We came to CoreWeave in the spring with two questions, and neither of them was about CoreWeave. We wanted to know what a pure-play neocloud could tell us about circular financing, and what it could tell us about the quality of cash. It seemed like the right specimen: one business, one product, no legacy monopoly to blend the answer into.

The answers took us somewhere we did not expect, into economic and financial rationales, into houses and players, and, not at all to our surprise by the end, back to circular financing. This is our last piece on the company, and the second half of it is where the questions land.

Two disclosures first, one about the company and one about us.

Nothing here identifies an error or an inconsistency, and we did not go looking for one. Advance rates move, spreads move, and a borrower that agrees to changed terms has agreed to terms. An equity cure right is ordinary in project finance and says nothing unusual about this borrower.

The ladder and the return calculation below are ours. The filings do not present a list of funding sources in order, and they do not compute a return on invested capital. Grouping the sources and building the ratio are both choices we made, and we have set them out so anyone can rebuild or disagree with them. Every figure carries its source, every quotation carries its section, and the accessions are public.

What the next dollar costs

The three loans are the clearest thing in the record, and all three are filed. Each one lends against equipment, and each one states, in its own definitions, what fraction of that equipment it will lend against. That fraction is the advance rate, and it turns out to be the number worth watching.

The facility got smaller by 69 percent. The spread more than doubled. And the equity required per dollar of equipment tripled.

The spread changes the price of the debt. The advance rate changes how much debt exists. A dollar of the company’s own money bought ten dollars of equipment in March and buys three dollars and thirty-three cents in August.

An advance rate is a cash-to-program ratio with a signature on it

We published a ratio earlier this summer, cash on hand against the annual capital program, and it ran near one to fifteen at March 31 and near one to six at June 30. That measure is OURS, and it describes what the company chose to hold.

An advance rate measures the same thing, and a lender sets it. At an advance of 90 percent the borrower brings ten cents of its own money for every dollar of equipment, which is one to ten. At 70 percent it brings thirty cents, which is one to three and a third.

Placed on one axis, the two measures move the same way, with the lenders moving further. The company’s ratio went from about one to fifteen to about one to six across two quarters. The contractual ratio went from one to ten to one to three and a third across five months, and it has been tighter than the company’s own position at every point along the way.

The advance rate governs only the equipment those facilities finance, so reading it as a company-wide requirement would overreach. As an illustration and nothing more, holding the whole 2026 program to the August standard would call for $10.5bn to $11.7bn of cash against $5,524m held at June 30.

What that does to the alternation

The year so far has gone equity, debt, debt, equity, debt, debt, debt. Seven capital events in one hundred and ninety-six days, all filed: a placement in January, a facility in March, notes in April, a placement in April, a facility in May, notes in June, a facility in August.

Setting the equity raised in each window against the slice the next facility requires gives the ratio below, where the slice is the facility size divided by its advance rate, less the facility size. The measure is OURS; both inputs are filed.

Twice covered, then four fifths covered, then not covered.

The measure does not trace dollars and does not assert that either placement was raised for the facility that followed it. The filings do not permit that. It is a ratio between two filed quantities, and the sequence between them is a sequence.

Two things moved in opposite directions to produce it. The slice per dollar of debt tripled. And the pace of equity raising went from $3,000m in three months to nothing in four.

So growth is rationed by the slice

$10,199m of committed delayed draw capacity remains undrawn. Drawing it in full requires $2,536m of equity beside it, and the requirement has to be built facility by facility, because a blended rate misstates it: $622m against DDTL 4.0’s $5,600m, $800m against DDTL 5.0’s $1,999m, $1,114m against DDTL 5.5’s $2,600m.

Two of those commitments expire this year.

About $1.9bn of equity is what keeps $4.6bn of committed borrowing from lapsing unused before December. It is two thirds of one placement already completed.

The size of the ask

Guided capital expenditure for the second half is $20,883m to $24,883m. Scheduled debt principal is $4,413m. Cash and equivalents were $5,524m and availability under existing facilities was $10,014m at June 30, both filed, and the August facility adds $2,600m.

That $10,014m is a different quantity from the $10,199m above, and the two should not be read as one number. The filed figure is a June 30 measure of the revolver and the delayed draw facilities together, before the August facility existed: about $2,000m of revolver capacity, $5,600m undrawn on DDTL 4.0, $1,999m undrawn on DDTL 5.0, and a residual near $415m on older facilities the filing does not size separately. The $10,199m counts only the three facilities this piece follows, as they stand today, and leaves the revolver and that residual out.

Granting the company an operating cash conversion at least as favorable as the first half’s, and crediting every committed dollar as drawn, the committed resources fall short of December by $3.8bn to $7.8bn.

The equity slice is not added to that figure. Guided capital expenditure is the whole of the equipment purchase, and the slice sits inside it. What the slice constrains is the form of the money: at least $1,914m of whatever is raised has to be equity, because equity is what releases the commitments expiring September 30 and December 31.

Against 551,536,602 shares outstanding at July 31, and at the two prices this company has transacted at, the full gap is 35m to 71m shares at $109.00, which is 5.9 to 11.5 percent of the register, or 43m to 89m shares at $87.20, which is 7.3 to 13.9 percent. The $1,914m floor alone is 17.6m shares at $109.00, or 3.1 percent. The 2027 and 2028 programs are not guided and are not modeled here.

Where the lines point

The next entry does not have to be equity. What the record shows is four independent things pointing the same way, none of them requiring a forecast, and a set of alternatives that are open and priced.

The arithmetic. Equity in this structure answers to two calls. One is the slice each lender declines to advance. The other is the loss the operation runs.

Across the eighteen months from December 2024 to June 2026, the company spent $24,426m on capital expenditure and lost $2,533m, while paid-in capital rose $7,989m. Equity funded about three dollars in every ten of losses plus capital expenditure. The other seven came from debt, customer prepayments and vendor financing, each of which ranks ahead of the residual.

A shorter version of the same sum, for the first half alone, would need the advance rate on every facility drawn in the period. Three of those rates are filed and the older facilities’ are not, and a single blended rate misstates the answer for the reason set out above. So the eighteen-month share is the figure that travels.

The loan documents. All three agreements set a debt service coverage test at the borrower, first measured in November 2026, February 2027 and August 2027, and all three name a cure in Section 7.03. What differs is how many.

The March and May agreements name one, “the receipt of Equity Proceeds (which shall be in the form of common equity or other equity in a form reasonably acceptable to the Administrative Agent).” The phrase “Additional Master Services Agreement” appears nowhere in either document. The August agreement adds a second: “solely with respect to the financial covenant set forth in Section 6.12(a), the Borrower entering into an Additional Master Services Agreement (or renewing an existing Master Services Agreement) that satisfies the Eligibility Criteria,” at a projected coverage of 1.35 and a projected contract value ratio of 2.40.

So the newest and tightest facility is also the one whose covenant can be cured with a signature rather than with cash, and a miss is narrower in consequence than it first appears: Section 7.03(b) provides that the lenders “shall not be obligated to make any Credit Extension under any Facility” while a breach continues, which turns off the draws rather than accelerating the loans.

The covenant is therefore unlikely to be the binding constraint. On two of the three facilities, the only named way through it is equity.

The filings. An automatic shelf registration statement was filed June 5, 2026, and no prospectus supplement has followed it. We pulled the registrant’s complete 2026 filing index and filtered for every 424, S-1, S-3, 8-A and free writing prospectus form type; the shelf itself is the only result. The registered route to public equity has been in place for ten weeks and stands unused.

The history. Seven capital events in one hundred and ninety-six days, two of them equity, both in the first half of the sequence, and no equity since April 15.

None of that forecloses the other rungs. Prepayments, vendor financing and the unsecured market are all open, all took more money in the first half than they held at the start of it, and each is more expensive or more heavily occupied than it was. What the three facilities record is that the lenders now set the amount, the ratio and the price of the debt. Whatever those three leave uncovered has to come from somewhere the lenders do not price, and the list of those places is short.

Subscribe now

What the capital already in has returned

That is what the next door costs. The other half of the question, and the one we came for, is what the doors already open have paid.

The two halves are one fact seen twice. A lender cutting its advance from ninety cents to seventy and a business failing to earn its cost of capital are both readings of capital intensity outrunning earning power. The lender is simply the first party in line to price it, and it prices in writing, on a date, in a filed exhibit.

Last month Microsoft’s chief executive published an application scoring the four large builders on return on invested capital, built from a Morgan Stanley analysis by Brian Nowak: 29.7 percent average, four of four clearing a 9 percent hurdle, under the line “capital is visible, attribution is not.” That framework is built for companies whose AI spending is one line inside a much larger business, which is what makes the attribution problem the card names. We wanted to run the same style of arithmetic on a company where the buildout is essentially the whole balance sheet, and we wrote the test down in a working note on July 30, before CoreWeave’s second-quarter numbers existed, promising ourselves the result would print whichever way it came out. Here it is.

Invested capital is total assets less non-interest-bearing current liabilities, which is a standard construction and the one we used on the other companies in this comparison. Debt, finance leases and operating lease liabilities stay inside it, because they carry a cost. A second cut removes construction in progress, which is capital committed and not yet producing, and that cut runs in the company’s favor. Operating income is filed. The tax normalization is ours: 21 percent on profits, no benefit taken on losses.

One divergence from the July 30 note belongs here rather than in a footnote. That note specified a second cut we did not use, net property and equipment plus net working capital, which on the June 30 balance sheet gives $35,339m of invested capital and a return of negative 1.09 percent. It is the most adverse of the three constructions available, and the figures below use the one that favors the company.

What the capital base demands comes first, ahead of what it delivered, because the demand is the larger number and it needs no forecast at all. A return on invested capital is a margin multiplied by turns, and the turns are a June 30 fact.

At 0.178 turns of revenue per dollar of capital in service, clearing 9 percent requires an after-tax operating margin of 50.6 percent. On the second-quarter revenue rate annualized, turns are 0.197 and the requirement falls to 45.7 percent. The company’s own stated target, given on the first-quarter call, is a low double-digit adjusted operating margin by year end, on a measure that excludes stock compensation and intangible amortization and is stated before interest and tax.

Held the other way, on the June 30 base and with no further capital, clearing 9 percent at a 20 percent after-tax margin would take $23,553m of revenue, which is 2.29 times the second-quarter rate annualized. At a 15 percent margin it takes $31,403m, or 3.05 times.

A word about that 9 percent, because it is borrowed and it is doing a lot of work. It is Morgan Stanley’s, set for four investment-grade companies with very low costs of capital, and it appears here only because using one yardstick is what lets three different companies sit on one scale. For this company it is a lenient yardstick. CoreWeave’s own filed weighted cost of debt is 9.27 percent excluding the convertible notes, and the facility it signed in August prices near 9.8. The bar that matters to a CoreWeave shareholder sits above the line, not on it, which means every figure below is measured against a standard easier than the one the company pays.

And the denominator keeps moving. Invested capital rose from $15,609m to $64,257m across those eighteen months, a factor of 4.12, while annualized operating income fell $710m and crossed from positive to negative on the way. Guided capital expenditure of $35bn to $39bn this year adds more, so the margin the business has to reach is a target that recedes while the argument about it continues.

Against that, the level. Return on capital in service comes to 2.06 percent for 2024, negative 0.15 percent for 2025, and negative 0.74 percent for the first half of 2026 annualized. Taking a 21 percent benefit on the losses moves the last figure to negative 0.58 percent. Measuring pre-tax throughout moves 2024 to 2.61 percent. No convention available reaches 9 percent, and none reaches the company’s own filed cost of money, which is 6.34 percent weighted across the whole stack and 9.27 percent excluding the convertible notes.

The one period in the filed record with positive operating income returned about a fifth of the hurdle.

Revenue grew faster than capital, at 4.86 times, and that is the one line here that runs the company’s way. Capital turnover improved with it, from 0.154 to 0.167 to 0.178. The improvement is real. The level is the constraint.

Which is the point worth holding onto, because it disposes of the argument most often made on the other side. Demand is not the constraint. Remaining performance obligations grew 70.8 percent in the half, to $103.7bn, and revenue has run at 2.12 times year on year for four consecutive quarters. This company can sell what it builds. What the return calculation measures is something else: whether it can sell what it builds at a price that covers the capital required to build it. So far that answer sits below zero, and it sits below zero while demand is strong, which is the harder version of the finding.

What the arrangement costs to run

Arranging money costs money. The filings state the amount, and until now we had walked past it.

The debt note carries unamortized discount and issuance costs of $427m on recourse debt and $56m on non-recourse debt at June 30, against $242m and zero at December 31. The cash flow statement shows $86m amortized through income over the same half. So $327m of new debt friction was incurred inside six months, and $18m of issuance costs on the two equity placements sit beside it.

That is $345m of identified friction in six months, against $19,729m raised, or 1.75 percent of it. Carried against invested capital and annualized, it runs at 1.07 points a year.

So the hurdle is not 9 percent. It is about 10.1, and that is a floor, because arranger and agency fees sit in Fee Letters that the credit agreements reference and do not attach, and because the discount on the August facility falls in the third quarter.

Carried back into the return calculation, that friction lifts the required after-tax margin from 50.6 percent to 56.7. It is collected at close, whether or not the return arrives, and it recurs with each arrangement.

Two comparisons make the size legible. The $345m of six-month friction exceeds the $324m of operating income the company earned in its one profitable year. And on arithmetic we published earlier, a payback of 27 years at the spread we found, against a six-year depreciable life, means the equipment turns over about four and a half times inside one payback period. Each turn is financed, and each financing carries the friction again.

Who is on the other side

The first half of 2026 raised $3,073m of paid-in capital, of which $2,982m was two private placements.

NVIDIA bought $2,000m at $87.20 on January 23. NVIDIA supplies the equipment these facilities finance, holds a master services agreement as a customer, and was already a shareholder. Jane Street bought $1,000m at $109.00 on April 15, alongside a $6.0bn compute commitment announced with it. Jane Street is a customer and was already a shareholder.

