured	Total	Implication vs AI
Telecom / fiber	1996 to 2002	Cumulative capex + debt over ~6 years	~$0.5T to $1T (mid ~$750B)	AI physical capex over just three years already exceeds the midpoint, and may clear the top of the range by end-2027.
U.S. subprime	Peak 2007	Outstanding stock, not spend	~$1.3T	AI cumulative capex through 2027e (~$2.2T) is larger; AI cumulative debt through 2027e (~$0.6T to $0.9T) is still smaller.
AI DC buildout	2025 to 2027e	Hyperscaler / AI data-center capex (physical)	~$2.0T to $2.5T (mid ~$2.2T)	The relevant size question is no longer "is one year big?" It is whether funding, power, and demand keep this path intact.
AI debt financing	2025 to 2027e	Broad AI-related debt issuance (financing)	~$0.6T to $0.9T base	Still below both historical comparators on a debt-only basis. Capex is mostly equity-financed cash flow today; that mix can change.

Capex path sources: 2025 ~$380B Big-5 aggregate from recent analyst summaries; 2026e ~$625B to $725B guidance cluster (shown at $700B); 2027e Goldman Sachs ~$1.1T with a bull case to $1.4T. Longer-horizon context (not charted): Goldman ~$5.3T hyperscaler capex 2025 to 2030; McKinsey ~$5.2T to $6.7T global data-center investment need by 2030. Debt 2026e/2027e figures are scenarios, not Street consensus prints. DC securitization alone is smaller: $27B in 2025 (KBRA), with JPMorgan / Bloomberg discussing $30B to $40B per year potential in 2026 to 2027.

Currently: logarithmic display.

Issuer and ABS stocks (detail)

components verified

CoreWeave DDTL 1.0 to 5.0

Identified commitment capacity components verified

$27.1B

CoreWeave total debt

Net of discount and issuance costs, Mar. 31, 2026, 10-Q verified

$24.9B

S&P-rated U.S. DC ABS

Outstanding, Mar. 31, 2026, per S&P paywalled

>$23B

Beignet / Meta project notes

Single project's secured debt unverified

$27.3B

These are issuer- and project-level stocks, not the buildout total. Commitment capacity includes undrawn amounts. 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. One project (Beignet) now equals the entire 2025 securitization market.

2025 debt issuance detail

one-year flows

Broad AI-related debt issuance

2025, "low hundreds of billions" shown at ~$200B approx

~$200B

Hyperscaler public bonds

2025 issuance, per Reuters unverified

~$120B

Data-center securitization

2025 new issuance, KBRA verified

$27B

S&P-rated DC ABS issuance

2025, narrower rated universe paywalled

$9.25B

Kept as a detail panel, not the headline comparison. The gap between KBRA's $27B and S&P's $9.25B shows how "market size" moves with the measurement perimeter.

What the numbers actually say. On a cumulative physical-capex basis through 2027e, the AI data-center buildout is already in the same league as (and likely larger than) both the fiber boom and peak subprime stock. On a debt-only basis it is still smaller, because so much of the spend is still equity-financed cash flow. The credit question is therefore not "is the buildout small?" It is whether funding structure, holder quality, and contract durability hold up as the stock compounds and as more of the spend migrates into debt markets.

Section 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

Observed

Contract 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.

Observed

Investment 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.

Important

Capacity 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.

Section 3

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.

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)

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.

Section 4

Contract quality is a waterfall, not a label

"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.

Strongest

Contract-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

Middle

Capacity 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

Weakest

Usage-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.

Section 5

Funding transmission: where a repricing becomes a forced sale

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.

Section 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

Dimension	Subprime / 2008	Telecom / fiber	AI infrastructure, July 2026
Primary error	Underwriting 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.
Collateral	Homes 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.
Obligors	Millions of households with weak documentation.	Carriers and startups, many speculative.	Concentrated corporate and AI-lab counterparties. Better disclosure, far higher single-name concentration.
Funding	Repo, 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.
Amplifier	Forced 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 mode	Mortgage 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.

Section 7

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.

