Who has to pay when AI returns disappoint?

Protection can contain a loss—or move it

Disappointing AI investment returns do not automatically become creditor losses. Some contracts genuinely protect lenders. But the protection can leave a large customer with uneconomic purchases, a sponsor with lower returns or a supplier with a contingent obligation. It does not make the original investment good.

The strongest favorable finding is specific: Morningstar DBRS identifies Meta as the take-or-pay customer supporting CoreWeave’s DDTL 4.0 and expects full amortization from contracted operating cash, without refinancing or re-leasing. That is a stronger proposition than selling aging GPUs to repay a balloon. It is an April underwriting expectation, not observed September collections or current covenant compliance.

Sources: Morningstar DBRS initial credit analysis; DDTL 4.0 executed credit agreement; CoreWeave DDTL 4.0 overview.

The adverse case is specific too. Delayed acceptance, service deductions, unhedged costs, reserve replenishment or a pause in further lending can create pressure before final maturity and without AI use collapsing. The relevant number is the borrower’s cash available on its payment dates, not a group-wide earnings figure divided by a debt headline.

The whole-pool conclusion remains a real commercial expansion with uneven investment prospects. This investigation strengthens the case for containment in selected well-supported projects, while documenting continuing financing dependence during expansion and correlated exposures whose ultimate holders remain incompletely known. Lower retail AI prices do not automatically reprice a fixed cloud invoice.

Scope: the U.S.-centered commercial generative/frontier-AI build-out, with global finance and suppliers where material. This is documentary economic analysis—not a legal opinion, crash forecast, investment recommendation or census of AI lending.

Sources: DDTL 4.0 executed parent guaranty; Filed form of residual-value guaranty; Meta June 2026 Form 10-Q.

Expansion and repayment still require real financing

CoreWeave reported $4.653bn of revenue and $3.663bn of operating cash flow for January–June 2026. OCF already reflects working-capital timing and expensed cash interest. It is not gross customer collections, and the consolidated accounts are not the project borrower’s covenant certificate.

Sources: CoreWeave June 2026 Form 10-Q.

This bridge shows financing dependence during expansion, not that every operating project loses money. Refinancing an existing loan produces a principal outflow and financing proceeds; counting every such repayment as recurring amortization would overstate structural weakness. Ignoring principal because adjusted EBITDA is large makes the opposite mistake.

The full cash identity is different again: $3.663bn operating inflow, $14.874bn investing outflow and $13.985bn financing inflow increased cash including restricted cash by $2.774bn to $6.904bn. Of that ending total, $1.380bn was separately restricted. Financing inflow is not customer revenue or an unrestricted parent cash buffer.

An advance is one receipt, even when revenue arrives later

A customer prepaying $100 supplies cash now. Recognizing $20 of that advance as next year’s service revenue does not create another $20 receipt. The supplier must still fund the service. Advance payment can protect construction funding without making later revenue equivalent to later cash.

CoreWeave’s deferred-revenue balance rose $1.507bn, while the cash-flow adjustment was $1.365bn; $0.554bn of revenue came from the opening deferred balance. These are different measures, not interchangeable gross-prepayment estimates. The missing bridge is net deposited customer collections after credits, matched to the borrower’s own costs and accounts. No “recurring OCF” is manufactured by stripping the whole deferred-revenue movement.

Sources: CoreWeave June 2026 Form 10-Q.

A priced offering is still not a cash receipt

The June principal schedule remains $35.551bn, versus a $35.068bn carrying amount. It includes $4.413bn in the remainder of 2026 and $6.184bn in 2027, or $10.597bn together. Those annual buckets are not a September next-24-month cash schedule. Later August drawings were disclosed separately; the intervening repayments, spending and other changes do not support a complete cutoff roll-forward.

The 17 September proposal progressed to a $3.7bn priced convertible offering on 18 September. Expected settlement was 22 September 2026, subject to conditions. Estimated cash after purchasers’ discounts and capped calls is $3.1457bn before other expenses, assuming closing and no exercise of the $0.5bn purchaser option. The full face remains a claim: its 2.875% coupon implies $106.375m of annual cash interest. Capped calls can mitigate specified dilution; they do not repay principal.