Every dollar of equity raised in the half came from a counterparty that also transacts with the company commercially. Both were reported under Item 3.02 as unregistered sales. No motive is imputed and none is needed. The observation is about who was on the other side, and it is filed.

The record carries a third instrument of the same family, and the three are not alike.

In March 2025, on reported terms, OpenAI received $350.0m of stock in connection with a contract of up to $11.9bn, and no cash reached the company. In January and April of 2026, on filed terms, NVIDIA and Jane Street paid $2,000m and $1,000m for stock at $87.20 and $109.00, and the cash did reach it.

One issued stock and took no cash. Two sold stock and took cash at a price. One transaction of each kind establishes no rate for either.

The supplier is the load-bearing party

Of the counterparties on that list, one holds a position that the others do not, and following it changes how the whole arrangement reads.

The collateral under all three delayed draw facilities is equipment, and the definition each agreement uses to size a draw is the “Funding Date GPU Amount.” So the advance rate is a lender’s judgment about the equipment and about the borrower together, and the filings do not separate the two. What can be said without separating them is that the asset the lending market is pricing is the supplier’s product.

The plumbing runs the same way. Under the August agreement each draw request states the purchase orders it funds; where the borrower acquires servers from its parent, the lenders receive “a copy of each invoice from the Parent to the Borrower”; and within seven business days of a borrowing, proceeds may be distributed to pay those invoices. The parent sells the servers to its financing subsidiary, the lenders fund seventy cents of each invoice dollar, and the cash moves out to the supplier. That sequence is in Sections 2.03, 4.02 and 6.06 of the credit agreement.

Now the sizes. Technology equipment on this company’s balance sheet grew from $9,146m gross at December 2024 to $33,823m at June 2026, an increase of $24,677m. Over the same eighteen months the supplier bought $2,000m of stock. The supplier’s share of that equipment is not disclosed, and even so the order of magnitude is the finding: its commercial exposure to this company continuing is measured in tens of billions, and its equity exposure is two.

That is the asymmetry, and it is the reason the arrangement holds. An equity holder is paid out of the residual, last, if anything is left. A supplier is paid on delivery, at its own margin, and keeps the stock besides. The supplier’s reason for wanting this company to keep buying sits on the supplier’s income statement, not on this one, and it has very little to do with what the residual is worth.

Which is also why the return calculation above is only half the picture. Run the same arithmetic on the supplier and it comes out at 58.6 percent on its fiscal 2026, against negative 0.74 here.

The rest of the ladder

Every other source of money is open. Each also has a ceiling, and the filings describe where it sits.

The starting condition is the company’s own: an operating loss of $193m and a net loss of $1,366m for the half, and “We have generated significant losses from operations, as reflected in our accumulated deficit of $4.0 billion as of June 30, 2026.”

Working capital supplied 40 percent of the half’s operating cash, and 92 percent of that came from customer prepayments. Payables and accrued liabilities rose $2,661m on the balance sheet but contributed only $304m to operating cash, because the rest belongs to the equipment program: accrued purchases of $5,520m, accrued interest of $410m, other of $494m. This rung grows with the buildout rather than independently of it, and its operating elasticity ran about $300m in a half whose program exceeded $20bn.

Customer advances are the rung with no coupon, and they are also a claim. Deferred revenue grew $1,507m on the balance sheet and the cash flow statement credits $1,365m of it. Cash of $5,524m stands against $9,692m of deferred revenue, so the company holds 57 cents for every dollar of compute paid for and not delivered, and each further dollar enlarges an unsecured claim ranking ahead of the residual. Olga Usvyatsky of Deep Quarry and Francine McKenna of The Dig traced the $1.3bn reclassification inside that arc in a joint piece on August 15, and readers who want that question should go to them for it.[1]

Collateral is largely committed. About $21.5bn of assets sit inside the special-purpose structures, against property and equipment of $46,736m, and the filing does not characterize the remainder. A sale of collateral is a mandatory prepayment event, so a sale from inside pays the lenders first.

The other debt rungs are open and priced. The revolver carries $2.0bn of remaining capacity with $533m of letters of credit reducing availability, at a quarter point on the undrawn part, and those letters are issued “primarily in support of certain lease obligations,” which puts bank credit behind the landlords. OEM and vendor financing stands at $5,102m, up $937m in the half. Four note issues cleared in the half at 9.27 percent weighted, excluding the convertibles.

A fourth delayed draw facility is possible; three cleared in five months and nothing in the record says a fourth could not. A fourth on August’s terms at August’s size would require $1,114m of equity before a dollar of it could be drawn. This rung cannot open itself.

Why it continues, and rationally

None of the above forecasts an ending, and the arithmetic of the half runs the other way. This is the part of the story a bear tends to skip, so it is worth giving in full.

The accumulated deficit grew $1,366m in the first half while paid-in capital grew $3,073m, so the corpus was refilled at 2.25 times the rate it was drawn. Book equity rose from $3,335m to $5,024m even as the deficit rose. Book equity per share rose from $6.64 to $9.12, up 37.2 percent, because the placements were struck at roughly nine times book, and selling equity above book raises book per share for the holders who stay. That is a standard result and it holds here.

The clock on the residual does not start until issuance stops, and issuance in every form has continued. What the residual holds against all of that is the ninth position of nine, behind a $1,300m repayment gate.

What this says

We came to this company to understand the pure-play neocloud and its place in the financing chain that the large builders sit at the other end of. That question has an answer now, and the answer is not that the arithmetic fails to add up. It is that the arithmetic which decides this arrangement is computed on other balance sheets.

Measured here, the capital in service returned negative 0.74 percent annualized in the most recent half, against a filed cost of debt of 9.27 percent excluding the convertible. The friction of arranging the money lifts the hurdle above 10 percent before the first dollar of return. The denominator grew 4.12 times in eighteen months and the numerator went the other way. Forty-four cents of every dollar ever contributed to this company has been consumed, and that fraction did not move across a half in which $3,073m of paid-in capital and $26,079m of liabilities arrived.

Measured elsewhere, each participant’s own trade closes. The lender holds a 70 percent advance rate against contracted equipment at SOFR plus 550, senior, secured, dated. The arranger was paid at close. The customer prepays and ranks ahead of the residual.

And the supplier, which is the party that matters most, was paid on delivery at its own margin and kept the stock besides. On the same construction used here, it earned 58.6 percent on its invested capital last fiscal year. Its commercial exposure to this company continuing is measured in tens of billions of equipment; its equity exposure is two. Whatever reason it has for wanting the buying to continue, that reason is computed on its income statement, and the value of this company’s residual has very little to do with it.

Every one of those counterparties gave consideration and every one of them was paid. None of their returns requires this entity to clear a 9 percent hurdle, and none of them is computed on this entity’s income statement.

Which puts the position plainly. This company sits in the middle of the chain, by position and by choice, and across every period measured here it has not earned a return above the cost of the capital it commits. The party on one side of it earned 58.6 percent on the same construction. The party on the other side cannot be measured from outside, and we would rather say so than guess. They are not victims of the arrangement, since they signed every agreement described here. On this evidence they are also not the ones collecting from it.

So the corpus holds value to the participants independently of its own return. That is the mechanism, and its four parameters are now on the record: the advance rate that releases the debt, the slice of equity that releases the advance, the friction taken at close, and the coverage calendar that dates the next tests. Those four are what we came looking for, and the circular financing we came asking about turns out to be the thing holding all four in place.

We should say the harder version of that plainly, because it is where we ended up. We could not construct a version of this business that works without the circle. We offer that as a limit on our own imagination as much as a finding about the company, and we would rather be shown wrong than be right about it.

None of it stops the entity from continuing, and the rationale for continuing is solid. Every rung of the ladder is open and each is more expensive than it was, which describes a cost rather than a wall.

What the record does say plainly is that the price of continuing is paid in a currency the residual holder supplies: equity that releases debt, friction collected at close, and a return that has not yet appeared on capital that quadrupled.

The larger arithmetic sits one level out, with the participants whose round trips do close, and that is where we go next.

What it does not say

It does not forecast a default, a missed covenant, or a failure to raise. This company has raised capital every two to three months for two years and the record says it can.

It names nobody as the writer of the next check.

It imputes no motive, and every arrangement described here is stated as an effect rather than as a design. Every counterparty named gave consideration for what it received.

And it does not claim the return calculation is the company’s own. The filings do not compute a return on invested capital. The construction is ours, the inputs are filed, and both are printed above so the arithmetic can be disagreed with on its own terms.

How this could be wrong, and why the list is short

We usually close with a falsification block. Most of what would normally go in one belongs elsewhere here, and the reason is a method point worth stating.

The return, the friction and the counterparty findings measure periods that have closed. A restructuring charge, a sale of the company, an amendment, a wind-down, or a decade of better returns would each change what happens next without changing what happened. An event after the measurement date is not a falsifier of the measurement. What can falsify a measurement is a restatement, a construction objection, or an arithmetic error, and those are the three that follow.

A restatement of any filed input moves the result. The 2026 figures are unaudited, and the round trip in the contract balances meets its first audit at the fiscal-year 10-K.

The invested capital construction is ours, and it is the place to attack the return. Three cuts are shown, including the most adverse, from the July 30 note. A reader who would cut differently should name which assets and rerun it. On the in-service cut, the hurdle needs an after-tax margin near 46 percent at the second-quarter revenue rate, and any construction that brings that requirement into reach falsifies the section.

One claim here is prospective and it carries the ordinary falsifiers. The gap arithmetic is a statement about the next four months. It is wrong if the capital program is slowed to match available cash, which the company controls and which its liquidity representation is written to accommodate. It is wrong if a facility clears at a higher advance rate than the last three. And it is wrong if an investment-grade customer pays a large advance without stock attached, or if a registered takedown places equity with buyers who have no commercial relationship with the company.

What is not disclosed

Current commitments for capital expenditure, the single number that reconciles the liquidity representation to the guidance.

Arranger, agency and rating agency fees, which sit in Fee Letters referenced in the agreements and filed nowhere, and which would convert the friction floor into a total.

The size of the variable consideration deduction taken against remaining performance obligations. The company names its components as “estimates of future potential credits to customers under availability of service agreements, amounts that may not be recognized as revenue due to delivery delays, and estimates of committed cloud computing capacity that the Company has the right to resell,” and sizes none of them, while the risk factors record a realized instance of “insufficient power to service a customer’s project” that required service credits.

That last item is the most interesting number left on this company, and the reason it is interesting is not ours. Olga Usvyatsky and Francine McKenna have the dated chronology around the variable consideration language, when it left the filings and when it came back, which is reporting the documents alone will not give anyone. Their thread and this one meet there, on the one figure neither of us can size.[1]

The split of cost of revenue between rent, metered power, personnel and power-plant depreciation.

The amortization schedules at each special purpose borrower, which set the coverage-test denominator.

Whether property and equipment outside the special-purpose structures is encumbered.

The identity of the single customer contract behind DDTL 4.0.

Standing Disclosure

Cape Fear Advisors holds no direct position, long or short, in CoreWeave or in any company named here; any exposure is indirect, through managed funds it does not control, which may now include index funds holding the public companies named. Anthropic is the developer of Claude, which is used in preparing this research. That nearness cannot be checked away, which is why no claim here rests on trust in the tool: every figure carries a public source, and the record grades the rest. Others named in this piece have ties to the filer: NVIDIA supplies the equipment the facilities finance, is a customer, is a shareholder, and is measured here on the same construction as the filer; Jane Street is a customer and a shareholder; OpenAI is a customer that received stock; JPMorgan and MUFG are arrangers and JPMorgan is administrative agent on the August facility, with Morgan Stanley holding those seats on the May facility and publishing the return framework cited here; Microsoft is a customer of the filer and published the application that framework appears in. Companies not named here may hold positions or supply relationships that bear on the filers discussed, and that possibility is part of why every piece is re-checked for bias, ground facts, and filings rather than read against a fixed list. Figures are quoted from the filers without characterization, and the same standard of reading is applied to every filer named.

NOTES

[1] Olga Usvyatsky and Francine McKenna, “What CoreWeave’s $1.3 Billion deferred revenue reclassification might be telling us,” Deep Quarry, August 15, 2026. Deep Quarry is Olga Usvyatsky’s publication and The Dig is Francine McKenna’s; the piece ran jointly on the former.

Corrections and provenance, August 16, 2026. One correction to our own published work. The DDTL 5.5 ledger of August 14, 2026 described that facility’s cure right as taking only equity. Section 7.03 of the August agreement names a second cure, a qualifying new or renewed Master Services Agreement, and the published sentence was incomplete. Re-reading the March and May agreements for this piece establishes the fuller picture: those two name the equity cure alone, and the phrase “Additional Master Services Agreement” appears in neither. The correction runs against our own earlier reading and is stated here rather than left to a footnote elsewhere.

Two further items are stated because a reader checking the filings will meet them. The April 14, 2026 Form 8-K reports “$1,750,000 aggregate principal amount” of 9.750% Senior Notes due 2031 while the June 30 balance sheet carries that line at $2,750m, and the same 8-K writes the convertible notes as “$4,000,000,000”; the balance-sheet figure is used here. And the revolving credit facility reconciles across both dates to a commitment near $2.5bn: December carried $1,000m drawn, $294m of letters of credit and $1.2bn remaining, and June carried none drawn, $533m of letters of credit and $2.0bn remaining.

Method. FILED figures come from the Form 10-Q for the quarter ended June 30, 2026, accession 0001769628-26-000366, the Form 10-K for fiscal 2025, accession 0001769628-26-000104, the Form 8-K filings cited by accession, or the credit agreements filed as exhibits to them: DDTL 4.0 at Exhibit 10.1.1 to accession 0001769628-26-000129, DDTL 5.0 at Exhibit 10.1 to accession 0001769628-26-000236, and DDTL 5.5 at Exhibit 10.1 to accession 0001769628-26-000357. GUIDED figures come from the August 11, 2026 earnings materials. REPORTED figures come from third-party or company publication and are labeled where used. NVIDIA figures come from its Form 10-K for the fiscal year ended January 25, 2026, on the same invested capital construction and the same tax normalization. The second cut, which removes construction in progress, is applied to CoreWeave and not to NVIDIA; applying it to NVIDIA would raise NVIDIA’s figure, so the comparison as drawn understates the distance between them. Everything else is OURS, including the coverage ratio, the ladder, the equity slice arithmetic, the invested capital construction, the tax normalization and every return figure, for both companies. The 9 percent hurdle and the 29.7 percent comparison are Morgan Stanley’s, per Brian Nowak’s analysis as rendered in the application published by Microsoft’s chief executive on July 30, 2026, and described there as illustrative.

