Who funds the next purchase?

The answer

The evidence supports real, repeatedly funded investment demand more strongly than a story of fictitious sales. It supports the next expansion less strongly than it supports payment for the current build. NVIDIA has a large operating business, substantial receivables and disclosed extended customer terms. Meta has operating cash from a predominantly advertising-funded franchise. CoreWeave has customer contracts, customer cash and completed financing, but its equipment purchases cannot be described as funded entirely by cash retained from ongoing service. These are different funding mechanisms, not a division between “real” and “fake” customers.¹²³⁴⁵⁶

The new investigation changes the diagnosis in three concrete ways.

First, supplier growth is accompanied by more credit extended to customers. Across NVIDIA's selected receivable, deferred-revenue and customer-program accounts, its selected net customer-account position rose from $30.576 billion to $49.256 billion during the first half of fiscal 2027. That $18.680 billion increase remains after crediting the growth in deferred revenue and customer-program liabilities. NVIDIA also explicitly describes extended payment terms for certain investment-grade customers. This is not proof of arrears: a receivable within an agreed extended term can be performing. It does mean that some recognized growth arrives before cash and must be financed somewhere.²

Second, a strong operating payer can keep buying while the cash left over narrows. Meta's reconstructed second quarter generated $31.862 billion of operating cash, against $31.078 billion of cash capital spending including finance-lease principal. The resulting $784 million is positive, but much smaller than the first quarter's $12.386 billion. After the quarter's actual equity-award tax settlements and dividends, the same boundary is negative $4.850 billion. This does not establish a payment failure or a borrowing requirement of exactly that amount: accumulated assets, financing and other uses also matter. It establishes that repeating this particular purchase-and-distribution programme is not effortless self-funding from an unchanged quarter's operating cash.³⁴

Third, CoreWeave's lower second-quarter operating cash is not explained by a collapse in its pre-working-capital subtotal. Operating cash fell from $2.984 billion to $679 million, while the subtotal before the five disclosed working-capital movements rose by $244 million. A $2.549 billion adverse change in those working-capital contributions more than explains the cash decline. Yet the second quarter still paid $6.422 billion for property and equipment and $3.884 billion of debt principal. Cash remaining after those uses was negative $9.627 billion, before other investing uses. Completed financing, not a fiction about customer cash, closed the historical account.⁵⁶

There is favorable evidence of an actual repeat decision. In March 2026, Meta entered a new CoreWeave order and exercised an option under an existing order; the disclosed initial commitment of approximately $21 billion includes both. That is more informative than an automatically extending contract or a rising share price. It is still a conditional, multi-year service commitment—not a $21 billion current payment, a pure new-order increment, or a payment certificate for NVIDIA.⁷

The supported vulnerability is therefore the funding and economic attractiveness of the next purchase, particularly when customer concentration, extended credit and capital-market access interact. Existing creditors can be paid while new equipment orders slow. A customer can rationally buy again without proving that every earlier asset, provider or ownership price earned an adequate return. The stronger favorable interpretation is that productive service, operating franchises and suitably structured capital are financing genuine expansion. The stronger adverse interpretation is that maintaining the previous growth rate requires more credit, smaller net prices or more new risk-bearing than later cash returns justify. This evidence does not establish how prevalent the adverse combination is across AI assets.

What is being followed—and what is not being added together

2.1 The supplier and the two buyer funding models

The organizing supplier is NVIDIA. Meta is the operating-payer contrast because its own engineering record documents NVIDIA deployments, its accounts expose the cash-resource boundary, and the filed CoreWeave relationship contains an identifiable further purchase decision. CoreWeave is both a purchaser of infrastructure and a seller of cloud service. Its disclosure says current customer contracts specify NVIDIA GPUs; that establishes a technology-procurement link, not the NVIDIA content of every dollar of capital expenditure.⁸⁹⁷⁶

The economic chain can therefore include advertisers → Meta → a service provider such as CoreWeave → equipment vendors, as well as Meta's own equipment purchases. CoreWeave can also sell to other customers, and NVIDIA can supply through original-equipment manufacturers or other direct purchasers. These legs must not be added to estimate independent end demand. The same underlying customer resource may support several reported revenues at different production stages.

This is a bounded named-chain alternative to a complete allocation of NVIDIA's end customers. Neither NVIDIA's anonymous large customers nor CoreWeave's lettered concentration accounts are assigned guessed identities. Meta's inclusion does not make it “Customer A” at either issuer. Nor does the presence of a named commercial agreement identify the payer behind an anonymous balance with a similar percentage.

2.2 Five numbers that answer different questions

A cash account below uses reported cash flows. An account reconciliation shows how disclosed balances and revenue fit together without inventing unobserved transactions. A repeat-purchase requirement asks which additional resources or changed uses would make a specified future programme balance. These labels remain beside the results, not only in a remote disclaimer.

2.3 Dates and the research advance

NVIDIA's fiscal-year sequence is followed from the January 2023 opening stocks through January 2024, January 2025, January 2026 and July 2026. Full years are not compared with the last half-year as if their durations were equal. The latest filing's explicitly recast prior-half market-platform figures supply the matched growth comparison. Its headquarters-based geographic attribution is not treated as the location of ultimate use.¹⁰¹¹¹²

For the two calendar-year buyers, second-quarter flows equal first-half flows less the matching first-quarter flows. Balance-sheet stocks are read at their own dates. This supplies a new historical decomposition; it does not claim that a new quarter appeared since the earlier payer investigation. The earlier half-year resource accounts remain useful and dated. The earlier customer-capture study already reported Meta's Q2 cash-after-capital amount of $784 million and its year-on-year comparison; that amount is not a new discovery here. The financing study also already reconciled CoreWeave's half-year cash after capital and principal. The advance here is the Q1/Q2 decomposition, the compensation/distribution and repeat-use boundary, the customer-liability classification bridge, and their connection to supplier credit, an actual further purchase decision and operative terms.¹²¹³

The report does not refresh equity prices or rerun the preceding ownership models. It tests an economic condition inside them: whether real purchasers and payment structures can support continuing supplier growth. The full sector-wide end-buyer allocation, invoice ledger and future returns are not observed.