The separate authority to distribute up to 35 million shares is not proof of sales. Prospective capital access is favorable financing evidence, but neither a priced transaction nor a share-sale authority is current customer cash.

Sources: CoreWeave June 2026 Form 10-Q; CoreWeave $3.7bn note-pricing release; CoreWeave 17 September current report.

Start with the customer—not the second-hand GPU price

The legal borrower is CoreWeave Compute Acquisition Co. VIII, LLC. DDTL 4.0 has an $8.5bn facility ceiling and $2.837bn of principal reported at 30 June 2026. A delayed-draw ceiling is not cash already drawn. The issuer describes installation funding, later stabilization-related financing, monthly amortization and a final maturity of 31 March 2032.

DBRS identifies the supporting service agreement as a Meta take-or-pay capacity contract. In its base case, contracted operating cash fully amortizes the loan without re-contracting or refinancing. Availability deductions and power costs, rather than Meta’s utilization alone, are key sensitivities. DBRS judges installation riskier and acknowledges reserve limitations relative to its methodology.

Sources: Morningstar DBRS initial credit analysis; CoreWeave June 2026 Form 10-Q; CoreWeave financing announcement; CoreWeave DDTL 4.0 overview.

The favorable path: equipment is delivered and accepted; service remains available; Meta pays for reserved capacity even if its own use disappoints; collections cover costs and scheduled principal; debt declines before liquidation is necessary. Meta could earn a poor return on the capacity it bought while the project’s lenders are repaid.

The confidential Meta service agreement and underwriting model were not independently inspected. The public credit agreement redacts important customer and schedule details. A rating agency’s access to private information supports an attributed credit opinion; it does not supply this investigation with current invoices, accepted-unit counts or compliance certificates.

Installation financing is described at 90% of cost, with up to approximately 102% after stabilization and cash-flow sizing. Returning initial parent capital can leave less original equity beneath later losses. It is therefore wrong to assume a permanent installation-equity cushion. Loan-to-cost, book-asset coverage and net realization after enforcement are different ratios.

Sources: CoreWeave DDTL 4.0 overview; DDTL 4.0 executed credit agreement.

A covenant exclusion is not proof of an expected balloon

The 1.15× maintenance test starts after commitment termination, with specified transition rules, and excludes the initial special amortization payment and final bullet from its debt-service definition. This does not mean a material final balloon is expected. The issuer and DBRS describe full amortization. Both facts can hold: the schedule is designed to repay principal, but the ongoing covenant alone does not prove that design is succeeding.

The exact monthly installment percentages and actual commitment-termination date are not established. The 30 June 2027 outside availability date does not establish that commitments end only then. Initial, delay-related and top-up adjustments matter. The calculations therefore leave scheduled principal visible as a variable rather than imposing straight-line repayment.

Sources: DDTL 4.0 executed credit agreement; Morningstar DBRS initial credit analysis; CoreWeave DDTL 4.0 overview.

The physical requirements behind the favorable payment path

The 22 September capacity review emphasizes the operating terms behind the same favorable DDTL 4.0 proposition: hot and cold spares, Dell warranty coverage followed by a replacement strategy, and power hedging at two North Dakota facilities rather than the entire cohort. These protections can preserve service and contracted payments, but spares, maintenance and later replacements still consume resources. The rating’s uptime history and projected coverage do not become current workload telemetry, actual compliance or an owner-return estimate. Physical operating detail and the attributed rating source.

What collections would cover the obligation?

Use a matched three-month period. Let C be net deposited customer cash eligible for the test; K the corresponding prior-ranking cash costs under the contract’s expense treatment; and I, A, H cash interest, included scheduled principal and net included hedge payments. Excluding new cure equity and other eligible receipts:

Required customer collections = K + 1.15 × (I + A + H)

The ordinary payment floor is K + I + A + H. Reserve replenishment and excluded payments need separate liquidity tests. Neither formula determines the invoice legally due from Meta: the service agreement controls that. Contractual net operating income starts from specified account deposits and has adjustments, including cure-equity treatment; it must not be relabeled entirely independent customer earnings.

Sources: DDTL 4.0 executed credit agreement.

The 5–7% rates are assumptions, not actual weighted cash coupons or the filing’s accounting effective rate. Approximately 5.9% initial fixed-tranche pricing and the floating SOFR spread supply context only. Actual amortizing balances, daily accruals and hedge settlements can differ. The missing A and H are not assumed to be actual zero.