Two disclosures first, one about the company and one about us.

Nothing here identifies an error, an inconsistency, or a bad actor, and we did not go looking for one. A credit agreement that ranks its claimants is doing what credit agreements do, and a borrower that agrees to those ranks has agreed to ordinary terms. The subject here is the order, and what the order costs the participant at the bottom of it.

The ledger is a construction of ours. The document does not present itself as a list of participants, and grouping it this way is a choice. Every quotation carries its section, and the accession is public.

What $2.6 billion buys

Term SOFR plus 5.50 percent, maturing September 1, 2031, commitments ending December 2026, rated Ba2 and BB+, arranged by JPMorgan and MUFG. The borrower is a limited purpose entity; the parent guarantee is full recourse. Collateral is the machines the loans buy plus customer contracts “averaging approximately three years in length,” against a loan of about five.

Sized against what the company spends, it is small. Cash paid for property and equipment in the first half of 2026 was $14.1 billion, about $2.35 billion a month (FILED). The loans are capped at 70 percent of the capital expenditure they finance, so $2.6 billion supports about $3.7 billion of building. That is about seven weeks of the current program, of which the loans themselves fund under five (ours).

Seven weeks. The rest of this piece is the price of them.

The order of payment

Section 2.20(b) applies the cash “at the following times and in the following order of priority,” nine steps. Three of the participants named in it were not on our ledger before we read the section.

The rating agencies are paid at step two, in the operative text: “indemnities, administrative, rating agency and legal fees and expenses (including fees, charges and disbursements of counsel).” Step three is the lenders’ scheduled interest. The agencies rank above the coupon on the debt they rated. That is where the drafters put them, and nothing more is claimed here than the placement.

The parent, as manager, is paid at two different steps, and the drafting separates the two. Step one covers operating expenses for the next sixty days. Its proviso admits affiliate payments under the Management Agreement only as “reimbursements of the actual cost of Operating Expenses paid by such Affiliate on behalf of the Credit Parties.” Step seven then expressly includes “any costs and expenses in respect of the Management Agreement.” The manager recovers its costs first in the whole waterfall, ahead of the rating agencies and ahead of the coupon, and collects the fee at step seven. One agreement, two seats, split along the line between cost and fee.

The swap counterparties are paid at steps three and four, ratably with the lenders. The facility requires hedging, and the hedge providers rank with the debt rather than behind it.

The parent appears in this document four times: as guarantor, on full recourse; as seller, since where the borrower buys from the parent the lenders receive the parent’s invoices and proceeds may be distributed up to pay them within seven business days; as manager, at steps one and seven; and, through its ownership of the borrower, as the residual claimant at step nine. It pays at one of those and collects at three.

The ninth step is not a payment

Cash reaching step nine goes “to the Distribution Reserve Account,” and leaves only if all five Distribution Conditions are satisfied and the date falls after the Commitment Termination Date. Four of the five are what any lender writes: representations true, no default, no cash shortfall, liquidity satisfied. The fifth is the one that matters.

“The outstanding principal balance of the Loans is equal to or less than the Minimum DDTL Amount.”

“Minimum DDTL Amount” shall mean $1,300,000,000.

Half the facility must be retired, $2.6 billion down to $1.3 billion, before one dollar leaves the structure. The residual claimant’s step is a holding account, and the key to it is repaying half the loan.

The price of the seat

The entry price is a definition: “’Funding Date GPU Amount’ shall mean, with respect to any Credit Event, the amount equal to the product of (x) 70% and (y) Funding Date Capital Expenditures.” That is the whole definition.

So $2.6 billion of loans implies about $3,714 million of capital expenditure and about $1,114 million funded from outside the facility (ours). One qualification, run once and carried thereafter: the text caps the loans and requires the balance to be funded, without naming equity as the source. In practice, across a draw period ending in December 2026, the balance comes from the parent, and the parent’s own capital in 2026 is placement proceeds.

At 2026 placement prices, $1,114 million is roughly 10.2 million shares, about 1.85 percent of the 551.5 million outstanding at July 31, 2026 (ours). The discount reported during marketing adds about $78 million, arranger fees an amount not disclosed, and the undrawn balance half a point a year. Roughly $2.00 to $2.15 per share of hurdle, added by one facility, to buy seven weeks.

The stated split is also not the cash split. The 70 percent cap runs on the loan amount. Funded at 97 as marketed, the lenders advance about $2,522 million in cash against $3,714 million of capital expenditure. About $1,192 million comes from outside the facility rather than $1,114 million, and the split in cash is nearer 68 / 32 than 70 / 30 (ours, on the reported discount). Undisclosed fees sit on top, so 32 percent is a floor.

Three months earlier the same facility was cheaper in three ways

The published comparison has been the spread: SOFR plus 4.50 on the May facility, SOFR plus 5.50 on this one, same ratings from the same agencies, same unconditional parent guarantee. We wrote that comparison ourselves. Set side by side, the two definitions show the spread to be one of three changes.

May’s advance rate was 71.42 percent; August’s is 70.00, so the outside-funded leg rose from 28.58 percent to 30.00. And May’s loan amount was grossed up to cover “fees, premiums and expenses incurred in connection with the Transactions, including any fees and expenses incurred pursuant to the Fee Letters.” August has no such clause. Fees moved from inside the loan to outside it.

As effect rather than design: the marginal secured dollar became more expensive in three dimensions at once, and all three land on the same participant. At its own rate, the May facility required about $1,240 million from outside. The two together call for about $2,354 million against $2,982 million of private placements in the first half, so about 79 percent of the year’s equity is committed to two credit agreements (ours).

The third change answers a question the other two raise. What did the friction cost, and who paid it?

The bill has one number on it and four blanks. The number is the discount, reported during marketing as widening from 99 to 97, about $78 million on $2.6 billion, paid to the lenders at close. It is REPORTED and carried as marketed, because the credit agreement contains no discount of any kind, the phrase appears zero times, and the Fee Letter is referenced eleven times and is not filed. The blanks: arranger fees to JPMorgan and MUFG; rating agency fees at step two; administrative, agent and legal fees, also step two. All undisclosed. A half point a year runs on the undrawn balance, at a rate that is disclosed and against a schedule that is not.

The total is therefore not computable from the filing, and the visible floor is about $78 million. The second question the agreement does settle. In May the loan was grossed up to cover the fees, which means the lenders lent the money that paid them. In August there is no gross-up, so the fees come from outside the facility, which is the parent, which is the common stock. The friction did not only get larger. It changed pocket.

The same cover pages carry a fourth change. Morgan Stanley was Administrative Agent and joint lead arranger in May, with JPMorgan among eleven bookrunners; in August the two have traded places, MUFG holding the same seat in both. What the top seat earns is in the unfiled Fee Letter. The rotation is the fact.

The standing obligation

The coverage covenant is 1.35 to 1.00, first tested around February 2027. Section 7.03 supplies the remedy, and the remedy is the ledger in one clause: contributed equity is “deemed to increase the Net Operating Income” for the test. There is no limit on the cure until three months past the draw period, and the first test falls inside that window. Thereafter it is available three months in every four, for the life of the facility. A missed principal payment is not an Event of Default if cured in time, and an insufficient sweep “shall not be deemed to be a Default or Event of Default.”

When the asset fails to cover its own debt service, the residual claimant may write a check, and the document records the check as though the asset had earned it.

Whether that clause binds is a question the filings answer without assumption. The company’s own sentence for FY2025: “our cash flows dedicated for debt service requirements totaled $4.4 billion... our net cash provided by operating activities was $3.1 billion.” Negative $1.3 billion before a dollar of capital expenditure, and cash capital expenditure was $10,309 million. The first half of 2026 runs the same way.

A second reading comes from this shelf’s quality-of-cash work. FY2025 operating cash flow was $3,058 million, of which deferred revenue contributed $4,174 million. With the customers’ prepayments set aside, the operating line is negative $1,116 million; the same subtraction gives negative $153 million for FY2023 and positive $700 million for FY2024 (ours, from filed lines). None of that is an allegation. Prepayments are normal in contracted-capacity businesses, the company states the convention at 15 to 25 percent of contract value, and customers receive reserved capacity for the money. The subtraction says what the operating line is made of.

The two facts are correlated, which closes the point on the cure. Prepayments arrive with signings. If signings slow, operating cash flow falls, the coverage test misses, and the cure asks the residual claimant for equity at the moment the equity market is least disposed to supply it. The document permits the cure. It does not supply it.

Inside the common stock

The last row of Exhibit 2 is one line, and four different positions.

The count first, on a base that post-dates both strategic placements. The cover page of the Form 10-Q for the quarter ended June 30, 2026, as of July 31: 458,871,690 shares of Class A at one vote each, 92,664,912 shares of Class B at ten votes each, no Class C. The multi-class risk factor states that “As of June 30, 2026, our Co-Founders collectively hold all of the issued and outstanding shares of our Class B common stock.”

So 16.8 percent of the shares carry 66.9 percent of the votes (ours, from the filed counts). With the two announced strategic purchases netted out, the rest of the Class A is 77.4 percent of the shares carrying 30.8 percent of the votes.

The column that does the work is the one asking what else each holder has.

NVIDIA has three accounts where every other shareholder has one. The share: $2.0 billion at $87.20 in January 2026, about 22.94 million shares (FILED, Note 16). The supplier account: the machines. The concentration table names three suppliers at 23, 20 and 17 percent of purchases and identifies none of them, so the supplier seat is a fact about the product rather than about the table. And a customer account, in the “Master Services Agreement between CoreWeave, Inc. and NVIDIA Corporation, dated April 10, 2023,” carried on the FY2025 exhibit index alongside the Microsoft, OpenAI and Meta agreements (FILED). Whether that contract sits inside remaining performance obligations, and whether it arrived with a prepayment, is not disclosed.

No tracing of dollars is required, and none is available; nothing ties any investment to a use of proceeds. A supplier account ranks ahead of a residual claim by its nature: a seller is paid on delivery, a shareholder is paid last or not at all. One party holds a claim on the residual and two claims that do not wait for it.

The choice of account has a price, which is the Forge House argument with an advance rate attached. Under a 70 percent cap, equity of $2.0 billion supports about $6.67 billion of capital expenditure, of which about $4.67 billion is borrowed. The same $2.0 billion delivered as a discount on the chips is $2.0 billion less capex, $2.0 billion less collateral, and no equity, so it creates no borrowing capacity at all. Same cash, same silicon, same party, and the choice of form decides about $4.67 billion of borrowing. The covenant architecture then makes it recursive, because a price concession cures nothing and the cure socket takes only equity.

Jane Street has two accounts, and one of them does three jobs. The share: $1 billion of Class A at $109.00, announced April 15, 2026. The customer account: in the same release, “$6 billion to use CoreWeave’s AI cloud platform,” described as expanding an existing relationship (FILED). Jane Street is not a lender, an agent or a manager, and it is paid at no step of the waterfall.

The customer account is the one that multiplies. The contract is revenue. It is also collateral, and a large share of it, about 35 percent of this facility’s committed contracts, second to Anthropic at about 40 percent (REPORTED, Dubitsky, August 4). And the counterparty’s own credit rating sets how much of that contract counts in the covenant.

That third job lives in the parent guarantee, which requires contracted revenues of at least $1,000,000,000. The test is weighted: “the sum of (i) the reasonably projected contracted revenues from such contracts with counterparties which have an Investment Grade Rating (as defined in the Revolving Credit Agreement) and (ii) the product of 0.75 and the reasonably projected contracted revenues from such contracts with counterparties which do not have an Investment Grade Rating.” A customer’s dollars count at 100 cents or at 75, depending on that customer’s own rating. A counterparty’s creditworthiness is a term of the loan that funds the machines that serve it.

The definition sits in a different agreement, and the guarantee points at it “as amended... from time to time,” so it moves without this facility being touched. Amendment No. 4, dated November 10, 2025, rewrote it. The text now reads: “’Investment Grade Rating’ means a corporate family rating of at least ‘BBB-’ or higher from S&P or ‘Baa3’ or higher from Moody’s or, if no rating of Moody’s or S&P then exists, the equivalent of such rating by any other Nationally Recognized Statistical Ratings Organization.” The fallback to other agencies opens only where neither Moody’s nor S&P rates the counterparty, and both rate Jane Street: S&P at BB in June, Moody’s at Ba1 on the parent and Baa3 on four operating companies in July (REPORTED). Which of those entities signed with CoreWeave decides the answer, and no filing shows it.

The co-founders hold all of the ten-vote class, and employment, and compensation is an operating expense, which is paid before any residual. Control does not wait for a waterfall at all.

They are also the one holder in the exhibit moving the other way. Between February 2 and July 29, 2026 the three sold 19,184,744 shares for gross proceeds of about $1.86 billion, and 98.9 percent of those shares were sold under Rule 10b5-1 plans adopted between May 2025 and March 2026, every one of them before this facility was marketed (FILED, Forms 4; the aggregation is ours). Those are secondary-market sales. The proceeds neither came from nor went to the company, and no part of them touches this financing. The co-founders retain 92.7 million shares and about two thirds of the votes, so it is a reduction rather than an exit.

The ledger records direction, not conduct. Over the same seven months, inside one class of stock, NVIDIA bought, Jane Street bought, the co-founders sold, and the last row held. The seller keeps a great deal of option value. What separates it from the last row is that it also holds the cash.

The last row holds the share, and the column says nothing besides the share.

The covenant counts revenue, and revenue at this company costs money

That $1 billion runs against the intuition the number invites. The test is a revenue test, not a margin test, a cash test, or a profit test. Contracted revenue satisfies it, and contracted revenue at this company has arrived with a loss attached: FY2025 revenue of $5,131 million against a net loss of $1,167 million, or roughly 22 cents of net loss per dollar of revenue (FILED; the ratio is ours).