NVIDIA: growth, customer credit and the limits of a collection bridge

3.1 The growth is not one interchangeable customer category

In the latest matched first-half presentation, revenue rose from $90.805 billion to $177.837 billion. Data Center revenue rose from $80.208 billion to $164.269 billion. Within the recast market-platform presentation, Hyperscale contributed $45.333 billion of the increase, AI Clouds, Industrial & Enterprise $38.728 billion, and Edge Computing $2.971 billion. These three changes sum to the $87.032 billion total increase.²

The decomposition is useful because it prevents the growth being attributed wholly to a single kind of purchaser. It is not an estimate of independently funded demand. NVIDIA changed the market-platform presentation in fiscal 2027 and subsequently moved a company from one category to another, recasting the comparison. A category can include different funding models; a classification change is not a new purchase. The anonymous customer's identity and the exact operating-versus-financing source of each category remain unallocated.

The annual Data Center totals in the companion preserve their original fiscal dates. They are not spliced into a fixed “Hyperscale customer cohort,” and Compute & Networking segment revenue is not substituted for Data Center market-platform revenue. Those different definitions can overlap without being identical.

3.2 A multi-period customer-account reconciliation

Define a selected net customer position:

N = net accounts receivable − total deferred revenue − accrued customer programs.

This is an analytical grouping of disclosed accounts, not a balance-sheet line, net loan portfolio, or valuation of customer credit. Its change shows how much more of the selected recognized/customer-account position lies ahead of an offsetting liability. Subtracting the increase from revenue gives the corresponding bridge:

B = revenue + opening N − closing N.

The formula deliberately retains customer-program liabilities rather than assuming that all net sales can be treated as gross invoice cash. It also retains both current and noncurrent deferred revenue. The historical results are:

Sources: reported balance-sheet stocks, deferred-revenue notes and revenue; calculations in the companion.¹⁰¹¹¹²

The material observation is the larger selected net customer-account position, not a claim that $159.157 billion was banked. The latest half-year increase of $18.680 billion exceeds the previous full year's $14.204 billion despite the shorter interval. That comparison describes amounts and durations; it is not a growth-adjusted default statistic. Higher sales naturally require more working capital, and extending sound customers' terms can be commercially sensible.

The ending position moves from $2.059 billion at January 2023 to $6.581 billion, $16.372 billion, $30.576 billion and $49.256 billion at the successive dates. Offsetting liabilities are real claims on the business or on future service, not costless financing indefinitely available for another purpose. Conversely, the existence of the assets is not itself an observed loss.

3.3 What would be needed to call the bridge collections?

A complete collection account would require the movements that can affect these balances without being the corresponding cash receipt: acquisitions, currency translation, bad-debt allowances and write-offs, returns, contract assets, noncash consideration, rebate settlement and classification changes. Sales taxes or customer cash passed to another party could also affect a gross-receipt interpretation. The public tables do not supply a complete matched rollforward for all these terms.

The narrower receivable check is encouraging but insufficient. Net AR increased by $24.593 billion while the cash-flow statement reports a $24.590 billion receivable use. The $3 million residual is small. The analogous annual residuals are $0 million, $3 million and $2 million. These checks reduce one reconciliation uncertainty; they do not prove that the rest of the customer account has no noncash adjustment. The bridge therefore has no asserted upper or lower bound on actual receipts until the missing adjustment signs and quantities are established.

Write the complete but unidentified join as:

actual comparable net customer cash = B + J,

where J is the signed net of the missing adjustments under one consistent receipt definition. It is not set to zero. If an acquisition adds a receivable without current-period sales, for example, mechanically subtracting that receivable increase understates collections unless the acquisition effect is restored. A write-off can create the opposite distortion. A cash credit and a noncash equity incentive need different treatments. Numerical precision cannot supply that missing classification.

Actual cash evidence remains stronger than a bare revenue number. NVIDIA reports operating cash of $28.090 billion, $64.089 billion and $102.718 billion across the three full fiscal years, and $74.421 billion in the latest half. The early build-out record also identifies $557 million received ahead of invoice due dates in FY2024. Neither measure equals gross independent customer receipts, but both are inconsistent with treating all the reported activity as unpaid paper.¹⁰¹¹¹²

3.4 Advances are funding once, not revenue cash twice

The latest deferred-revenue table exactly reconciles $2.572 billion opening + $17.709 billion additions − $13.869 billion recognized = $6.412 billion closing. The additions include a reported, rounded $15.6 billion of customer advances, while recognition includes $13.0 billion associated with advances. The ending current-liability note identifies approximately $2.8 billion of advances, against $160 million at the preceding year-end.²

This supports a genuine favorable mechanism: customers can help finance production or secure access before revenue is recognized. It does not justify adding $15.6 billion to reported operating cash, nor adding the later $13.0 billion again as another receipt. The rounded advance subtotals also should not be forced into an exact rollforward by inserting an invented fee or refund.

The economic cost of an advance depends on what the supplier owes in exchange: equipment, support, development, credits, availability or other performance. An advance is not all profit at receipt. It can fund productive work and still create a future interval in which revenue is recognized with little incremental customer cash.

3.5 Extended terms change the financing burden without proving bad credit

NVIDIA says it normally receives payment shortly after delivery, but has extended terms of 90 days to one year for certain investment-grade customers. Its discussion links working-capital use to extended terms on large multi-quarter agreements.² No disclosed invoice-aging schedule in this investigation establishes that the July receivables were overdue. “Receivables rose” is therefore not a sufficient noncollection finding.

The stronger adverse hypothesis is narrower: if new sales require longer terms faster than customer advances or other funding grow, the supplier finances more of the buyers' purchasing interval. Credit quality can remain sound while cash conversion slows. A strong customer may eventually pay every invoice and nevertheless reduce its next order if the equipment's returns disappoint. The financing risk can thus reach new supplier orders before any existing invoice defaults.