Sources: CoreWeave financing announcement; CoreWeave June 2026 Form 10-Q.

In the illustration, an extra $10m of prior costs raises both thresholds by $10m; an extra $10m of included principal raises the coverage threshold by $11.5m. These are reverse tests for actual records, not an assertion that invoices, service costs or principal currently equal the example.

The projected cushion is not measured headroom

If a period actually starts at DBRS’s 1.20× base-case ratio, if its basis matches the contractual test, and if debt service is unchanged, a 4.17% decline in net operating income reaches 1.15×; a 16.67% decline reaches 1.00×. These are conditional identities around an underwriting assumption. They are not today’s cushion, permissible revenue haircuts or outage estimates. Fixed costs can magnify a revenue change into a larger NOI change; a contractual deduction may not begin at the first minute of an outage.

Sources: Morningstar DBRS initial credit analysis; DDTL 4.0 executed credit agreement.

Reserves can tighten before a payment is missed

June disclosure reports $155m of current collateral, primarily restricted cash—not a verified $155m available reserve. Under the favorable cash-only assumption that all of it could support the required reserve, the 6% interest sensitivity leaves at most $112.445m for relevant operating costs, installments and hedges. The bound is $119.538m at 5%, or $105.353m at 7%. Additional drawings or less usable cash tighten it; eligible letters of credit may help.

This is not a reserve-breach finding. Asset composition, account allocation, actual installments and eligible letters of credit are unknown. Before commitment termination the reserve uses the next three months; afterwards separate category maxima across remaining three-month windows can require more than the largest combined quarter. A high-cost quarter and a different high-principal quarter can both affect the required reserve.

The companion’s synthetic check illustrates that last distinction: candidate windows with interest 10/8, principal 20/30, hedges 1/2 and operating costs 40/35 total 71 and 75 units. Separate category maxima total 82. These are invented units for the contractual method, not borrower balances.

Sources: CoreWeave June 2026 Form 10-Q; DDTL 4.0 executed credit agreement.

A cure is financing support, not stronger customer receipts

The parent guaranty covers specified conduct-related losses, including fraud, misappropriation and certain voluntary or consented bankruptcy actions. It is not a general guarantee of ordinary principal and interest, nor permission for a costless voluntary walkaway.

Different cure clocks apply. The defined Cash Shortfall Event concerns §2.08(a) payments and has a 30-day deadline after the relevant Monthly Payment Date. The DSCR cure period is 15 days after required financial statements; the separate liquidity cure allows 15 business days after non-compliance. After the three-month post-termination sunset, §7.03 cures are limited to three fiscal months in twelve. During the specified windows, acceleration for those breaches is restrained, but lenders need not keep advancing funds.

Blocked distributions or new sponsor capital can be the first economic burden. A successful cure may preserve a working service and prevent destructive liquidation. It does not prove customers paid more, and its cash is not counted again as independent demand.

Sources: DDTL 4.0 executed parent guaranty; DDTL 4.0 executed credit agreement.

If contracted cash fails, what could collateral repay?

Liquidation is the backup stress, not DDTL 4.0’s expected repayment route. This test uses actual June principal of $2,837m and approximately $3,300m of noncurrent book collateral, primarily net equipment. It does not use the separate $14.117bn hypothetical investment cohort. Later drawings, repayments, depreciation and reserve consumption change both sides.

Let q be gross noncurrent realization as a fraction of book, x realization costs, and c recovered current cash. With other claims initially set to zero:

Resources for principal = $3,300m × q × (1 − x) + c

At 75% realization, recovering all $155m as cash leaves approximately $330.8m of principal unpaid; recovering none leaves $485.8m. Full recovery requires about 85.6% or 90.5% of book respectively. The rounded book amount and hypothetical inputs do not justify interpreting these as observed resale prices or expected losses.

Claim priority makes the displayed case optimistic. It assumes no prior accrued interest, enforcement fees or applicable premiums, and no secured-swap claims sharing the principal tier. If prior claims are H and pari-passu swap claims T, the full-recovery fraction becomes (D + H + T − c) / [B × (1 − x)]. Adding an illustrative $100m in each category raises the 75%-realization/$155m-cash principal shortfall to about $512.7m. The code allocates the remaining pool proportionally between principal and the competing swap claim rather than subtracting all swap loss from principal.