Section 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.
Headline comparison uses multi-year cumulatives; one-year debt flows and issuer stocks are demoted to detail panels.
Forward 2026e/2027e figures are Street estimates or scenarios and are tagged as such; 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.

Source ledger

Sources and verification status

verified fetched and checked July 20, 2026 research base supplied fact base unverified not independently checked analysis framework or judgment

verified CoreWeave Form 8-K, DDTL 4.0 (Mar. 31, 2026)
$8.5B, SOFR + 2.25% floating (0% floor), fixed-rate formula, June 2027 commitment termination, March 31, 2032 maturity, CCAC VIII collateral and equity pledge, limited "bad acts" guarantee, 1.15x DSCR, MUFG and Morgan Stanley as arrangers. Names no customer, rating, or anchor investor.
verified CoreWeave DDTL 5.0 closing press release (May 18, 2026)
$3.1B, SOFR + 4.50% after 50 bps tightening, Ba2 / BB+, two large non-investment-grade customers, ~5.5-year maturity, CoreWeave Financing DDTL V, LLC, first publicly syndicated HPC-backed DDTL.
verified CoreWeave Q1 2026 Form 10-Q
XBRL debt tables: total debt $24,859M net at Mar. 31, 2026; DDTL 1.0 $1,438M at 15% effective; DDTL 2.0 $4,425M at 11%; DDTL 2.1 $3,000M at 9%; DDTL 3.0 $1,700M at 9%; OEM and software-license financing $5,036M at 10%.
verified KBRA, Data Center Leases: Variations on Established Themes (Mar. 10, 2026)
$27B data-center securitization issuance in 2025; hyperscale net-lease structures; termination, contraction, and assignment rights as key credit variables; power as the economic driver.
unverified CoreWeave 2025 Form 10-K
Cited for DDTL 1.0 margin (9.6196%), DDTL 2.0 commitment ($7.6B, S + 6.0% to 13.0% by customer credit), DDTL 2.1 terms, and six-year equipment life. Consistent with the verified 10-Q balances but not independently fetched.
research base CoreWeave DDTL 3.0 (Jul. 2025)
$2.6B at SOFR + 4.00%, OpenAI-linked; from the supplied research base, consistent with the verified 10-Q carrying amount.
research base DDTL 4.0 ratings and anchor
A3 / A(low) ratings and Blackstone Credit & Insurance anchoring, per press coverage in the research base. Not in the 8-K text.
research base Market aggregates
JPMorgan / Bloomberg coverage of data-center securitization at $30B to $40B per year potential for 2026 to 2027; 2025 AI-related debt issuance in the low hundreds of billions; GPU advance rates commonly 50% to 70% of FMV; three-year residual outcomes debated at roughly 10% to 60% of cost.
unverified Cumulative AI buildout path (2025 to 2027e)
Hyperscaler / AI DC capex: ~$380B (2025 Big-5 aggregate, analyst summaries), ~$625B to $725B (2026e guidance cluster, shown at $700B), ~$1.1T (2027e Goldman Sachs, bull case to $1.4T); derived cumulative midpoint ~$2.2T. Longer-horizon context: Goldman ~$5.3T hyperscaler capex 2025 to 2030; McKinsey ~$5.2T to $6.7T global data-center investment need by 2030. AI-related debt cumulative base path ~$750B assumes ~$200B / $250B / $300B across 2025 to 2027e (scenarios). Not independently verified line by line.
unverified S&P Global Ratings, North America Data Center ABS Roundup Q2 2026
More than $23B outstanding S&P-rated U.S. DC ABS and $9.25B of 2025 rated issuance, Registration wall prevented direct verification.
unverified Beignet Investor LLC (S&P), Amazon 2025 10-K, NVIDIA capacity agreement 8-K, IMF GFSR Oct. 2007, Reuters hyperscaler bond coverage
$27.3B Meta / Blue Owl project notes; five-to-six-year server lives; $6.3B NVIDIA residual-capacity backstop through April 2032; ~$1.3T peak subprime estimate; ~$120B 2025 hyperscaler bond issuance. Not independently checked.
analysis All model outputs, the funding-transmission framework, contract-tier hierarchy, crisis scores, and comparative judgments
Analytical constructions on the fact base, not observed market data.