The decisive next observation is not another debt-to-revenue slogan. It is the realization of the extended-term receivables, their aging and adjustments, together with the volume and net terms of subsequent purchases by a consistently defined group. Good collections followed by rational repeat orders would strengthen the durable-growth case. Good collections followed by smaller orders would support the current-sale case while weakening the extrapolation.

Concentration: which obligation would weaken first?

4.1 The denominator must travel with the percentage

The supplier's concentration is material, but the public records do not identify a complete end-customer map. NVIDIA distinguishes direct purchasers—including distributors, original-equipment manufacturers, system builders and cloud providers—from the organizations ultimately using the equipment. A distributor's invoice can aggregate several end buyers, while one end buyer can procure through several intermediaries. The anonymous disclosures cannot be joined to Meta, Microsoft or CoreWeave merely because a percentage looks similar.¹⁰¹¹¹²

These sums describe the disclosed large-customer buckets. They are not a stable five-customer cohort, a complete concentration index or five independent risks. More customers crossing a disclosure threshold can change the sum without an identical change in economic diversification. Nor does 70% of period-end receivables mean 70% of current revenue is unpaid or financed by speculative customers. The stock and flow have different denominators and dates.

The more defensible finding is that a large part of the supplier's collection exposure sits in a relatively small disclosed set, while the financing and end-use allocation of that set remains incomplete. That increases the importance of customer-specific terms and balance sheets. It does not establish that the identified balances are impaired. NVIDIA's extended investment-grade terms are a concrete reason why concentrated receivables can rise without delinquency.²

4.2 CoreWeave's revenue concentration and receivable concentration are different again

CoreWeave's June-quarter disclosure assigns 36%, 26% and 10% of revenue to three anonymous customers: 72% in total. Its first-half disclosure assigns 40% and 23% to two customers. At June 30, two customers account for 32% each of accounts receivable. In the preceding December snapshot, the disclosed large receivable shares were 68% and 11%. These are the filing's own categories; they do not identify Meta as any particular letter or provide a stable independent-demand cohort. The filing explicitly cautions that lettered customers may differ from those represented by the letters in earlier reporting periods.⁶

The annual and interim records also distinguish committed contracts, remaining performance obligations, revenue backlog and current recognized revenue. The June report describes approximately 98% of revenue as related to customer commitments, including capacity delivered before a commitment’s start date, and reports $103.7 billion of remaining performance obligations. The separately communicated $104.2 billion revenue-backlog measure should not silently replace that accounting measure. Neither is an available cash balance or a margin after hardware, power, financing and customer credits.⁶¹⁴

The favorable interpretation is important: a large customer can commit to a meaningful service for a long period, allowing construction and financing at a scale that fragmented spot demand would not support. Concentration can reflect the cost and scale of AI infrastructure rather than an invented market. The adverse interpretation concerns dependence on a few purchasers' future budgets, completion decisions and contract economics. These interpretations are not equally resolved by the available disclosure: current committed business is well evidenced; the complete resources behind every next purchase are not.

4.3 A slowdown has several possible first effects

For a strong operating customer, disappointing returns can first reduce new equipment orders or optional capacity, while accepted invoices remain paid. For a financed cloud provider, a delay can first reduce drawable financing or available supplier credit, before an existing customer defaults. For a customer receiving service credits, the first effect can be lower net future consideration, without a breach. For a distributor, a change in end-customer mix can affect inventory and purchase timing without identifying the final buyer's investment return.

Those are distinct mechanisms. They cannot be added as separate losses on the same underlying capacity. In particular, a reduction in CoreWeave's next hardware purchase and NVIDIA's corresponding lower sale are two sides of one transaction, not two independent demand shortfalls. The same discipline applies to Meta's cloud payment and CoreWeave's receipt. The remainder of the investigation follows their separate cash accounts without summing them into an industry total.

Meta: a real repeat buyer with less current cash left after investment

5.1 The purchasing connection is documented, not inferred from a logo

Meta's own March 2024 engineering account describes two clusters with 24,576 H100 GPUs each, used for generative-AI development, alongside the earlier Research SuperCluster with 16,000 A100 GPUs. The article identifies actual infrastructure and its intended work; it is not a procurement invoice or a causal financial-return study. Its separate year-end hardware ambitions remain targets, not observed delivered-and-paid quantities.⁸

NVIDIA's February 17, 2026 announcement then describes a multiyear, multigeneration Meta relationship: production deployment of Grace CPUs and planned Blackwell/Rubin infrastructure, with further potential Vera adoption. This supports a continued commercial and technical relationship across generations. The future quantities and deployment plans are supplier statements, not a verified schedule of October cash payments.⁹

A different purchase is documented in CoreWeave's filing. On March 31, 2026, the companies entered a new order and Meta exercised an option under an existing order. Their approximately $21 billion aggregate initial commitment covers the described service periods through December 20, 2032 and April 10, 2032, respectively, subject to delivery, availability and contractual termination provisions.⁷

The exercised option matters: this is stronger evidence of a new purchasing decision than an automatic extension or an unchanged old backlog. The full $21 billion is nevertheless not all a single wholly new order, all cash received at signing, or a disclosed equipment invoice to NVIDIA. It buys cloud service from CoreWeave; the supplier's subsequent equipment, facility and financing payments remain separate legs.

The comparison therefore does not depend on knowing NVIDIA's anonymous Customer A. There is a documented named purchasing chain and an observable payer account. What remains unavailable is a complete allocation from that chain to NVIDIA's reported direct-customer revenue and collections.

5.2 What the new quarterly match establishes

The cash comparison uses the original March and June filings, with second-quarter cash flows calculated as first-half cumulative flows minus first-quarter flows. The periods and corporate perimeter match. No September-quarter actuals are inferred, and a June stock is not called an October available balance.³⁴

USD millions; cash uses are shown as positive amounts.

This is not simply the already documented first-half budget with a later date. The new match shows where within that half-year the balance changed. Second-quarter 2026 operating cash grew approximately 24.65% year on year, while capital cash grew 82.68%. The amount left after capital spending fell by $7.765 billion against Q2 2025 and by $11.602 billion against Q1 2026. Stronger operating cash and a much smaller immediately retained amount coexist.