Sources: DDTL 4.0 executed credit agreement; CoreWeave June 2026 Form 10-Q.

Transfer costs, downtime, customer assignment rights, power and lease access, and an operating platform determine what collateral can realize and how quickly. A functioning contracted service can be worth more than disassembled equipment at auction. A technically usable GPU can still realize too little or too late. Those values remain measurements to obtain, not facts supplied by this haircut table.

A continuing asset is not the same as an appraised enforcement recovery

The capacity investigation does not supply a market bid or empirical resale curve for this collateral. It adds a different costed case—IREN’s Microsoft deployment—to solve how much net sale or continuation value is needed after a first contract. Its $2.81bn/$3.42bn end-value requirements belong to a normalized all-capital return model, not to DDTL 4.0’s $2.837bn principal claim or the enforcement waterfall above.

A usable powered building may retain value after GPU replacement, but transfer rights, cooling, service continuity, maintenance and the cost of the next equipment generation matter. A principal-recovery percentage, a building’s gross rent and complete investment NPV cannot substitute for one another. Facility and integrated-compute continuation; cash and ownership boundaries.

NVIDIA’s support has gates, remedies and a different time horizon

NVIDIA’s August disclosure describes support for 4.25 GW at SB Energy’s PORTS-Pike project, with a separate 3.8 GW option at NVIDIA’s discretion. The $105bn main ceiling is phase-dependent, with lease commencements expected to begin in 2028. It is a residual-value arrangement following defined failure—not a present loan, an unconditional payment of every missed rent invoice or money already received by SB Energy.

The nine-phase description is not an activated, building-by-building callable schedule. The project forms name NVIDIA as guarantor but leave individual landlord and tenant identifiers partly blank or redacted. The commercial OpenAI/SB Energy relationship is identified in NVIDIA’s disclosure. Energy Global LP’s prepaid forward and SE Global Holdings’ share subscription are separate instruments and parties.

Sources: NVIDIA 17 August current report; NVIDIA July-quarter Form 10-Q; Filed form of residual-value guaranty.

Delivery, claim and payment are separate steps

The filed lease permits specified rent abatements and service/power credits. Lawful relief for failed delivery is not automatically wrongful tenant nonpayment. The principal Covered Loss Amount is gated by ready-for-service conditions; expense provisions and release rules have their own conditions. The forms do not support describing the arrangement as an unconditional construction-completion guarantee.

Reserves and security deposits enter the sequence. NVIDIA can choose among assumption, re-letting, sale, rejection and specified deferral alternatives, including deferral of up to one year while paying specified costs. Certain enforcement conditions require concurrent pursuit of insurance and the tenant-parent guarantee. Reimbursement is directed to the tenant or tenant parent; recovery cannot simply be retained as duplicate compensation.

The main loss formula, discount rate, phase values, important cost definitions and several notice periods are redacted. Some cost provisions sit beside or affect the main cap. The $105bn headline therefore cannot be reconstructed into an all-in same-day cash call, and dividing it evenly across nine phases would invent exposures.

Sources: Filed form of PORTS-Pike lease; Filed form of residual-value guaranty; SB Energy exhibit-only amendment.

A guarantee can ultimately pay yet leave interim liquidity needs with project reserves, equity or separate financing. Conversely, preserving a working asset may be much better than abandonment. Neither outcome is proved by the cap alone. The public forms do not support a reliable dollar loss scenario for the actual project.

NVIDIA has meaningful resources: it reported $56.586bn of cash, equivalents and marketable debt securities at 26 July and $74.421bn of first-half operating cash flow, alongside other substantial cash uses. Those are not funds escrowed for PORTS-Pike, and one half-year’s cash flow is not guaranteed from 2028 onward. No ratio of the cap to current cash is presented as a solvency test.

Sources: NVIDIA July-quarter Form 10-Q.

The prepaid-forward correction remains controlling

NVIDIA entered into a $1.5bn prepaid-forward contract with Energy Global, LP. The agreement requires payment within three business days after its date. The documents inspected for this investigation do not verify receipt.