A covenant met by signing more contracts is met by taking on more delivery obligation, and delivering it consumes cash. The 75 percent haircut runs the same way from the other side: a customer without an investment grade rating must contract about $1.33 of revenue to produce $1.00 of covenant credit. The test can be satisfied by volume.

One caveat. Unit economics on any particular new contract are not disclosed, and a marginal contract need not carry the average. The narrower claim stands: a revenue covenant reads as protection, and on this filer’s record revenue is not the same thing as cash, and is not yet the same thing as profit.

Two ladders, both nine, measuring different things

This shelf already runs a nine-rung ladder, and the coincidence needs a word. That one measures distance from cash: cash on the barrel at the bottom, then trade payables, supplier finance, customer prepayments, component intermediation, vendor financing, equity in the counterparty, contingent support, and structures at the top. It is a taxonomy, and it ranks nothing. Section 2.20(b) measures order of payment inside one structure. Different axis, same number, and the two do not read across.

They touch at exactly one place. This entire facility is a single instance of the top rung, a structure, and inside it sit participants who reached it by the rungs below: customer prepayments funding the operating line, vendor financing on the equipment, equity in the counterparty on both the supply side and the demand side. One structure at the top of the ladder, holding four lower rungs inside it, with its own nine-step order deciding who is paid first.

The inversion

In the standard treatment a longer horizon raises the value of levered equity, because the option has more time. Here the extension is bought by the equity holder, so lengthening the horizon raises the strike.

That result is a statement about structure rather than outcome. A benefit exists: the machines produce revenue and the residual claimant owns the residual. The costs are certain, immediate and quantified. The benefit is contingent, deferred, behind eight steps and a $1.3 billion repayment gate, on contracts averaging three years against a loan of five. The Re-lease Eligibility Criteria require a Qualified Customer, US dollars, assignability, an approved site, a tenor “equal to, or longer than, six (6) months,” and terms “not materially worse, taken as a whole,” self-certified by the borrower’s own officer. There is no stated price floor.

The same arithmetic runs at stack scale. Approximately $10.0 billion was committed and undrawn at June 30, and drawing it at this facility’s rate would require about $4.29 billion of outside funding. That is an illustration at one facility’s rate, not a computation across the stack; the two agreements we have read use different rates, and as an illustration it holds. Undrawn capacity is presented as liquidity, and under an advance-rate structure it is a right to borrow that can be exercised only by first spending capital the company has not raised.

The whole thing, from the last seat

Read from the ninth step, the transaction comes to five sentences.

Real dollars went out: about $1,114 million to draw the facility, plus about $78 million of discount as marketed, plus fees nobody outside the deal can size. More dollars will have to be raised, because a cap of the same shape, at rates that vary by facility, governs every further draw, and about $10.0 billion sits committed and undrawn. Everyone ahead in the order has to be repaid, on dates the document sets, secured by the machines and guaranteed by the parent. The losses through the period have to be covered from somewhere else, because operations after debt service ran negative $1.3 billion in FY2025 and the covenant that protects the lenders counts revenue rather than cash. And the return, if there is one, sits nine steps away behind a gate that opens when half the loan is gone.

None of that is a forecast. Every clause of it is in the agreement or in the filings, and the last clause appears in the agreement twice, once as the order of payment and once as the condition on the reserve.

Seven weeks of building, about 1.85 percent of the company, and a standing obligation to fund the covenant three months in four for five years. Every other participant at this table has a date.

Related, on the Quality of Cash shelf: CoreWeave, Adding It Up: One Filing, Two Announcements; CoreWeave: The Sum of the Participants; CoreWeave: What Has to Happen Next; NVIDIA, the Forge House; Quality of Cash: Circular Financing and the AI Bubble.

Standing disclosure

Cape Fear Advisors holds no direct position, long or short, in CoreWeave or in any company named here; any exposure is indirect, through managed funds it does not control, which may now include index funds holding the public companies named.

Anthropic is the developer of Claude, which is used in preparing this research, and Anthropic is reported to be the largest customer in the collateral pool of the facility read here, at about 40 percent of its committed contracts. That is the closest this nearness has come to the subject of a piece, and it cannot be checked away. What answers it is reproducibility rather than any assurance about the tool: every quotation carries its section or definition, every figure carries its accession or its REPORTED label, the derivations show their arithmetic, and the whole reading can be run again from the public record by anyone who opens the same documents.

Others named have ties to the filer or to each other. NVIDIA is a shareholder, the supplier of the machines the facility finances, and a customer under a master services agreement carried on the same exhibit index as the OpenAI, Microsoft and Meta agreements. Jane Street is a shareholder and a customer, and its contracts are part of this facility’s collateral. JPMorgan and MUFG are the arrangers and JPMorgan the administrative agent; Morgan Stanley held those seats on the May facility and is a bookrunner on this one; U.S. Bank is depositary and collateral agent. Moody’s, S&P and Fitch are rating agencies discussed here both as raters of this debt and as a paid participant in its waterfall.

Companies not named here, among them the lenders under the syndicate, the customers behind the redacted marks in the agreements, and the suppliers behind the concentration table, may hold positions or supply relationships that bear on the filer discussed, and that possibility is part of why every piece is re-checked for bias, ground facts, and filings rather than read against a fixed list. Figures are quoted from the filers without characterization, and the same standard of reading is applied to every party named.

NOTES

1. DDTL 5.5: CoreWeave, Inc. Form 8-K dated August 7, 2026, accession 0001769628-26-000357, Exhibit 10.1 (credit agreement, dated as of August 7, 2026), Exhibit 10.2 (parent guarantee), Exhibit 99.1 (press release, furnished). Quotations are verbatim and carry their section or definition.

2. DDTL 5.0: Form 8-K, accession 0001769628-26-000236, Exhibit 10.1, dated as of May 15, 2026. The source for the “Funding Date GPU Amount” definition, the plural “Fee Letters,” and the cover page’s agent and arranger list. The interval between the two agreements is eighty-four days. In “The Sum of the Participants” we gave it as eleven weeks; it is twelve, and the longer interval is the one that runs against our own reading.

3. Amendment No. 4 to the Revolving Credit and Guaranty Agreement, dated November 10, 2025, filed as Exhibit 10.7 to the Form 10-Q for the quarter ended September 30, 2025, accession 0001769628-25-000062. The exhibit is a blackline; the definition quoted is the amended text as marked. The original is in the agreement dated June 21, 2024, Exhibit 10.18 to the Form S-1.

4. Jane Street: Form 8-K Exhibit 99.1, April 15, 2026, for the investment and the commitment. Its agency ratings and its share of this facility’s committed contracts are REPORTED. The contract shares are Rod Dubitsky’s, from “CoreWeave’s Near Debt Debacle, Jane Street and Fitch’s Curiously Timed Jane Street Upgrade”, August 4, 2026.

5. NVIDIA: FY2025 Form 10-K, accession 0001769628-26-000104, Note 16 for the investment and Exhibit 10.38 for the Master Services Agreement dated April 10, 2023, first filed as Exhibit 10.31 to Form S-4/A, 333-289742, September 25, 2025.

6. Share counts and votes: the cover page of the Form 10-Q for the quarter ended June 30, 2026, accession 0001769628-26-000366, as of July 31, 2026, and that filing’s multi-class risk factor for the statement that the co-founders hold all of the Class B. That base is used rather than the FY2025 cover page because the April 2026 placement post-dates the January 31 count. Class B outstanding fell from 106,660,052 at January 31 to 92,664,912 at July 31, consistent with the charter’s conversion-on-transfer provision and with 14,727,337 of conversions reported on Forms 4 in the same window. Those two figures differ by 732,197 shares, which the forms do not reconcile; the likeliest residual is conversions dated inside the window but settling outside the two cover-page dates, and we state the difference rather than net it. Sale figures are our aggregation of every Form 4 filed under the issuer’s CIK from January 15 to August 15, 2026, restricted to code S dispositions with transaction dates between the two cover-page dates, with proceeds computed as shares times the reported price per share. The Rule 10b5-1 status is the checkbox on each form, with plan adoption dates from the forms’ own footnotes. The NVIDIA and Jane Street lines in Exhibit 3 are their announced purchases at the announced prices; no filing confirms continued holding, and any Class A held by the co-founders is not netted out of the last row, which understates the concentration rather than overstating it. Percentages are ours.

7. Operating and capital figures: FY2025 Form 10-K and the Form 10-Q for the quarter ended June 30, 2026, accession 0001769628-26-000366. The deferred revenue subtraction uses the cash flow statement’s line rather than the balance-sheet change, which avoids importing non-cash movement.

8. The discount at 97: Bloomberg, July 29, 2026, reporting marketing terms before the August 10 close. Carried throughout as marketed.

9. The nine-rung distance-from-cash ladder is from “Quality of Cash: Circular Financing and the AI Bubble”, August 1, 2026.

10. Rounding: our derivations round away from the argument. The capital program comparison is 6.9 weeks, given as about seven. The per-share hurdle and the share percentage are computed on the July 31, 2026 count of 551.5 million shares rather than the December 2025 count, which is the larger denominator and the smaller result. The dilution is 1.853 percent, given as 1.85 rather than rounded up.

Analysis: Cape Fear Advisors.

On August 3 we published four questions and wrote that the filings, not the call, would answer them. The filings have printed: the earnings release on August 11, the Form 10-Q the same evening. Here are the answers.

First, the frame. The filings present a company that burns cash, depends on debt, and grows fast: revenue of $2,575 million in the quarter, a net loss of $626 million, operating cash flow that covers net interest expense 1.06 times on our derivation, and $35.6 billion of total indebtedness with about $10.0 billion more committed and undrawn, per the 10-Q. The value story, in cash and not in relative terms, is that the company will ultimately return more than its cost of capital. By the company’s own math that point is years out, and progress toward it is measured against metrics the company defines: backlog, contracted power, and a family of adjusted measures. Two other statements in the same filing complicate the reading. The auditor’s critical audit matter sits on revenue recognition. And material weaknesses in internal control over financial reporting “continued to exist as of June 30, 2026.”

And then there is the $1,469 million that appeared under a new caption in one quarter and disappeared in the next. It lives at the intersection of those two complications. We will come to it.

The second question: what resolves the $1.3 billion?

In August we asked what would become of the amount reclassified out of deferred revenue in the first quarter: does it “earn into revenue as the machines deliver, settle as a note, a reversal, a return of cash, or a charge, resolve in a way no filing has shown, or simply stand another quarter?” The answer is a reversal. Of the outcomes on that list, it is the one the filings performed without narrating.

The record, in sequence. The FY2025 10-K carries no caption called “customer liabilities” anywhere; other current liabilities stood at $162 million, not itemized. The Q1 10-Q created the caption: its balance sheet components note showed customer liabilities of $1,469 million at March 31, and its revenue note narrated the arrival, a change in deferred revenue “primarily driven by reclassification of $1.3 billion to customer liabilities.” The Q2 10-Q does not contain the phrase at all. The line’s parent, other current liabilities, fell to $70 million, again not itemized. The caption was born on March 31 and was gone by June 30. Its entire life was lived in interim statements, which are reviewed, not audited.

The caption is gone. The money is not. The reclassifications are non-cash, so each period’s gap between the change in the deferred revenue balance and the deferred revenue line inside operating cash flows measures the non-cash movement. In the first quarter, the balance fell $662 million while the cash flow line contributed positive $575 million: a non-cash exit of about $1.2 billion, matching the narrated $1.3 billion within ordinary residuals. Across the first half, the balance rose $1,515 million against a cash flow line of $1,365 million: a non-cash movement of only $150 million for the six months combined. The subtraction is ours and it is short: the second quarter alone contains a non-cash movement of roughly $1.4 billion back into deferred revenue, against the $1,469 million balance that vanished. The money sits where it started, under the name it had before it left.

One precision point. These movements are non-cash; they inflated no quarter’s operating cash flow in either direction. The cash is real. The question was where the obligation stood, and under what name, and the answer is that it made a round trip through a caption that no longer exists. Out of deferred revenue with a sentence. Back without one.

The third question: does the largest customer’s share keep falling?

It fell again. The 10-Q reports the top three customers at approximately 36, 26, and 10 percent of quarterly revenue, against a single customer at 71 percent in the prior-year quarter. The 10-Q does not attach names to the percentages. It does attach commitments: an order form with OpenAI of approximately $6.5 billion running through May 2031, an order form with Meta of approximately $21.0 billion signed in March, and, per the release, more than $25 billion of net new commitments in early Q3 against a backlog of $104 billion. One sentence in the risk factors is new to us: “our current customers have contractually specified our use of NVIDIA GPUs.” Diversification of names, concentration of chain.

The first and fourth questions: which rung, and does cash arrive from outside?

In August we asked which rung the quarter’s solution would sit on, a new one or cash off an old one, and whether cash arrives from outside. The answers arrived together. Cash arrived, mostly borrowed, at a rising price, on a new rung. Financing activities supplied $14.0 billion in the first half, net: $16.7 billion of debt proceeds and $3.0 billion of private placement equity, less $5.2 billion of debt repayments and $0.5 billion for capped calls. On our derivation, $10.1 billion of that net arrived in the second quarter, against $3.7 billion of operating cash flow for the half. Most of that operating cash arrived on the first quarter’s working-capital tide, chiefly a receivables collection above $1 billion; the second quarter contributed $679 million. And first-half debt service, $5.2 billion of principal plus $982 million of interest per the 10-Q, exceeded the half’s operating cash flow by about $2.5 billion. The company’s new facility, DDTL 5.5, priced at SOFR plus 5.50, one hundred basis points wider than the facility signed in May at the same ratings from the same agencies, with the release naming what changed: earlier facilities were backed by customer contracts running through the debt’s maturity, and the new one carries a five-year term against three-year contracts. That facility is the new rung. On the equity side there was one outside check: Jane Street’s $1 billion investment in April, announced alongside a $6 billion commitment to purchase compute. We flagged that transaction in a correction to our August 11 piece, and it counts on both sides of this question, outside capital and a commercial commitment arriving in the same announcement.