The negative last cell does not mean Meta was unable to pay or that the quarter's purchases were economically irrational. It means that the specified capital and distribution uses exceeded that quarter's operating cash by $4.850 billion. Existing liquid assets, actual financing, investment realizations and other cash movements can support the difference. The statement does not identify how much of any individual hardware or cloud invoice was paid from advertising receipts versus a bond issue.

Equity-award withholding is retained because it is an actual cash use outside operating cash. It is not silently treated as an optional open-market buyback. Nor is the accounting stock-compensation expense subtracted again as if it were another current cash payment: this is a funding account, not a fully diluted equity valuation. Future compensation and renewal remain real economic costs when judging the return on continuing investment.

The lease convention is also consistent. Operating lease cash and finance-lease interest already enter operating cash; finance-lease principal is included once in capital cash. Newly obtained noncash lease assets do not become an additional same-quarter cash use. Their future payments remain obligations, so the convention does not imply that leased infrastructure was obtained free of cost.³⁴¹⁵

5.3 The whole cash account prevents a false funding-crisis inference

The derived Q2 account reconciles $31.862 billion operating cash, $49.653 billion investing cash use, $15.967 billion financing cash and a $7 million adverse currency effect to a $1.831 billion reduction in cash including restricted accounts. That cash fell from $31.102 billion at March 31 to $29.271 billion at June 30.³⁴

Financing includes $24.910 billion of net debt proceeds, already reflected in the June account. Investing includes $17.754 billion of net marketable-security purchases, as well as capital and other uses. It would be wrong to label the entire debt receipt an NVIDIA or CoreWeave payment, or to add it again to June resources. Buying a security also reduces cash while retaining another financial asset; it is not the same as consuming that amount in operations.

June nonrestricted cash of $15.462 billion plus marketable securities of $74.798 billion gives a $90.260 billion resource proxy before the remaining claims and market/timing effects. Separately restricted cash totals $13.809 billion and is excluded from that proxy. The reported purchasing-related escrow is subject to conditions and later application; it is not surplus cash for an arbitrary next invoice, and no named NVIDIA/CoreWeave allocation is inferred.⁴

The established payer investigation already showed these first-half resources, restrictions and May borrowing. The new result does not claim them as newly discovered money. It shows that a favorable broad half-year resource position can coexist with a much tighter second-quarter internally funded purchasing margin.¹⁵

5.4 A bounded requirement for repeating the spending rate

A useful reference question is: what would another interval with the observed Q2 capital cash, award withholding and dividends require before other uses? This is a hypothetical repeat of a completed quarter, not a forecast for Q3 or company guidance.

The requirement is:

30,116 + 962 + 4,281 + 1,353 = 36,712 million of operating cash.

Against observed Q2 operating cash of $31,862 million, the difference is $4,850 million, or 15.22% more operating cash. Omitting the dividend reference still leaves $3,497 million to be met through stronger cash generation, lower applicable uses, eligible existing assets or additional net funding. Existing assets can legitimately fund a multiyear investment; this test does not require every quarter to be self-financing.

More generally, in millions:

additional net finance + eligible asset draw + operating-cash improvement ≥ 4,850 + additional capital/award/dividend uses + other net uses.

All changes are relative to the stated Q2 reference. Additional operating costs already incorporated in the operating-cash change are not charged again under other uses. Financing proceeds are net of the costs assigned to that interval; future interest and principal still create claims. A release of restricted purchasing cash is usable only for its permitted purpose and with its corresponding purchase recognized, not as new earnings.

This makes the next-purchase problem concrete without assuming any volume needed to restore NVIDIA's growth. There is a demonstrated operating engine and genuine financial capacity. What is not established is that the same engine can repeatedly fund an expanding capital programme and the existing claims without using assets, changing allocations or raising capital. Nor does the reference establish that every cash-capital dollar is cancellable. Accepted bills, commenced leases and noncancelable contracts can restrict the available response.⁴¹⁵

5.5 Why paying again could still be rational

Meta earns most of its revenue from advertising, not from a separately disclosed sale of AI infrastructure. Its first-half filing reports $114.387 billion advertising revenue within $117.111 billion total revenue. These are broad operating receipts recognized as revenue, not a measured AI contribution or a stand-alone return on this hardware.⁴

The earlier customer-capture investigation supplies a genuine favorable mechanism: improved recommendation and advertising systems can make an existing commercial service more valuable, even when no customer receives a separate AI bill. The provider-reported experiment had a defined population and additional compute requirements; it did not establish a consolidated AI cash margin. The buyer research likewise showed that valuable output, vendor substitution and released cash are different gains.¹³¹⁶

A rational next purchase can defend or expand that franchise, reduce another cost, or provide valuable development capacity. Its incremental benefit must exceed the relevant hardware/service, facilities, power, engineering, distribution, tax and replacement costs over the useful period. An existing profitable franchise deserves credit; assigning it no continuing value because the AI increment is not separately reported would be no more neutral than assigning every advertising dollar to AI.

The narrower adverse case is that the next generation or additional capacity costs more than the cash it adds or protects, even while Meta pays every contracted bill. That would first pressure future orders, capital allocation or ownership returns. The present evidence cannot isolate the full causal return on the $21 billion service commitment or the next NVIDIA generation. It can establish both a real repeat decision and the resources that must support it.

CoreWeave: operating cash, advances and new capital perform different jobs

6.1 A falling cash-flow line is not a measured fall in underlying service demand

CoreWeave's original March and June statements permit the same cumulative-minus-quarter comparison. Revenue rose from $2.078 billion in Q1 to $2.575 billion in Q2 2026, while operating cash fell from $2.984 billion to $679 million. It would be tempting, but incorrect, to treat the entire cash-flow decline as lost collections or collapsing unit profitability.⁵⁶

The matched reconciliation explains the difference. All amounts are USD millions; cash-flow adjustment signs follow the filings.

The $2.305 billion fall is therefore the combination of a $244 million improvement before those five movements and a $2.549 billion adverse change in those movements. Accounts receivable went from a cash-flow benefit to a use; the payable/accrual line changed direction too. The positive deferred-revenue adjustment increased rather than disappeared.