This is the correction already established in the demand investigation, not a new discovery of delinquency. The 20 September inspection does not reverse it and adds no $1.5bn cash receipt. The separate $1.5bn share purchase remains conditional on a qualifying IPO and subsequent closing; its primary-proceeds threshold excludes sales by existing holders. Contract title, parties and section identify the instruments because the later exhibit index and filenames number them differently.

Sources: Energy Global LP / NVIDIA prepaid-forward contract; SE Global Holdings / NVIDIA share-purchase agreement; SB Energy exhibit-only amendment.

Read the exact correction and its original consequence.

Meta’s resilience is real, but not unlimited economic insurance

Meta’s unsecured parent notes draw repayment support from its broad enterprise rather than the resale of one project’s equipment. The May issue raised $25bn face across six tranches, coupons 4.55–6.45%, and maturities from 2031 to 2066. June note principal was $84bn, with none due in the remaining 2026 bucket and $2.75bn due in 2027. “No financial covenants” does not mean no contractual obligations.

Long-dated funding and operating resources allow a poor incremental project return to reduce shareholder value or cash reserves before it threatens bond payment. They also help explain the stronger credit proposition of a Meta-backed capacity contract. Paying that contract is not a guarantee of CoreWeave’s other debts.

Sources: Meta June 2026 Form 10-Q.

A bounded second-half cash sensitivity

The starting general-use pool is approximately $90.260bn of June cash and marketable securities. H1 OCF was $64.088bn; cash PPE of $49.113bn plus $1.805bn lease principal gives $50.918bn of capex under management’s definition. The annual $130–145bn capex guidance uses that same lease-inclusive definition.

Assume H2 OCF is flat versus H1 or 30% lower. Retain the annual capex range. Repeat H1 RSU-tax settlement and dividend cash uses totaling $11.403bn. Add no new borrowing, equity or asset sales. These second-half choices are assumptions, not management guidance.

Residual = June liquidity + assumed H2 OCF − (annual capex − H1 capex) − repeated RSU-tax/dividend uses

Even the selected adverse combination leaves $29.637bn after the specified uses; it does not mechanically exhaust the starting pool. That is a serious countercase to equating a large investment plan with immediate funding failure. Every scenario nevertheless consumes liquidity, so none shows spending entirely self-financing.

Operating interest and taxes already affect OCF; lease principal already sits in capex; May borrowing already sits in June liquidity. The separate approximately $10.8bn restricted escrow is outside the starting general-use pool—not subtracted twice. There were no H1 share repurchases, so there is no H1 buyback cash outflow to “cut” again. Other financing uses, acquisitions, new restrictions, market movements and intermediate payment timing are not fully modeled.

Sources: Meta second-quarter results; Meta June 2026 Form 10-Q.

Fixed commitments limit the option to slow spending

At June, Meta disclosed $349.31bn of non-cancelable commitments, including cloud and infrastructure, with approximately $53.52bn due in 2026 and $81.65bn in 2027. It also reported $278.99bn of uncommenced leases and an additional $68bn entered in July. Their timing and categories differ; they are not one immediately payable debt pile.

Assuming all future spending is cancellable understates pressure. Adding every commitment again to a cash model whose operating or capital spending already includes those purchases overstates it. The detailed mapping is unavailable. In a moderate slowdown Meta may keep paying suppliers while accepting lower returns; persistent deterioration can consume liquidity and constrain future choices.

Sources: Meta June 2026 Form 10-Q.

Where could the shortfall travel?

Contracted cash catches up. Acceptance, availability and customer payment support amortization; reserves suffice; supplier support remains unused or is validly released. This path has a documented repayment proposition behind it, not only hopes about future adoption.

Useful technology, disappointing equity. The buyer pays for reserved capacity despite weak utilization, the operator pays debt but retains little, and shareholders face lower distributions or dilution. A strong project rating does not contradict that outcome.

Performance or timing failure becomes a funding interruption. Reduced or delayed collections coexist with costs, installments and reserve needs. Distributions stop or a cure is funded; if protection fails, restructuring or realization exposes creditors to claim priority and net asset value. These paths can coexist across different firms.