The quality-of-cash ratio moved with the tank. Cash on hand against the annual capital program ran near one to fifteen at March 31 and near one to six at June 30. The improvement is real. It is not earned; it is borrowed. Operations contributed $679 million of the quarter’s roughly $10.8 billion of cash inflows, and the rest was raised. The dot grew because the tank was refilled, not because the machine filled it.

A fifth reading, ours: how is the future presented?

The August piece did not ask this one; the print put it in front of us. The CFO called the first quarter’s one percent adjusted operating margin “the trough.” The second quarter printed five percent. One data point is not a trend; the claim survived its first test. Adjusted EBITDA margin printed at 59 percent. Active power reached 1.5 gigawatts against roughly 3.7 contracted, both ends of the ratio disclosed, the only base-attached utilization figure we have seen among the filers we read.

Beside those numbers sit these. Adjusted operating income of $128 million against net interest expense of $640 million, a five-fold gap between the profitability measure management leads with and the single largest cash cost beneath it. The measures that improved are measures the company defines. The measures that did not improve require no definition: operating cash flow of $679 million in the quarter, on our derivation, against $640 million of net interest.

The $1.4 billion, taken as a whole

In July we wrote that the three statements, each correct under its own rules, present three different businesses, and that the reader who wants the one company must hold all three at once. The round trip is that piece’s second verse. It is a revenue recognition issue because deferred revenue is the waiting room for revenue, and the auditor’s flagged judgment is “the identification and treatment of contract terms that may impact the timing and amount of revenue recognized.” A billion and a half dollars left that waiting room, lived one quarter under a caption the filing left undefined, and returned. Every future quarter’s revenue draws from a balance this movement passed through twice.

It is an internal controls conundrum because the same 10-Q states that material weaknesses, in IT general controls, segregation of duties, and personnel, continued to exist at June 30. The round trip was narrated in one direction and silent in the other, in periods no auditor has audited, under a control environment management itself flags. The first audited statements that must carry this arc, and the first auditor attestation on the controls that produced it, arrive together in the FY2026 annual report. That is now the most important date on our calendar for this name.

Nothing in this piece identifies an error, an inconsistency with the accounting rules, or a bad actor. Interim reclassifications are permitted. Captions come and go. The subject is the asymmetry: a departure with a sentence, and a return the reader must derive.

The sum of the participants

A note on what this piece does not contain: a valuation. A discounted cash flow model run on the consolidated firm does not produce one. On the filed trajectory, the crossing point where returns on invested capital clear the cost of the capital sits beyond any horizon a model can defend. That is not a fault of the company, and it is not a fault of the method. It is the method reporting that the answer is not yet in the numbers.

What the numbers do contain is participants. A supplier that is also a shareholder. Lenders across seven named facilities since 2023. Customers holding order forms, and one holding stock. Landlords carrying $16.3 billion of operating lease liabilities. A public float. Each row’s cash terms are filed, and each row presents different bases, different time frames, and different remedies. The bull case and the bear case need the same addition: what each participant puts in, what each takes out, and where each return accrues. The equity’s line in that addition is computed last, after every senior row clears, and it stays out of reach until the assumed debt inflows, and the round trips beneath them, have run their course. Against the model’s time frame, those differences are the analysis. Some rows will complete their round trips entirely before the equity’s line is measurable at all. In the newest credit agreement, cash steps through a nine-rung waterfall, and distributions are the ninth rung.

And one round trip is not the journey. The facilities carry five-year terms against a question that answers in decades; in July we counted the refinancing requirement near four and a half turns before the machines pay for themselves and for the money that bought them. This year the turns repriced: SOFR plus 4.50 in May, SOFR plus 5.50 in August, one hundred basis points in eleven weeks, like for like. Each repricing raises what the machines must earn before anything reaches the ninth rung, and each one moves the equity’s answer date further out, with the price of the wait set again at every turn. Every participant carries that repricing, and none carries it alike. The lenders carry it with collateral, covenants, and remedies, repriced quarterly. The landlords carry it with leases. Some participants carry it with returns that accrue on their own ledgers, outside this firm’s. The float carries it with a mark. The round trip this quarter documented is what one turn looks like from the outside. The ledger of who rides them is a piece of its own.

What has to happen next, still

In July we counted twenty-seven years against a schedule that gives the machines six. In August we asked four questions. One is retired: the $1.3 billion resolved as a reversal. The other three renew each quarter, and two dates now hold them. The FY2026 10-K, where the round trip meets its first audit and the controls meet their first attestation. And roughly February 2027, when the newest facility’s debt service coverage test begins to run. Two clocks, as before. The financing clock is running, and repriced one hundred basis points heavier. The return clock is running, and this quarter it did not move.

Leave a comment

Standing disclosure

Nothing here is investment advice. Cape Fear Advisors holds no direct position, long or short, in CoreWeave, NVIDIA, or Meta Platforms; any exposure is indirect, through managed funds it does not control, which may now include index funds holding the public companies named. Anthropic is the developer of Claude, which is used in preparing this research, and CoreWeave has named Anthropic among its customers (Note 5); no claim in this piece rests on trust in the tool, since every figure names its filing, the derivations show their arithmetic, and the reading is reproducible from the public accessions. NVIDIA, discussed above as supplier and shareholder, is the maker of the GPUs the customer contracts specify and an investor in CoreWeave; Meta and OpenAI hold the order forms cited; Jane Street holds the investment and purchase commitment cited; each is discussed only in its filed role. Companies not named here, among them the lenders under the facilities and the customers behind the concentration percentages, may hold positions or supply relationships that bear on the filers discussed, and that possibility is part of why every piece is checked against ground facts and filings rather than a fixed list. Figures are quoted from the filers without characterization, and the same standard of reading is applied to every filer named.

NOTES

1. “CoreWeave: What Has to Happen Next,” Cape Fear Advisors, August 3, 2026. The four questions are from that piece as published; the second is quoted in full above, and where a section heading compresses the published wording, the published text controls. Also “CoreWeave, Twenty-Seven Years” (July 12, 2026) and “CoreWeave, Taken as a Whole” (July 2026).

2. “CoreWeave, Adding It Up: One Filing, Two Announcements,” Cape Fear Advisors, August 11, 2026, including the same-day dated correction regarding Jane Street’s April 15 investment.

3. CoreWeave Q2 2026 Form 10-Q (accession 0001769628-26-000366): balance sheet components note; revenue note; statements of cash flows and supplemental disclosures; risk factors; controls and procedures. Q1 2026 Form 10-Q (accession 0001769628-26-000222): Note 7 itemization of customer liabilities; the reclassification sentence in the revenue note. FY2025 Form 10-K (accession 0001769628-26-000104): absence of the caption; other current liabilities presentation.

4. All derivations (the quarter-alone cash flow decomposition, the non-cash round trip arithmetic, the coverage and quality-of-cash ratios) are ours from the filed figures and carry residuals under $100 million; the filed figures control wherever they differ. The seven named term facilities: DDTL 1.0 ($2.3 billion, July 2023); 2.0 ($7.6 billion, May 2024); 2.1 (an increase of $3.0 billion, September 2025); 3.0 ($2.6 billion, July 2025); 4.0 ($8.5 billion, March 2026); 5.0 ($3.1 billion, May 2026); 5.5 ($2.6 billion, August 2026), per the Q2 10-Q’s risk factors. The family numbering is not chronological; the dates are as filed, and the maturity schedule in the debt note (DDTL 3.0 due August 2030, DDTL 2.1 due March 2031) corroborates the order.

5. CoreWeave named Anthropic among its customers in its Q1 2026 earnings release (accession 0001769628-26-000220, May 7, 2026): “Signed multi-year agreement with Anthropic to support the development and deployment of Anthropic’s Claude family of AI models.”

Two disclosures before the exhibits, one about the company and one about us. About the company: nothing in this piece identifies an error, an inconsistency, or a bad actor. We did not find one, and we did not go looking for one. Every document here is taken on its face as accurate, as complete under the rules that govern it, and as consistent with the others; the release is responsive to all filing and disclosure requirements; and the omissions noted below are expressly permitted. The subject of this piece is not conduct. It is what each kind of document can carry, and what a reader can learn by laying them side by side. About us: the spine and the matching are choices of ours. It is possible our framework excludes something, and possible that language we selected from the agreements would read differently beside language we left out. We are not reconstructing these documents for any legal purpose. A reader who wants to check the choices can: every column names its exhibit, the quotations from the agreements carry their section numbers, and the whole accession is public.

One more fact belongs up front: the layers of an 8-K do not carry equal weight, and that is by rule, not by choice of ours. The press release is furnished under Item 7.01 and, in the filing’s own words, “shall not be deemed ‘filed’ for purposes of Section 18 of the Securities Exchange Act of 1934... or otherwise subject to the liabilities of that section.” The summary and the agreements are filed. The rules sort a company’s statements into tiers of liability, and the company sorted its statements accordingly. We are reading the sort. The distinction between the columns below is one the rules themselves draw.

Comment: The filing opens with the parties and the purpose. The release expands the purpose into platform expansion and customer support; the agreements specify the mechanics: loans capped at 70 percent of the capital expenditures financed, the parent’s invoices delivered to the lenders where the borrower buys from the parent, and proceeds distributable up to pay those invoices within seven business days. The parent sells the servers to its own subsidiary, the lenders fund 70 percent of the invoice, and the proceeds may return to the parent. Every customer name in the agreements is replaced by the mark “[*].”

Comment: Each layer describes a company, and they do not describe the same one. The release’s About section describes CoreWeave, Inc. The agreements describe the borrower, a limited purpose entity covenanted to be separate and independent, whose notice address is the parent’s suite in Livingston. The 8-K describes both in a single clause and contains no further description of either company’s business.

Comment: The dates are the same in every layer. The release compares the debt’s five years to the contracts’ three and reads the difference as lender confidence; the agreements set the order of cash: amortization beginning only after the draw period, on percentages in a schedule the filing omits, a sweep to the $1.3 billion floor, distributions ninth of nine, and deposits from any parent company accepted at any time. Cash exits the structure last and enters it freely.

Comment: A debt sale produces two numbers, a quantity and a price. All three layers carry the price; one layer characterizes the demand. For scale, one comparison from outside the accession, pinned in the Notes: the company’s facility of the same ratings priced in May at SOFR plus 4.50 percent. The marginal secured dollar costs 100 basis points more than it did three months ago.

Comment: The words “guarantee” and “guarantor” appear zero times in the release, which attributes the ratings to collateral and structural protections. The filed 8-K names the unconditional parent guarantee in its first sentence on the subject, and the guarantee itself is Exhibit 10.2, full recourse, primary obligor, capped only by fraudulent-conveyance and solvency law. The support is named in the layers that are filed.

Comment: The release does not address the covenants directly; its left column above is empty for that reason. The 8-K names the test and its start date, which puts the first check of the coverage ratio around February 2027, earlier if the commitments end before December. The agreements name the cure: equity deemed income for the test, without limit through a window that contains the first test, and in three of any four consecutive months thereafter.

Comment: The release does not address these either. The 8-K describes them as customary, and they are; the agreements state what customary resolves to here: covenants borrowed from the revolver, terminating on an investment-grade rating, with a cross-default confined to non-affiliate working-capital and revolving debt.

Comment: This row runs the other direction, and the empty cell above is the 8-K’s: the filed summary contains no sentence on renewal or re-leasing. The subject lives at the outer layers. The release, fairly, directs the reader to the document that contains the criteria, and the right column is what the reader who follows the pointer finds: a standard certified by the borrower’s own officer, a fallback approval by lender silence, and a backstop prepayment sized on the borrower’s own model. The Renewal Deadline itself is one of the marks: “[*].”

Comment: The filing closes by reading itself. Each layer holds exactly what its rules provide. The release is furnished and not deemed filed. The summary is qualified in its entirety by the exhibits. And the exhibits omit what the exhibit rules expressly permit, information that is not material and competitively sensitive, under the filing’s own index notes quoted above. The omissions are permitted and routine. Their effect, stated without any conclusion attached: the marks fall on the customer names, the data-center sites, the numerator of the 2.40:1.00 collateral test, the hourly GPU rate, and the renewal deadline, so the facility’s tests are public in their thresholds and not computable from the filing.

A few remainders, from the same accession. The credit agreement is signed for the borrower by its Treasurer; the guarantee is signed for the parent by its Chief Financial Officer; they are the same officer, wearing each entity’s hat in its own document. The borrower’s first standalone audited financial statements will cover fiscal 2027, deliverable up to 150 days after year end, and the audit opinion may be qualified for impending maturity or a covenant breach without breaching the reporting covenant. The monthly sweep that pays the loans down takes the cash that is there; if there is less than the full amount, the shortfall itself is not a default, a standard feature of waterfall structures, which look to the collateral accounts rather than to a promise of sufficiency. And the release counts “more than $30 billion of debt and equity capital year-to-date”; the composition, as we count it from the company’s announcements, is one equity placement in January and debt since, itemized in the Notes.

The same facility, described three times, by one company, on one day, each register following its own rules and each consistent with the others. The Forge House read one set of numbers wearing two accounting forms; this filing is one transaction wearing three genres. The form decides what the reader sees, and every form here was chosen and permitted. For a reader who wants the repeatable version: the 8-K first, then its exhibits, and the release last. The announcement is the whole accession. The press release is one exhibit of it.

Standing disclosure

Cape Fear Advisors holds no direct position, long or short, in CoreWeave or NVIDIA; any exposure is indirect, through managed funds it does not control, which may now include index funds holding the public companies named. Anthropic is the developer of Claude, which is used in preparing this research; no claim in this piece rests on trust in the tool, since every column names its exhibit, the quotations from the agreements carry their section numbers, and the reading is reproducible from the public accession. NVIDIA, named in the release quoted here, is the maker of the GPUs that run in the servers the facility finances, and is an investor in CoreWeave (Note 2). The customer names in the agreements are redacted; whether Anthropic, or any counterparty of Anthropic’s, stands behind one of the “[*]” marks is not knowable from the filing, and we state that unknown rather than assume it away. Companies not named here, among them the banks arranging the facility and the customers behind its contracts, may hold positions or supply relationships that bear on the filers discussed, and that possibility is part of why every piece is re-checked for bias, ground facts, and filings rather than read against a fixed list. Figures are quoted from the filers without characterization, and the same standard of reading is applied to every filer named.