The first row is not normalized operating cash, EBITDA or cash profit. It retains the filing's net loss and noncash adjustments, and excludes timing movements that are part of actual funding. Depreciation, compensation, equipment renewal and the future service owed against advances cannot be treated as economically free merely because some are noncash in the current reconciliation. The result establishes why the quarterly operating-cash change does not by itself diagnose the change in demand or sustainable margin.

This supplies consequential new historical evidence: the quarter with lower reported operating cash still had higher recognized revenue and a positive deferred-revenue adjustment. That contradicts the simple claim that the entire cash decline demonstrates disappearing advance support or a matching collapse in sales. It does not prove that all receivables were collected, all customers remain strong, or the pre-working-capital amount will persist.

6.2 A classification change can look like disappearing customer funding

The commercial model helps explain why timing matters. The June filing describes advance payments on its active contracts and reports a weighted-average advance of 15%–25% of total contract value at the December 2023, 2024 and 2025 reference dates. This is an attributed historical contract-set description, not the advance percentage on every new order or a disclosed cash payment against Meta’s $21 billion commitment. A change in the timing or size of advances can alter funding while the underlying service order remains real.⁶

The deferred-revenue stock illustrates the danger more sharply. It fell from $8.185 billion at December 31, 2025 to $7.523 billion at March 31, 2026—a $662 million reduction—while the first-quarter cash-flow statement shows a positive $575 million deferred-revenue adjustment.¹⁴⁵

The March note reports approximately $1.3 billion reclassified from deferred revenue to customer liabilities following a contractual change. The simple stock-versus-cash-flow difference is $1.237 billion. Adding back the rounded $1.3 billion leaves a $63 million residual. This is a useful reconciliation of direction and scale, not an exact invoice-level cash bridge: the reclassification is rounded and recognition, timing and other adjustments still require allocation.

The broader stock check reaches the same limit. Disclosed customer liabilities rose from $137 million to $1.469 billion. Deferred revenue plus that liability therefore rose by $670 million, rather than falling by $662 million. The combined increase differs from the cash-flow deferred-revenue line by $95 million. It would be wrong to call that residual new cash, a refund or missing money without the relevant accounting entries.

By June, deferred revenue had risen to $9.692 billion. The filing's much smaller aggregate “other current liabilities” does not disclose the exact continuing customer-liability component at the same resolution as March. This investigation therefore does not infer a $1.469 billion cash refund, complete reversal or zero June liability. The separate $1.469 billion OEM-related balance reclassified into debt is a different disclosed item; the identical number does not establish that the two reclassifications are the same transaction.⁶

Two further checks retain the uncertainty rather than conceal it. Across the first half, the accounts-receivable balance movement and cash-flow adjustment differ by $21 million; the deferred-revenue balance movement and cash-flow adjustment differ by $142 million. June contract assets of $179 million also belong in any complete recognition-to-cash account. “Not material” at an earlier date is not a measured zero.¹⁴⁶

The practical conclusion is positive and limited: the public accounts support continued customer-related funding and a material noncash reclassification, not the assertion that the March deferred-revenue decline proves customers withdrew that amount. They do not yield a complete gross collection ledger or identify the ultimate origin of every advance.

6.3 Investment and repayment still require much more than operating cash

A cash-generating quarter is not necessarily a self-funding expansion. The matched actual payments show the scale of the distinction:

The $9.627 billion Q2 amount is a retrospective requirement at the stated uses, not a current funding hole or an estimate of economic loss. It is the amount that would have to come from existing eligible resources or other net sources after observed operating cash, if those capital purchases and repayments were held fixed. Actual financing did occur. Some repayments replace existing debt with new claims; they are not all evidence that the operating investment lost money.

The complete Q2 account reconciles $679 million operating cash, $7.166 billion investing cash use and $10.071 billion financing cash to the $3.584 billion cash increase. Within financing, derived Q2 net debt proceeds were $13.457 billion and net equity proceeds $997 million, offset by $3.884 billion principal, $492 million capped-call purchases and $7 million other financing uses.⁵⁶

June total cash of $6.904 billion includes $1.380 billion restricted cash, leaving $5.524 billion nonrestricted cash. Those accounts are not all available to any one project. The September financing already established in the earlier research is not added to them without intervening uses. Neither the quarter's financing total nor the primary equity proceeds are customer revenue.⁶¹⁷¹⁸

There is also a disclosed favorable financing resource, not merely hope for an uncommitted capital raising. The June liquidity table lists $10.014 billion available under existing secured revolving and delayed-draw facilities, alongside $5.524 billion cash and $15 million marketable securities: $15.553 billion of reported liquidity. That is not $15.553 billion of cash already held. For example, the DDTL 4.0 disclosure ties borrowing and debt sizing to eligible infrastructure purchases, depreciable equipment cost, projected debt-service coverage and project conditions. The aggregate availability cannot be assigned freely to every parent repayment or every new GPU order. Management says its existing resources and borrowing availability should meet the following year's anticipated obligations; that is an attributed liquidity assessment, not this report's cash-flow forecast or independent compliance certificate.⁶

This distinction materially strengthens the favorable explanation: some continued investment can use already arranged financing rather than depend entirely on an unspecified future lender. The remaining test is whether the particular purchase is eligible, the conditions are satisfied on time, and the service supports the resulting claim. Comparing the $10.014 billion commitment stock directly with the $9.627 billion hypothetical repeated-quarter requirement would not establish that the latter is fully financed, because timing, entity, eligible uses and intervening draws differ.

This accounting convention does not charge interest twice. Cash interest already included in operating cash remains there. Capitalized interest embedded in capital purchases is not deducted again below the capital line. Noncash acquired equipment or the OEM balance reclassified into debt is not added as another current cash payment and another cash financing receipt. It remains relevant to future payment requirements, collateral and claim priority.