Do not impose one contract’s resilience on the whole group

CoreWeave’s DDTL 5.5 announcement describes approximately five-year financing against initial contracts averaging about three years, with a framework for replacement or renewal. That leaves a re-contracting task different from DDTL 4.0’s expected full amortization. The September issuer statement about higher annualized capacity prices concerns short contracts of three to six months, not a fleet-wide multiyear cash-margin forecast. Upstream scarcity can coexist with cheaper retail model use.

Sources: CoreWeave DDTL 5.5 announcement; CoreWeave capacity-price statement; CoreWeave $3.7bn note-pricing release.

Origination is not ownership

The DDTL 4.0 announcement names Blackstone Credit & Insurance as an anchor and identifies bank and institutional participation. Agent and arranger roles do not establish who retained the loan at 20 September. No complete current beneficial-holder schedule was found. A business-unit name does not place all the exposure on a corporate balance sheet, a particular insurance account or a retail redemption fund.

Sources: CoreWeave financing announcement; DDTL 4.0 executed credit agreement.

The FSB’s May private-credit report describes both stabilizing long-term fund structures and potential amplification through leverage, redemption-enabled vehicles, opaque valuations and links to banks, insurers and pensions. Subscription loans against investor commitments differ from portfolio/NAV loans against assets. These are general mechanisms, not evidence of this loan’s holdings.

A wider shock needs demonstrated links: a loan is impaired; an exposed holder owns enough of it; holder capital or liquidity cannot absorb the loss; and financing or redemption terms force further cash demands or sales. A well-capitalized closed-end holder could instead accept a mark-down and negotiate. Missing holder evidence prevents selecting a systemic outcome or assigning a crisis probability.

Sources: FSB Report on Vulnerabilities in Private Credit.

What would strengthen—or weaken—the financing case?

The most useful next evidence replaces the variables that decide payment and recovery: actual customer collections, borrower costs and installments, reserve support, transferable service value and the funding of the institutions bearing the exposure. A larger financing headline cannot substitute for these records.

What this changes in the wider assessment

A stronger containment case: the identified Meta take-or-pay contract and DBRS full-amortization expectation sharpen the favorable case beyond a generic GPU-backed loan. The covenant’s final-payment exclusion remains an inference limit, not proof of an expected balloon.

A new quantitative level: the cash bridge, collection/reserve formulas and collateral test are anchored to a real obligation. They do not replace the separate hypothetical asset-return model with an empirical borrower margin, and an investment NPV shortfall is not a lender-loss estimate.

A changed financing stage: CoreWeave’s September proposal became an upsized priced offering with future settlement. That is a status update, not new cash received. The NVIDIA forms deepen loss allocation but do not establish prepaid-forward receipt or reverse the already-stated correction.

Still unresolved: net project collections, actual covenant cushion, unredacted cost/amortization schedules, a full post-June group roll-forward, net collateral realizations, beneficial loan holdings and the SB Energy loss formula. The conclusion is stronger containment for some contracts, continued funding needs for expansion and incompletely mapped correlated exposures—not blanket safety or inevitable collapse.

The capacity review retains the NVIDIA obligation/receipt correction. It also does not establish that the CoreWeave notes settled on the previously expected September 22 date. The September 20 transaction-status statement remains dated; a separate closing record and subsequent cash movements would be needed to update receipt or the group balance.

Source boundaries and reproducible calculations

Sources for the financing investigation were inspected on 20 September 2026. The report distinguishes filings, executed agreements, filed forms, issuer descriptions and rating expectations. The financing investigation inspected original PDF tables and figures. It did not independently audit confidential accounts.

The public credit agreement redacts decisive schedules. The SB Energy registration-statement body exceeded the retrieval limit, although its index and exhibits were accessible. The DBRS argument is attributed; its private model was not replicated. A Moody’s page supplied no readable underwriting argument. Current beneficial holdings, actual letter-of-credit support and phase-specific guarantee activation were not established. Missing evidence is not evidence of default or of safety.

Download the source-linked financing inputs (CSV) · Download the calculations (Python). Run python financing-calculations.py financing-data.csv with Python 3.10 or later. The reader ledger retains the observed values and clearly separates assumptions. The code reproduces the cash bridge, thresholds, recovery priorities and Meta sensitivity; it is not a live covenant calculator.

The earlier whole-system inputs and demand observations remain separate, along with their calculations. No guarantee cap or unverified forward receipt is added to their cash measures.