NOTES

1. The May comparison: CoreWeave’s DDTL 5.0, $3.1 billion, priced at SOFR plus 4.50 percent, rated Ba2 by Moody’s and BB+ by Fitch, per the company’s release of May 18, 2026 (Form 8-K, accession 0001769628-26-000236, Exhibit 99.1). DDTL 5.5 carries the same ratings from the same agencies at SOFR plus 5.50 percent; the difference is 100 basis points.

2. The year's raises, as announced by the company: an equity placement of approximately $2.0 billion to NVIDIA, priced January 23 (reported in the Form 10-Q for the quarter ended March 31, 2026, accession 0001769628-26-000222); a $1 billion equity investment from Jane Street, announced April 15 alongside a $6 billion cloud agreement (accession 0001769628-26-000167); and, among others, an $8.5 billion delayed draw facility announced in March (accession 0001769628-26-000129), the $3.1 billion DDTL 5.0 in May, the $2.6 billion DDTL 5.5 closed August 10, and note offerings reported in the spring. The list is not exhaustive; the equity in the release's "$30 billion of debt and equity" phrase is, by our count, the January placement and the April investment. Correction, August 11, evening: an earlier version of this note stated that we found no equity announcement after January 23. The April 15 Jane Street investment is one, and the company's second-quarter release, published this evening, lists it among the quarter's highlights. The error was ours and was corrected the same day.

3. The first-test date is derived from the 8-K’s covenant description: testing begins the first full calendar month after the earlier of the commitments reaching zero and December 31, 2026.

4. The agreements are dated Friday, August 7, 2026; the press release is dated Monday, August 10, 2026; both travel in the accession cited below.

Source: CoreWeave, Inc. Form 8-K dated August 7, 2026, accession 0001769628-26-000357: Exhibit 99.1 (press release, furnished); Item 1.01 and Exhibits 10.1 and 10.2 (filed). All quotes verbatim.

Analysis: Cape Fear Advisors.

The house and the bubble

Set side by side, the two ends of the circle tell the story before any number does. Measured as cash against each company’s own annual capital program, NVIDIA is a house: roughly fifty billion dollars of cash and marketable debt securities against about seven billion of capital spending, close to seven to one, a disk of cash with only a small hole. CoreWeave is the bubble: about two billion of cash against a build running near thirty billion a year, about one to fourteen, a shell of commitment around a dot of cash. And the dot is mostly the seller’s money: about two billion of CoreWeave’s roughly two-and-a-quarter billion of quarter-end cash came in this quarter as NVIDIA’s placement, and even with that infusion the cash fell about eight hundred million over the quarter while the committed build ran on. It is the same circle, filled from opposite sides, the buyer’s small reserve of cash largely supplied from the seller’s large one. The seller collects; the buyer commits.

The house is unusual for a maker of physical things. NVIDIA carries a software company’s balance sheet because it does not build the fabs; it designs and sells, and the cash arrives at the register within about forty-five days. Its forward commitment is real but it does not sit in capital expenditure. It sits in roughly a hundred and fifty-five billion dollars of purchase commitments, mostly manufacturing, supply, and capacity, an instrument backed by the order book rather than by a plant, which is why the donut reads as full. The bubble is the ordinary condition of a company that must pour concrete and rack chips today against revenue that arrives over years, and whose cash today does not cover the year’s build.

The trade and the stake

On January 23, 2026, NVIDIA bought about two billion dollars of newly issued CoreWeave stock, at $87.20 a share, and it also sells CoreWeave the chips CoreWeave racks. That is the circle in miniature: the seller funds the buyer, and the buyer spends the funds back with the seller. It is not hidden and it is not improper. An equity investment measured at fair value is recorded as an investment; that is what the guidance requires. The question this piece asks is narrower and, we think, more useful: the same two billion of support could have been given in more than one ordinary form, and the form chosen decides what the reader sees.

We know, from “The Fourth House”, that no price concession in fact occurred: NVIDIA’s gross margin held near seventy-one to seventy-five percent straight through the placement, so the equity was support, not a discount. That is what makes the following a clean counterfactual rather than a guess. Re-expressing the equity as the price concession it was not is simply a way to see what the structure that was used accomplishes, and where it places what a concession would have shown. The exercise is deliberately inconsistent with the accounting that was correctly followed. It changes one thing, and it changes no cash.

The lever

With both companies’ actual first-quarter cash flow statements laid side by side in their three layers, the two billion moves from an investment to a price concession. For NVIDIA it slides from the investing section up into revenue: revenue, operating income, and operating cash flow each fall two billion, while the investing outflow falls by the same two billion. Net change in cash: unchanged, at $2.6 billion. For CoreWeave it slides from financing into a smaller build: no equity is issued, and capital spending is two billion lower. Net change in cash: unchanged, at negative $0.8 billion.

What does not move is the cash, and the sources and uses beneath it. What moves is the classification within the cash flow statement, the operating income, and the balance sheet. As reported, NVIDIA holds a two-billion-dollar investment in CoreWeave, and CoreWeave carries two billion more in property and in equity. Under the concession, none of those appear. The choice of form decides which side of the operating line the two billion lands on: an investment sits below it, where a mark can even lift net income, while a concession sits above it, as a cost. Same dollar, opposite side of the line.

The round trip

The round trip is the second number, and it is sharper. Because CoreWeave spends the money on NVIDIA’s own product, at roughly a seventy-five percent margin, about three-quarters of the two billion, near $1.5 billion, returns to NVIDIA as gross profit in the same quarter. The investment’s effective cash cost is therefore about half a billion dollars, and reported operating income carries about $1.5 billion of gross profit from that funded demand. This is an estimate: the specific NVIDIA sales to CoreWeave are not disclosed, so it rests on the funds being spent on NVIDIA product at about the company margin, within the period. There are two numbers here, not one, and they are the same two billion doing two different things. As the equity deal, the chips are sold and funded: NVIDIA books about one and a half billion of gross profit on them and holds the stake as an asset. As a concession, the same chips are given away: NVIDIA books no revenue and carries their cost of about half a billion, so the exchange subtracts half a billion from gross profit instead. The dollar is identical; the reported gross profit swings the full two billion between the two, from half a billion given up to one and a half billion earned. The two billion is the same thing, and it is not the same thing. Stated that way, it is the circle completed: the funder funds the customer, the customer buys the funder’s high-margin product, and most of the outlay comes home as income the same quarter, while the outlay itself sits on the balance sheet as an asset rather than in the income statement as a cost.

The round trip is the part of this the market has largely reached on its own. The rest of this piece is where the same structure choice travels next, into the segment line, and into the rate a reader takes off it.

The segment

Circular financing at scale does more than move numbers between lines. There is now a place where they sit. With its May results NVIDIA recast how it reports Data Center, splitting the segment into two sub-markets for the first time, and the neocloud builders, the very companies this financing stood up, went into the one called ACIE, for AI Clouds, Industrial, and Enterprise. Management put ACIE at thirty-seven billion dollars for the quarter, up about thirty-one percent from the prior one; the hyperscaler bucket grew twelve percent in the same quarter, so the financed builders sit in the faster-growing half. Netted against that bucket, the concession slows its reported growth to about twenty-four percent quarter on quarter, and its margin, on the company rate applied for want of a disclosed one, eases about a point and a half. The change is immaterial to the company as a whole and concentrated inside the segment.

The split came on the earnings call and in a nine-quarter table posted to NVIDIA’s site, not in the filed segment note, whose reportable segments did not change. The new framework adds detail at the sub-market level, and removes it where it bears on this reading: the AI clouds NVIDIA finances are now folded into ACIE beside industrial and enterprise buyers, so their own figure cannot be read off the page. AI cloud revenue more than tripled year on year, management said, but as an aside, not as a line. It does not change the numbers. It changes what can be counted. That is the quiet cost of the circle growing large enough to name itself: the place where the demand now sits is also the place where the financed part of it is hardest to trace.

The ladder

None of these forms is exotic. They are rungs on a ladder that runs from cash on the barrel, the only rung that is not a circle at all, up through trade credit, supplier finance, customer prepayment and component intermediation, vendor lending, equity in the counterparty, contingent support, and finally the wrapped structures at the top. Distance up the ladder is distance the cash must travel to come home. The structures in this relationship sit on named rungs of it: vendor financing at six, the equity stake at seven, the capacity backstop at eight. A price concession is the near-cash bottom, barely a circle. The law of the ladder holds here as everywhere: the cash is the same at every rung; climbing it moves the same support to a greater remove from the income statement, and the danger is not the height but the rung where the cash inside the loop is gone, replaced by the loop’s own paper.

The rung also decides what the buyer is left holding. Equity put cash on CoreWeave’s balance sheet; the price concession at the bottom would have been the same support given as a lower price, and would have left no cash there at all. The seller’s choice of form is, among other things, a choice about whether the buyer holds a reserve or a discount.

And the choice reaches past the current quarter. Bought at a two-billion discount, CoreWeave’s hardware would carry two billion less on the balance sheet, and so about two billion less depreciation across its life, lifting the buyer’s reported income in every period of the hardware’s life. The equity route leaves the asset at full cost and sets two billion of fresh equity beside it. So the seller’s choice of form does not only move the seller’s quarter; it sets the buyer’s reported earnings for years, the longest-lived instance in the exercise.

The second derivative

The reason to read this before the next print is that the choice of rung moves the one figure the careful observer relies on. The three readings, in order. The level is the segment’s revenue, about thirty-seven billion dollars for the quarter. The first derivative is how fast that level is growing, the roughly thirty-one percent quarter on quarter. The second derivative is how that growth rate is itself changing from one quarter to the next, whether the segment is speeding up or slowing down. Those watching for a change in the weather do not watch the level; they watch the second derivative, because that is where a turn shows first.

This exercise moves the first derivative directly. The same two billion, an equity deal rather than a concession, is the difference between the segment printing thirty-one percent and twenty-four percent, a seven-point gap in the growth rate. And because the second derivative is built on the first, the seven points pass straight into it: a quarter that would read as accelerating can read as accelerating faster, or a slowing one as holding, on the structure alone. The level barely moves, about five percent on the segment and immaterial to the company. The signal moves seven points. A reader tracking the turn has first to strip the structure choices out, or the number read was, in part, written by the structure. And the echo returns a quarter later, when the lower base reads as faster growth: the structure moves the signal both ways before it settles.

What this is

This is a hypothetical, run because the debate around circular financing made it a question to work through on paper, and for no other reason. The accounting and the presentation as filed are, in our view, accurate and required. We change one thing, the form of a single two-billion-dollar arrangement, and we grant that the change would be inconsistent with the guidelines that were followed. We draw no conclusion from it, and we leave the numbers to do the work.

The corners we rounded, stated rather than buried: the $1.5 billion round trip assumes the funds bought NVIDIA product at about the company margin within the period, which is not separately disclosed; the concession’s effect on operating income assumes a price reduction on volume that would have sold anyway; the segment margin is the company rate applied, since none is disclosed at the segment level; and the two-billion figure is measured before tax and against each company’s own totals. None of these changes the finding. No cash moves, and the classification, the operating income, and the balance sheet move with the rung.

Share

Standing disclosure

Cape Fear Advisors holds no direct position, long or short, in the securities discussed here. Any exposure is indirect, through managed funds it does not control, which may now include index funds holding the public companies named, among them NVIDIA and CoreWeave. Anthropic is the developer of Claude, which is used in preparing this research, and Anthropic is also a compute counterparty within the same circle of arrangements read here. That nearness cannot be fully checked away, which is why no claim here rests on trust in the tool: every figure carries a public source, and the record grades the rest. Others named have ties to Anthropic, among them Amazon and Alphabet, its two largest outside backers; and companies not named here may hold positions or supply relationships that bear on the filers discussed, which is why every piece is re-checked for bias, ground facts, and filings rather than read against a fixed list. Figures are quoted from the filers and from named parties without characterization, and the same standard of reading is applied to every party named.

Analysis: Cape Fear Advisors.

NOTES

Figures are filed or reported as tagged, and given as ranges rather than points where the underlying figure is an estimate.

The quality-of-cash exhibit is drawn on a near-term basis, cash against each company’s own annual capital program. NVIDIA cash and marketable debt securities are about fifty billion dollars, its marked equity holdings excluded; being fabless, its forward commitment sits mostly in about a hundred and fifty-five billion of purchase commitments, mostly manufacturing, supply, and capacity, a different instrument from capital expenditure. CoreWeave cash is about $2.2 billion; its annualized first-quarter capital spending, near thirty-one billion, sits at the low end of management’s own thirty-one to thirty-five billion guidance for 2026, so the build rate is not an artifact of one quarter. The ratio drawn as one to fourteen is about 13.7.

The two-billion-dollar investment is filed: NVIDIA purchased 22,935,780 CoreWeave Class A shares at $87.20 on January 23, 2026. It sits within NVIDIA’s Q1 FY2027 investing activities, inside the $18.6 billion of purchases of non-marketable securities, and appears in CoreWeave’s Q1 2026 financing activities as a $1.985 billion private placement, net of issuance costs, the fifteen million between the two being issuance costs.

NVIDIA’s gains from equity securities for the quarter were $15.9 billion, net, added back in operating cash flow as a non-cash item. About $13.4 billion of that is the period’s unrealized gain on publicly held positions still held at quarter-end, the figure carried in “The Fourth House”; the balance, about $2.5 billion, is the rest of the period’s equity-security gains. The non-marketable holdings carry $5.3 billion of gross unrealized gains, but that is a cumulative, life-to-date balance, not part of this quarter’s total. The companion figures are the same line read at two levels, not two in tension.

The ladder exhibit is reproduced from our framework piece, “Quality of Cash: Circular Financing and the AI Bubble”, where its per-rung instances are anchored to their filings. Here it stands as the taxonomy; the placement of the structures in this relationship onto rungs six, seven, and eight is made in the text above.