6.4 What another purchase would need

CoreWeave's June filing says its current GPU contracts use NVIDIA equipment. That establishes a material technology/procurement dependence; it does not make the entire cash-property/equipment line a direct NVIDIA invoice. Buildings, networking, other capitalized items, intermediaries and timing differences remain inside or alongside the investment programme.⁶

The last two quarters demonstrate that repeat construction and hardware purchases were supported by a mixture of customer cash timing, supplier/payable timing, new equity, debt and accumulated resources. A further worthwhile build can rationally use the same categories. The economic test is whether the resulting service, net of credits and complete costs, supports its own claims and the funding can be obtained at the required milestones.

For a hypothetical repeat of Q2's selected uses, the natural starting requirement is $9.627 billion before other net uses. It is not a forecast: the next principal schedule, capital deliveries, customer advances and refinancing transactions need not repeat. More retained operating cash reduces the requirement; larger advances supply cash now but leave future service or repayment obligations; new debt and equity provide resources and create claims. Lower investment can reduce future funding needs, but cannot automatically erase an already accepted supplier bill or a mandatory installment.

The accounts do not show a self-financing steady state at this expansion rate. They also do not show a completed default or an inability to fund the observed period. This is a more informative conclusion than either “positive operating cash proves the expansion pays for itself” or “external finance proves the demand is fake.” The next section explains why the actual customer and supplier contracts affect which of those resources remains available.

The contract changes what a funded order means

7.1 Meta's current commitment is more than an intention, but less than unconditional cash

The filed March 2026 order announcement explicitly retains service-delivery, availability and termination conditions. The older published Meta master agreement supplies important context, but the actual order takes precedence. The complete 2026 order forms, prices, delivery schedules and all amendments were not obtained. The master agreement must not be used to invent terms omitted from the new order.⁷¹⁹

Within that disclosed master agreement, termination for an uncured material breach or qualifying insolvency differs from simply deciding that another service is cheaper. If Meta properly terminates under the relevant cause provision, the text provides a pro-rata refund of unused prepaid fees. If CoreWeave properly terminates, it provides for the remaining fees, with prepaid amounts available for setoff and proceeds from the relevant server transfer or sale deducted from the amount owed. This is payment protection with recovery offsets, not permission to collect the same loss twice.¹⁹

Force majeure has its own provisions: the specified prolonged event can permit termination, refunds or relief from paying for unperformed service. Important periods are redacted. Ordinary payment terms are also partly redacted; reserved service follows its order form, while on-demand use is billed monthly in arrears. There is no basis for inventing a universal net-30 collection period.

Economic consequence: a valid current order can protect the provider against low utilization without protecting it against every service or delivery failure. Existing payment rights can remain valuable even if Meta later refuses to exercise another option. The March exercised option is therefore favorable repeat-purchase evidence, but not proof of a perpetual renewal at the same price.

7.2 NVIDIA's service-purchase promise is not the same protection

The September 9, 2025 CoreWeave filing describes an NVIDIA arrangement with approximately $6.3 billion initial order value, under which NVIDIA purchases residual unsold capacity subject to delivery and availability, through April 13, 2032. That is a genuine commercial support mechanism, not an unconditional guarantee of all CoreWeave debt or evidence that the full amount was paid at signing.²⁰

The published NVIDIA master agreement makes the distinction concrete. Its cause-termination clause releases the customer from payment for service that would have been provided after an effective qualifying termination. It preserves liability for service delivered before termination and provides for specified prepaid refunds and wind-down arrangements. That differs from the remaining-fee treatment in the published Meta agreement. The applicable order form ranks above the master terms; redactions and uninspected order-specific changes prevent a complete legal comparison of the two current commercial packages.²¹¹⁹

This is not an assertion that NVIDIA can cancel at will, that an actual termination occurred, or that CoreWeave has no enforceable protection. It is evidence against treating “large committed customer” as one interchangeable guarantee. The relevant questions are which performance is required, when fees become due, what remains after a qualifying termination and whether recovery offsets apply.

7.3 The supplier can also support the customer's unsold capacity

NVIDIA's July filing describes a broader AI-cloud model introduced in its second fiscal quarter. Selected partners procure NVIDIA infrastructure; NVIDIA enters cloud-service commitments; the partner can stop supplying that capacity to NVIDIA and sell it to third parties at more advantageous rates. The reported commitments total $36 billion, typically over six years, and decrease as third parties or NVIDIA itself use capacity. Conditional revenue-sharing rights may also arise.²

The preceding ownership investigation already retained the $36 billion schedule. The advance here is the allocation mechanism, not rediscovery of the headline amount. The provider obtains a supporting purchase commitment while retaining a route to better third-party sales. NVIDIA can earn hardware revenue, obtain useful research/development capacity and potentially share later revenue, while accepting exposure if third-party uptake disappoints. These effects can be economically sensible together; they are not automatically an accounting fraud or an unqualified arm's-length demand test.²²

The favorable hypothesis is productive coordination: a provider with real prospective users cannot initially finance enough infrastructure; support enables delivery; third-party customers then use and pay for it, reducing the supplier's support obligation. The adverse hypothesis is that a hardware sale depends on a costly supporting service promise because outside customers will not absorb the capacity at an adequate net price. NVIDIA might still receive cash for the hardware while bearing a separate later service-purchase burden.

The public record does not allocate every dollar of this programme to named partners or identify its complete overlap with other disclosed commitments. Consequently, this investigation does not add $36 billion, the separate $6.3 billion initial CoreWeave order, other cloud commitments and guarantee ceilings into a single exposure or industry-demand number. Nor does it assign the same capacity's third-party sale and NVIDIA support purchase as two simultaneously collectible revenues when the contract reduces one as the other occurs.