Sources: NVIDIA Form 10-Q, Q1 FY2027 (ended April 26, 2026), accession 0001045810-26-000052. CoreWeave Form 10-Q, Q1 2026 (ended March 31, 2026), accession 0001769628-26-000222. NVIDIA Q1 FY2027 results and segment commentary; NVIDIA and CoreWeave releases on the January 26, 2026 investment. Standards referenced: ASC 606, 321, 230, 280, 450/440, 850. Companion reading: “The Fourth House”, and, on the buyer’s side, “CoreWeave: What Has to Happen Next”.

OpenAI, to scale

The measure we lead with is a ratio: cash to committed uses. For OpenAI, cash is the money we can see, an estimate in the range of $50 to $90 billion; committed uses are what the company has agreed to spend, about $1.4 trillion as announced. The ratio sits near one to twenty, about a nickel of cash for every dollar committed. On the page the circle is the commitment and the dot at its center is the cash, and the dot looks like a speck, just one twentieth of the circle. Two further headlines sit beside it, the round’s mark of about $852 billion (reported) and the stakes OpenAI holds and carries at a mark, about $50 billion. They are drawn to scale, and they are not added to the commitment, because they are different kinds of number. The picture looks more like a bubble than a donut. What that means is the rest of the work.

How we got there

The ratio needs a structure behind it, and we borrowed one that readers already know. We took the shapes of income and cash flow statements and put the cap table down its spine (1), one player to a line, and filled each line from a filed or announced relationship. Where a figure is filed we say filed; where it is announced or reported we say so. Each dollar is counted once. Cash is carried as a range, because OpenAI is private and we can see its cash only through its counterparties and its own statements, so we place the dot at the midpoint. The statement does not foot, by design; the two columns, cash on one side and headline on the other, are the finding.

What the statement shows

The ledger reads by its two columns: cash on one side, headline on the other, and the distance between them is the finding.

The structure is circular financing, and circular financing is common; money moving in a loop earns no grade by itself. Inside the loop, though, there are full circles, closed relationships where cash that leaves as investment returns as revenue. The picture we lead with borrows a word from those loops. A full circle in the quality of cash shows a business; a full circle of financing shows only the loop, and a loop can turn without value forming inside it. The question the statement is built to answer is where, inside these loops, the cash and the value are made, and where a dollar only passes through, counted again at each stop. We avoid that concern on the loops by taking each commitment once and letting the statement refuse to foot.

The dot is drawn as one substance, but cash has qualities, and we have counted them before: cash owed and waiting on a solvent buyer, cash that needs years of a machine earning before it arrives, cash that holds only while a mark holds its price, and the fee a croupier lifts at settlement while the money is still in motion (The Croupier Counts First). Which quality the build would throw off is the whole of the catch-up question. A nickel today rises toward a dollar only if the cash the commitments produce is the kind that holds, and the machine-dependent kind, the neoclouds’ kind, needs years of earning to become cash at all.

Some of what leaves is hard cash, owed regardless of the loop and paid down the rungs of the stack: payroll, and the overhead that keeps the people working, the office and lab leases, the power that stays on in the research buildings. Part of what reaches OpenAI as cash is itself borrowed, raised by a partner against its own stake, so even the cash is not quite cash. Payment terms in the system can run long, a lever the largest partners hold. And Microsoft is the worked example of the line the whole reading turns on: the transition from compute provided in kind to cash that has to be paid, as credits run down and the bill comes in money. On figures reported by The Information (2), the $24.1 billion Microsoft recorded from OpenAI in its fiscal 2026 (filed) (3) resolves to roughly $2.0 to $2.5 billion of revenue share, capped since April, and the balance Azure compute, and with a $6.0 billion receivable outstanding (filed) and part of the rest settled in prepaid credits, the cash Microsoft collected runs below the revenue it recorded.

The players themselves show the shape. Microsoft appears twice, as a funder and as a vendor, which is why it is set apart on the statement. The commitment Microsoft did not carry forward now sits with Oracle. NVIDIA appears on three lines at once, as a supplier, as an owner on the cap table, and, as reported, in talks to guarantee about $250 billion behind a partner’s build. CoreWeave is the pure play whose customers are the labs themselves, Microsoft among them, so its cash in is the circle. A handful of names, Microsoft, Oracle, NVIDIA, SoftBank, and CoreWeave, carry most of both columns, which is why the cap table works as the spine. Most of what comes in is not cash either, the announced round weighted toward marks, in-kind contributions, and guarantees, with a thin slice of real money.

What it leaves open

The first open question is revenue: whether it arrives to fill the dot. Reported annual recurring revenue is a small figure against the commitment, and the dot can fill in two ways, from a public listing that turns paper into cash, with a figure near $1 trillion reported for an eventual offering, or from a revenue ramp that turns bookings into collections. The statement shows the gap and the doors out; it does not say which one opens.

The financing lines carry their own questions. Microsoft’s transition is the largest: how much real cash OpenAI needs as compute credits run down, and how much of that Microsoft still extends through terms. SoftBank’s cash comes with a string, borrowed against the OpenAI stake, so the question is not only how much reaches the company but whose money it is if the mark falls and the loan is called. Oracle’s question is whether it has taken on more than it can hold, carrying a large capital program on negative free cash flow, at a rating one notch above high yield, both reported (4), while it books the commitment Microsoft passed on.

The last question is who rides on the marks. OpenAI’s valuation sits as an asset on other balance sheets, held by funds, by a partner under the equity method, and by employees, and a compression of the round’s mark travels to each of them. Beneath that sits the structure question, the Foundation and the for-profit arrangement, which decides who is owed first if the marks give way. We sit in that circle too: our work relies on assistance from Anthropic, one of the parties in it, and we do not think it possible to use an AI platform fully outside the circle and unbiased by participation, so we run the work forward and backward, looking for mistakes or possible bias.

A donut and a bubble

A single reading is a number without a scale, so we ran the same ratio on a company at the other end of the picture.

Apple holds about five dollars of cash for every dollar it has committed, a ratio near five to one. Its cash is the body of the circle and its commitment a small hole: a donut. OpenAI is the inversion, its commitment the shell and its cash just a speck floating in space: a bubble. Dark is cash and putty is committed uses in both, so only the proportion flips. The two committed figures differ in kind, and we say so: Apple’s is filed and contracted, from its Form 10-Q for the quarter ended June 27, 2026 (5); OpenAI’s is announced. Scored on contracted paper alone, on the Oracle cloud agreement of about $300 billion, the Azure purchase commitment of about $250 billion, and the CoreWeave contract of about $22 billion, all reported, OpenAI’s commitment falls to roughly $570 billion and the ratio rises to about twelve cents per dollar, a dime by rounding, on a contract list that is not complete. The distance between the two survives the change of basis.

We did not pick these two because they are the outer bounds. They sit near the center of the AI discussion, seen from two sides, and the comparison makes both of the questions people are already asking legible at once. A reader who looks at OpenAI and says bubble is easy to understand; so is a reader who looks at Apple, a company with quality of cash to spend and little build of its own, and asks what its AI plan is. Extreme is not the same as unsound. A company mid-build runs a low ratio as a matter of course, and the reading asks less whether the number is small now than whether it rises as the build lands. A low ratio does not settle the matter on its own, though, because losses re-sharpen the point: a profitable company mid-build would let its own earnings catch the cash up, while OpenAI’s reported losses, on the order of $14 to $21 billion a year, keep the point sharp.

Whether to run it again

This is not a market call. We take no view on the shares, and nothing here signals a decision to buy or sell. It is a tool for reading a company through its cash and its connections: where the money comes from, where it is owed, and where it would have to arrive for the picture to fill. A map of the circle names the players; this reads what runs between them, and it holds the number as a dial rather than a headline, one that can be checked again next quarter. The two doors, a public listing and a revenue ramp, are where a change would show first.

What would move the reading is specific. The dot rises if the offering lands and real cash comes in, or if a revenue ramp collects at a margin that clears the cost of the capital already committed. It does not rise on another announcement, another mark, or another commitment, however large. Those are the conditions we watch for, and their absence is a reading too.

Taken that way, a bubble becomes a question asked of one company at a time, in the place where the cash either arrives or does not. The prototype leaves a question about itself, whether the reading is useful enough to run across the ecosystem, and our hypothesis is that it is: a single ratio, taken on a common basis and watched over time, stratifies the players by how much of what they have committed their cash can reach, and by whether that reach is widening or narrowing as the build proceeds. So the next readings are the ones more likely in the middle of the spectrum, or new edge cases to bring in. Companies in transition tend to land in the middle as well, their dependencies and business models already diverging from what they were even six months ago, and Oracle, a software company becoming an infrastructure one, is the case we mean. We will take the same picture to them, on the same basis, and watch it move.

The full ledger

The ledger on one sheet, at legal-ledger size. It does not foot, by design.

Related, on the Quality of Cash shelf: The Croupier Counts First; The $235 Billion Cash Gap; CoreWeave: What Has To Happen Next; CoreWeave, Twenty-Seven Years.

Standing disclosure

Cape Fear Advisors holds no direct position, long or short, in the securities discussed here. Any exposure is indirect, through managed funds it does not control, which may now include index funds holding the public companies named. Anthropic is the developer of Claude, which is used in preparing this research, and Anthropic is also a counterparty inside the circle read here. That nearness cannot be fully checked away, which is why no claim here rests on trust in the tool: every figure carries a public source, and the record grades the rest. Others named have ties to Anthropic, among them Amazon and Alphabet, its two largest outside backers and rows on this ledger; and companies not named here may hold positions or supply relationships that bear on the filers discussed, which is why every piece is re-checked for bias, ground facts, and filings rather than read against a fixed list. Figures are quoted from the filers and from named parties without characterization, and the same standard of reading is applied to every party named.

NOTES

Figures are filed or reported as tagged, and given as ranges rather than points.

(1) The cap table and the roster of counterparties are drawn from OpenAI’s cap table as reported from leaked materials (NYT DealBook; The Information). A private company files no cap table; this is the market’s information position, not a claim of completeness.

(2) OpenAI revenue used for the revenue-share estimate is reported by The Information, reproduced with the inference and revenue-share figures at Where’s Your Ed At; first-quarter 2026 revenue about $5.7 billion. The arrangement changed in October 2025 and again in April 2026, when Microsoft stopped paying OpenAI its share while OpenAI continued paying Microsoft, subject to a cap (CNBC); the effective date is not disclosed, and the agreement, if filed after an offering, is likely to be heavily redacted. The $2.0 to $2.5 billion is our recalculation, on a June-fiscal-year basis, of about a fifth of a revenue-share-eligible base; the balance of the $24.1 billion is Azure compute.

(3) Microsoft’s $24.1 billion of fiscal 2026 revenue from OpenAI and the $6.0 billion receivable are filed (Form 10-K, fiscal year ended June 30, 2026; SEC EDGAR accession 0001193125-26-323660), which also records $13.0 billion of funding commitments, $11.9 billion funded. The disclosure and the surrounding accounting changes are examined in Olga Usvyatsky and Francine McKenna, “Microsoft’s 2026 Annual Report: What’s going on with all the accounting changes?” (Deep Quarry).

(4) Oracle free cash flow negative about $23.7 billion in fiscal 2026, and a BBB- rating, one notch above high yield (S&P Global Ratings, July 9, 2026); reported, from Oracle’s fiscal 2026 results and the S&P action.

(5) Apple cash and marketable securities about $147 billion and purchase obligations over one year about $28 billion, from Apple’s Form 10-Q for the quarter ended June 27, 2026 (filed; SEC EDGAR accession 0000320193-26-000020); credit ratings reported (Aaa, Moody’s; AA+, S&P). Apple’s twelve-month component-supply obligations are working capital, excluded from the multi-year committed figure; adding even a large within-year supply number leaves the ratio comfortably above one to one.

(6) OpenAI committed uses about $1.4 trillion and annual recurring revenue about $20 billion, stated by Sam Altman on November 6, 2025 (CNBC; TechCrunch); annual losses about $14 to $21 billion, reported. The round’s mark about $852 billion follows the roughly $122 billion round reported ahead of an offering (reported); an eventual offering near $1 trillion, reported.

(7) Contracted commitments used in the change of basis: the Oracle cloud agreement about $300 billion and the CoreWeave contract about $22 billion, reported; the Azure purchase commitment about $250 billion, reported at the October 2025 recapitalization and not itself in Microsoft’s 10-K. NVIDIA is reported to be in talks to guarantee about $250 billion of OpenAI-linked data-center obligations (reported), separate from its announced staged investment and ten-gigawatt systems partnership.

(8) OpenAI cash about $50 to $90 billion is our estimate, built from disclosed and reported funding net of in-kind, contingent, and borrowed components, with the dot placed at the midpoint.

(9) Dated tests ahead, as of writing: NVIDIA’s next quarterly filing in late August, CoreWeave’s print on August 11, SoftBank’s October tranche, and Microsoft’s next quarter. The reading updates as each lands.

Analysis: Cape Fear Advisors.

The bubble and the razor.

Quality of cash is a way of reading a company: what is cash, what only looks like it, and whether the two balance or resolve. It reads one company at a time, though the answers are often in other companies’ filings, and it does not call an industry a bubble or a stock a buy. The position of this series is plain: if there is an AI bubble, it will show itself company by company, filing by filing, and not as a weather system over the whole sky. CoreWeave, the largest of the pure-play AI clouds, is where the reading is clearest, because nothing else in the frame softens what shows. This piece is written before its second-quarter numbers arrive.

This is a reading, not a verdict. Quality of earnings taught a generation to separate reported profit from the cash behind it; quality of cash sits next to that skill and asks, of the cash itself, how much of what a company records as cash is cash, and how much is a claim, a prepayment, a mark, or a balance still deciding what it is. It takes no relative view and puts no value on the shares. It runs on one principle: cash is the currency that has to balance or resolve in the end.