NVIDIA also describes preliminary financing-platform arrangements intended to mobilize third-party capital. A memorandum or intended financing scale is not a completed cash source for the next invoice. Actual underwriting, executed commitments, draw conditions and deployment still have to connect. NVIDIA's own explanation acknowledges that capital, land, power and facilities can constrain customers even when management believes underlying compute demand is strong.²

7.4 Equity funding, customer incentives and service credits must remain separate

NVIDIA's January 23, 2026 primary purchase of 22,935,780 CoreWeave shares at $87.20, approximately $2 billion, is completed issuer funding. It creates an ownership claim. It is not a secondary seller's receipt, and it is already inside the quarter's financing account rather than an additional resource to add on top. The public transaction supports the existence of an investment link; it does not trace the use of every subscribed dollar to an equipment payment.²³

The 8.75 million shares initially valued at $350 million issued to OpenAI in 2025 are a different instrument: noncash consideration associated with a customer contract. As the previous realization investigation established, that grant did not provide $350 million of cash to settle hardware bills. CoreWeave's accounting policy records customer incentives and recognizes the associated reduction in revenue over the relevant performance period. The economic bargain includes the equity value transferred, even where the customer's contractual cash payment is unchanged.¹⁴¹⁸

It would be equally wrong to ignore that incentive and to charge it twice. A revenue figure already reported net of the allocated incentive cannot then be reduced again by the same allocated amount as a second operating cost. Nor can the whole original grant be subtracted from every future invoice. A complete relationship return would match the grant, actual net receipts, service costs, remaining credits and ownership effects across dates. That full private relationship account remains unavailable.

CoreWeave's revenue policies also distinguish contract assets, invoiced deferred amounts, variable consideration and significant financing components. The one-year practical expedient for assessing a financing component is an accounting convention, not a claim that all customers pay within one year or that longer advances are free money. Contractual service credits may reduce consideration without being a customer payment failure. These policy distinctions explain why recognized revenue, a deferred balance and gross cash receipts cannot be substituted for one another.¹⁴

7.5 Loan-specific protection survives the broader funding concern

The earlier financing investigation identified CoreWeave Compute Acquisition Co. VIII, LLC's DDTL 4.0, with a Meta-supported take-or-pay proposition and projected full amortization in the rating analysis. That remains a serious favorable case. Paying for available contracted service can support a creditor even if Meta's own utilization or incremental return is disappointing. The underwriting was projected, not a current compliance certificate, and the protection is not a general guarantee of every CoreWeave obligation.¹²

The new quarter-level capital need does not invalidate that loan-specific case. Consolidated capital purchases include expansion beyond one protected borrower. Conversely, a protected loan cannot establish that all new projects have equally strong customers, funded reserves, delivery protection or matching debt maturities. Patient, appropriately structured finance can bridge a productive deployment; a headline backlog cannot establish those conditions for every claim.

A repeat purchase must pay for another useful service, not just another machine

8.1 The physical and cash clocks must agree

The completion research established why a funded project is not automatically a billable one. Shipment, installation, energization, customer acceptance and reserve release can occur on different dates. An eventual reimbursement cannot finance an earlier bill without a real bridge. The repeated-purchase analysis retains that mechanism without exporting one project's reserve formula or acceptance clause to every NVIDIA customer.²⁴

The new supplier filing provides the corresponding upstream constraint: customers can postpone orders when they lack deployable sites, power or capital. The purchase programme cannot simultaneously assume immediate supplier cash, delayed customer funding and no party bearing the intervening financing cost. Longer supplier terms, customer advances or a loan can bridge the interval, but each allocates a claim and risk.

For Meta, continued investment must coexist with its operating franchise, existing service purchases, award cash, leases and other commitments. For CoreWeave, it must coexist with power, colocation, engineering, installed equipment, debt repayment and the service owed against earlier advances. Whole-company cash is useful evidence of capacity, but not proof that cash is available inside the right restricted project entity at the right date.

8.2 Replacement is distinct from the initial construction bill

Earlier-generation equipment can remain useful. The continuation investigation documented selective re-contracting and facility reuse, while distinguishing a design's age from the age and realized profitability of a particular machine. That evidence argues against setting all older capacity to zero merely because a new architecture arrives. It does not establish that every original purchase price is recovered or that every site can be transferred at book value.²⁵

Equally, keeping current equipment running and purchasing an additional generation are different economic choices. The next purchase must cover its incremental capital, transition costs, downtime, integration and continuing operation. Reusing a powered site can reduce one-time construction needs; it does not eliminate equipment renewal, software, staffing, maintenance or power. A lease moves the cash calendar and claim ownership; it does not remove the underlying investment.

The efficiency investigation adds an important discipline: cheaper tokens or higher technical throughput do not automatically mean more paid, accepted work. Quality, review, latency, retries and utilization can change the cost per useful outcome. A better system can benefit both customer and provider, but the gain retained by a hardware supplier, a cloud seller and the final buyer need not be the same.²⁶

The present cash accounts do not estimate a universal replacement interval, a GPU residual curve or a required service-price elasticity. They identify which observed resources paid for the current programme and which conditions would support another one. The full life-cycle return is a further evidentiary question, not an omitted zero cost.

8.3 What a self-replenishing expansion would look like

A convincing favorable sequence would show delivery and collections on existing contracts; net service cash after ordinary costs and required investment; repeat customer decisions on disclosed net terms; and financing whose amount and duration fit the remaining service. Advances could remain part of that model. A steadily growing business need not stop taking advances to be sound. What matters is that future service and capital obligations can be performed without ever-larger unsupported concessions or unpayable claims.

An adverse sequence could instead preserve current payment while weakening recurrence: extended terms grow; a few customers consume earlier prepaid service; new orders need more supplier support; net realized prices fall after incentives; and later cash does not cover the enlarged investment and financial claims. A slowdown in new orders could precede an overdue invoice. That is a supported vulnerability to investigate, not a claim that this entire sequence has already occurred.

The discriminating evidence is therefore not simply whether a financing round succeeds. It is what the financed capacity delivers, what net cash the customers subsequently pay, how much renewal costs, and whether the next customer decision needs materially different commercial support.

What this changes about “Is AI a bubble?”

9.1 The strongest contemporary arguments do not all answer the same question

An October 14 Reuters report describes a CoreWeave staff memo in which management argued that its backlog more than supported its debt obligations amid market scrutiny. The memo itself was not obtained here; this is an attributed account of management's argument, not a new audited cash forecast. The article's reported market-price observations are not independently reconstructed or used as default probabilities.²⁷

The strongest version of management's case survives: committed service with capable customers can support financing, and the accounts show that substantial funding and investment actually occurred. The stronger conclusion—that the backlog by itself covers every obligation at the required dates—does not follow. Delivery gates, price credits, operating costs, investment, reserves, claim priority and the actual repayment calendar must be reconciled first. The new cash and contract analysis explains those missing steps without declaring the backlog fictitious.

The opposing claim that supplier investment or offtake support makes all commerce unreal is also unsupported. The primary share subscription, service promises, hardware sales and customer incentives are economically different transactions. A productive deployment can be financed through several of them. The concern becomes specific where current sales depend on obligations whose eventual net cost is not covered by useful third-party demand. That is a return-and-risk-allocation problem, not proof of fraud from a relationship diagram.

9.2 Evidence-weighted findings

The third and sixth findings concern specific inference errors, not findings of a default. The fifth concerns funding composition during expansion, not whether that expansion is profitable. Keeping those positive meanings clear matters more than attaching a generic “uncertain” label to every number.

9.3 The whole-question judgment

The accumulated case for genuine economic substance becomes stronger; the case for effortless, universally self-funded growth becomes weaker. A profitable supplier can collect cash while extending more credit. A profitable operating buyer can make a rational new commitment while using financial assets or long-term borrowing. A financed cloud provider can serve real customers and pay existing creditors while needing substantial new capital for the next build.

The most supportable vulnerability is concentrated repeat-purchase demand whose continuation depends on the joint behavior of customer cash, commercial support, delivery and capital markets. The current evidence does not measure its prevalence across every AI asset, and does not establish that all such dependence is speculative overinvestment. It does show why extrapolating current supplier growth from a backlog or profitable quarter alone is insufficient.

This sharpens, rather than replaces, the prior ownership result. That work found plausible, costed paths for NVIDIA and Freshworks and did not establish either security as a bubble. NVIDIA's sustained-scale requirement now has a more concrete empirical funding condition: actual purchasers must keep allocating resources on economically workable terms. The quarter comparisons do not recompute its share price, select a new discount rate or turn a funding need into a valuation loss.²²

Coverage remains U.S.-centered and mechanism-led. The accumulated work has examined supplier earnings and claims, selected large operating payers, application and buyer outcomes, construction and hosting, customer advances, secured and parent borrowing, capital recycling and selected holders. It has not produced a representative end-buyer census, complete private model-provider margins, a worldwide bank/insurer exposure register, comprehensive sovereign procurement economics or every alternative accelerator's cost curve. Meta and CoreWeave are connected cases, not two statistically independent observations of the entire market. Their revenues, obligations and investments are not summed into a “bubble size.”

A more favorable next observation would be timely realization of the extended receivables alongside further net-paid, economically supported purchases and less reliance on incremental support per unit of useful service. A more adverse observation would be repeated concessions, delayed collections beyond agreed terms or new commitments whose complete costs exceed the resources available on the operative dates. Current payment, recurring customer value and a justified ownership price remain connected questions—but none answers the other two by itself.

Methods, boundaries and reproducibility

The investigation uses original financial statements, policy notes, executed or filed commercial documents, original company operational descriptions and one attributed contemporary news account. The primary calculations are historical reconciliations and explicitly defined reference requirements. No probability model, industry-demand allocation, security-price target or current bank-account estimate is supplied.

The companion CSV distinguishes reported, reported rounded, contractual rounded, issuer operational, unobserved and assumption records. Financial amounts are in USD millions unless the unit says otherwise. A blank unknown is not an assumed zero. Stocks have an observation date; flows have their own period; transaction events retain their execution date. Numerical precision in the output demonstrates arithmetic, not precision in a rounded commitment or an unidentified economic allocation.

The code uses the Python standard library and Decimal arithmetic. It reads only a local CSV, makes no network request and writes no output files by default. It prints the results and check count as JSON to standard output. The calculation accepts a neighbouring input file or an explicit --data PATH.

The supplier account bridge is not measured gross receipts. Its adjustment term remains unidentified; the displayed bridge is not a statistical interval or a proven upper/lower bound. The near-match between receivable stocks and the cash-flow receivable adjustment is a useful control, not complete validation of the revenue-to-cash attribution. CoreWeave's five-working-capital decomposition is a reconstruction of reported cash, not a recurring-margin estimate. The Meta repeat requirement holds stated uses fixed to identify an additional-resource condition; it is not Q3 guidance or a declaration that those uses are all cancellable.

The material unresolved observations are specific: invoice-level aging and settlement; customer identity behind anonymous categories; unallocated contract assets, incentives and other changes; all operative 2026 order forms and amendments; the exact subsequent classification/settlement of customer liabilities; entity-level reserve and draw eligibility; a complete current cash roll-forward after the last statement; and the full incremental life-cycle return on the selected purchases. None is filled from a different company's margin, an unnamed customer percentage or an investment announcement.

The earlier research remains dated. The Energy Global/NVIDIA prepaid-forward qualification remains obligation established, receipt unverified, with no inference of nonpayment. The September CoreWeave closing remains established at its actual date, without becoming an unspent October resource. April's illustrative financing coupon and May's executed coupon remain separate; the historical CoreWeave put that terminated in September 2025 is not revived as a current liability. Prior observed investor receipts remain genuine, while a debt-funded parent transfer is not relabeled customer-funded profit.^28172418

Technical input and arithmetic checks do not independently establish the truth of corporate reporting, the causal benefit of an investment or a judicial outcome. No private payment certificate, confidential customer model, paid transaction dataset or current borrower-compliance certificate was obtained. Reported data and the apparent public contractual allocation carry those limits throughout the analysis.

Sources and calculations

Evidence boundary: 15 October 2026. The notes identify original documents, reproductions, attributed claims and access limits. Earlier investigations keep their own dates. Source links are deliberate links, not live data feeds.

Source-linked inputs (CSV) · Reproduce the calculations (Python)

The reader companions retain the supplied numerical records and model equations, with public provenance and descriptive filenames. Observations, accounting reconciliations and conditional scenarios remain different result types. Running the code checks arithmetic, not the empirical assumptions or completeness of the claim inventory.