Most of what it reads is ordinary. Circular financing, one party funding the customer that buys from it, is as common as the day is long, and only rarely is it the extraordinary kind that earns alarm. Whether all of this is a bubble is a label history assigns later and no help now; the disservice would be the blanket judgment, emptying the bushel for one bruised apple, or calling a whole technology hollow because some of its financing runs in a circle. A thing can be overheated and a real advance at once, and some companies will come out of it stronger. So this is one company’s reading, and not a call on the rest. A bubble, if the word applies to anything here, gathers across an ecosystem, and a tip belongs to the leading edge. CoreWeave presents itself as that edge, “The Essential Cloud for AI,” the thing that “sits between the models and the silicon.” That is the reason to read it first.

Something is happening

Something is happening at CoreWeave that repays a slow read, none of it fraud, all of it filed.

It has never made a profit, and it asks that the reader look past that. In 2025 it earned $5.1 billion of revenue and lost $1.2 billion; the quarter that opened this year lost $740 million more. The company leads its own report not with those figures but with a $99.4 billion backlog and an adjusted EBITDA of $1,157 million at a 56 percent margin. That margin is an adjusted figure: it is what remains after depreciation and interest are added back, which is to say after the cost of the machines and the cost of the money are removed. Those two costs are the whole question, and the headline is defined to set them aside. One rung down, on the measure that keeps depreciation, the margin is one percent, and the chief financial officer has called one percent the trough.

Two companies, and the auditor said so

Set beside each other, the balance sheet and the income statement show two companies. One is a $99 billion promise; the other is a loss. That is the reader’s pair. The auditor names a different one, the accounting question underneath it: not promise against loss, but service against lessor. Deloitte and Touche named, as a critical audit matter, the judgment of “whether the contracts with customers are accounted for as a revenue contract for cloud-based services or a lease contract for cloud computing equipment.” A critical audit matter is not a criticism and not a ranking of risk; it is the auditor marking the judgment that was hardest to make and commending it to whoever holds the statements. The hardest judgment, by the auditor’s own account, was which company this is: a service that earns revenue, or a lessor of equipment. The company concluded service, because “generally either there are no identified assets or customers do not control or direct the use of underlying hardware,” the test the standard sets. A consequence of the service conclusion, and not its motive, is that the machines sit inside depreciation rather than out front in a lease.

The same matter names “the identification and treatment of contract terms that may impact the timing and amount of revenue recognized,” and that is where a number to watch appears. In the first quarter, $1.3 billion left deferred revenue, the caption that resolves into revenue, and entered a new caption, customer liabilities, that the filing does not define. The balance went from $137 million to $1,469 million in three months, most of it drawn from the long-dated portion into a current one, so that money once earned across years is now due within twelve. A caption with no definition can be read only from what it is not, and from the one thing that names it: the line’s own tag calls it a customer prepayment, in other current liabilities. So it is the customer’s cash, paid in advance, moved out of the prepayment account that becomes revenue and into a current one that does not. How and when it resolves, the filing does not say. That balance is a treatment of contract terms affecting the timing and amount of revenue, which is to say it sits inside the very judgment the auditor flagged as the hardest in the audit.

Outside the four corners

On July 30, the chief executive of Microsoft posted a return-on-invested-capital demonstration built on Morgan Stanley’s analysis, the work of Brian Nowak, of four companies, Microsoft, Alphabet, Amazon, and Meta, that together earn about 29.7 percent on capital against a 9 percent hurdle. It is a table, and it was laid by the chief executive of CoreWeave’s largest customer. He can sit at it, because the 29.7 is a blend and the blend is cover: a soft quarter in the buildout folds into search and software margins and never reaches the consolidated line. He can also sit out the pure-play version of the game entirely, because he does not have to play it. CoreWeave has no such choice. It is the blend pulled apart, with no legacy business to fold a miss into and a cost of debt, 9.1 percent on average and 9.625 on its June unsecured notes, already above the hurdle before a dollar of equity return, so its number arrives without cover. The four corners state the rule; the company-specific arithmetic the table only implies, that a dollar of these machines cannot return its cash and pay for the money that bought it inside the machine’s life, this series has already walked.

And the largest customer has not grown with the supplier the test leaves standing, though not in the way a glance suggests. Its share of CoreWeave’s revenue held above seventy percent for two quarters, 72 percent in the first quarter of 2025 and 71 in the second, before falling to 45 by the first quarter of this year, which reads at a glance as Microsoft buying less. The levels say the opposite. On figures derived from the filings, Microsoft’s dollars rose about a third, from roughly $707 million in the first quarter of 2025 to about $935 million a year later, while CoreWeave’s revenue more than doubled. Microsoft grew; the company around it grew faster, and the marginal dollar that diluted Microsoft came from other customers, among them OpenAI, which committed in 2025 and holds CoreWeave’s stock. The concentration did not fall away from the circle. It rotated deeper into it.

And the same week, the same customer’s chief financial officer named the freedom that lets it. Amy Hood, asked how Microsoft manages the risk of overbuilding, told investors the company’s capital spending had pivoted toward what she called “short-lived assets,” which “that’s CPUs and GPUs that have relatively shorter lead times,” and that if demand changes, “you can just slow down what is, in fact, the largest component... and the driver of” cost of goods sold, because “when you have the ability to late bind some of the more expensive components... you have a big book of business that’s flexible.” That is among the largest buyers of these chips describing optionality rather than commitment: the freedom to slow its own purchases without slowing its own business. She named no supplier, and the remark was about Microsoft’s own spending; its reach is the margin, not the contracts already signed, which are committed, but the ones that come next. A committed backlog is safe because it is committed; the freedom Hood described governs what comes after it, the renewal and the next signature, the same exposure the falling price carries later in this reading, now in the buyer’s own words. The chief executive published the test; the chief financial officer described the freedom. And she called the chips what the residual question already calls them, short-lived assets.

The sign of progress the company offers against all this, margin climbing off its one-percent trough toward a low-double-digit target, is measured on the wrong ruler for that table. A margin is income over revenue; the hurdle is income over capital; and on a business spending 7.7 times its EBITDA on equipment, even a low-double-digit margin is a low-single-digit return on the capital, a fraction of nine. The mark the company told the market to watch cannot carry it back inside the corners, and by its own year-end target it does not.

Every rung, up to the structure

It has used about every ordinary tool of this financing that the filings let anyone see. It takes its customers’ cash in advance. Its chief supplier both sells it chips and holds its stock, and one of its customers holds its stock as well. It ring-fenced its best contract, a Meta agreement its rating agency characterizes as take-or-pay though the contract’s commercial terms are redacted, into a separate company and financed it at investment grade, while the parent itself is rated speculative by all three houses, Ba3, B plus, and BB minus. And the last of these tools is the top of the ladder, structure, which the auditor also flagged: its second critical audit matter is whether a joint venture with a third-party developer, formed to build a multi-phase campus, belongs on the balance sheet at all.

The future the company is selling is premium-priced compute, rented to the few counterparties with the means and the stated intent to build it themselves. Microsoft, named in the annual report at about 67 percent of 2025 revenue, has just described, through its chief financial officer, the flexibility that governs what it buys next. Meta, the anchor of the investment-grade facility, has announced it is entering the cloud business. And the rates in the $99 billion backlog were written in the scarcity of 2023 through 2025, into a market that has since fallen between 64 and 75 percent from its peak on independent indices, with one hyperscaler cutting its own published prices up to 45 percent in a single announcement. A falling price pulls two things at once, the rate a contract renews at and the value of the machine that secures the debt, though a memory shortage has firmed near-term rates and has to be read out before the trend is read.

What has to happen next

So something has to happen, or not. The questions are few, and each is answered in a filing. Which rung will the quarter’s solution sit on, a new one or cash off an old one? Does the $1.3 billion earn into revenue as the machines deliver, settle as a note, a reversal, a return of cash, or a charge, resolve in a way no filing has shown, or simply stand another quarter, which an unaudited interim statement is allowed to let it do? Does the largest customer’s share keep falling? Does cash arrive from outside? And how does a business that sits outside the four corners present a future it concedes is unlikely to bring it back inside them?

Occam’s razor asks for the simplest answer. The reading here is that no simple answer is on the table; every candidate resolution is a hard one, and the absence of a simple path is itself the thing to notice. There may be a third kind of move as well, assembled for the occasion, an instrument or a structure without precedent in this company’s statements. A reading written before the data cannot describe what has not yet been done, and will not pretend to. It can say only that more structure and cash from outside may not be the only two answers, and that the unfamiliar move would be the most telling of all, because a company reaches for one when the familiar answers no longer serve.

The other outcome has to be written as plainly, and on the scale the quarter can settle. The margin comes off its one-percent trough, as management said it would. The largest customer’s share steadies rather than falls further. The $1.3 billion earns into revenue the way deferred revenue is meant to. And cash, if it arrives, arrives from outside the circle, on terms the filing states. If the second quarter shows those, this reading says so first and in full: whatever the filing shows gets an even reading, and if the news is good it runs first. The larger question, whether the return ever clears nine percent on its own, no single quarter can answer, and the arithmetic above says it would be the harder climb for having no cover to lean on; that one the following filings grade, not this one.

And whichever of these it is, it may not arrive as the headline. Any company leads with its best news, and the answer to a question about last quarter’s cash can sit as the second line beneath a headline about something else, a new commitment, a larger number, a fresh partnership. That is ordinary; it is how stories are told. This reading only insists on looking under the headline for the line beneath it. The quality of cash is, in the end, the discipline of reading the second line first.

This reading presupposes no wrongdoing and finds none. A critical audit matter is the auditor’s ordinary flag of a hard judgment, placed there by the people who signed the statements. A high rung is a financing choice. A caption the market has no template for is a company writing its own presentation, the ordinary cost of going first. What the filings show is simpler and quieter: a company selling a story, that it is the essential cloud for the age, and a story that is not working out on the numbers, and that may, on the numbers, prove unable to. Selling a future the present cannot yet fund is the ordinary condition of a company asking to be believed before it can prove itself. The only question this reading holds open is whether the proof arrives, and cash is the place it would have to show.

Beyond the filings the series does not claim to see; where the record is silent, the blank is left a blank. The answer arrives on a date already on the calendar, in numbers the company files itself.

Standing disclosure: Cape Fear Advisors holds no position, long or short, in CoreWeave or in any company named here. Anthropic is the developer of Claude, which is used in preparing this research, and Anthropic is a customer of CoreWeave under a multi-year commitment; that commitment gives Anthropic an interest in CoreWeave’s health, and nothing in this reading serves or disserves it. The nearness cannot be fully checked away, which is why no claim here rests on trust in the tool: every figure carries a public source, and the record grades the rest. Others named here have ties to the company: Microsoft and Meta are its largest customers; NVIDIA supplies its accelerators and holds a position in it; OpenAI is a customer and holds a position in it; Morgan Stanley is among its arranging banks. Companies not named here may hold positions or supply relationships that bear on the filers discussed, and that possibility is part of why every piece is re-checked for bias, ground facts, and filings rather than read against a fixed list. Figures are quoted from the filers without characterization, and the same standard of reading is applied to every filer named.

NOTES

Sources are public filings; the figures re-derive from them.

1. FY2025 and Q1 2026 figures, the $99.4 billion backlog, power, adjusted EBITDA, and the chief financial officer’s “trough” remark: CoreWeave FY2025 10-K, the Q1 2026 press release, and the Q1 2026 earnings call.

2. The two critical audit matters and the “service, not lease” conclusion: Deloitte & Touche LLP, FY2025 10-K(auditor’s report and Note 1); discussed across this series, first in “Adding It Up” and most recently in “Taken as a Whole” (July 4, 2026).

3. The $1.3 billion reclassification out of deferred revenue, and the customer-prepayment line tag: Q1 2026 10-Q, Contract Balances.

4. The customer-level dollars are derived, filed concentration share times filed quarterly revenue: Microsoft about $707M in Q1 2025 (72% of $982M), about $861M in Q2 2025 (71% of $1,213M), about $935M in Q1 2026 (45% of $2,078M); shares and revenue from the FY2025 10-K and the quarterly filings.

5. The four corners: Morgan Stanley (Brian Nowak) hyperscaler return-on-invested-capital analysis, as rendered and posted by Microsoft’s chief executive on July 30, 2026.

6. Amy Hood’s remarks on short-lived assets and slowing purchases: Microsoft FY2026 Q4 earnings call (Q&A, in response to Bernstein’s Mark Moerdler), quoted from a transcription-service record of the call, to be corrected by name should the company’s official transcript differ.

7. The Microsoft, Meta, and OpenAI master services agreements and the OpenAI common-stock issuance (commercial terms redacted): S-1 and 8-K exhibits (Meta MSA; OpenAI MSA).

8. Parent ratings Ba3 / B+ / BB- and the A3 ring-fenced facility: the agencies’ actions, the DDTL 4.0 8-K, and the June unsecured notes 8-K.

9. The H100 rental decline (64 to 75 percent from the 2024 peak) and the AWS price reduction (up to 45 percent, effective June 1, 2025): AWS’s own announcement and independent indices, labeled as reported.

10. The residual and the capex-to-EBITDA reading: this series’ “Twenty-Seven Years”.

11. The self-descriptions “The Essential Cloud for AI” and “sits between the models and the silicon” are CoreWeave’s own.

Analysis: Cape Fear Advisors.

Every business runs on a circle. A coin goes down on the bar and the drink comes back, the circle drawn so tight it reads as a point: value out, value in, one instant, settled. A day of credit between the coin and the drink, or a promise, or a machine that earns its cost across years, opens the point into a loop. The money goes out and comes home later, from somewhere. The loop is ordinary, and small enough, most days, to pass without a name.

The cash register was the first circle. The drawer rings the sale before the cash is counted; it turns a coin in a hand into a number on a page, recorded and owed and waited on. Every arrangement since is that same loop, drawn larger, with more hands on it and more time inside.

Circular financing is as normal as the day is long. The chipmaker takes a stake in the lab, and the lab commits to years of chips; the landlord builds the data center and leases it back; the cloud funds the customer who fills the cloud. Drawn as a ring, the picture invites alarm. Read as a ladder, it is an address system, familiar at every rung. The lower rungs are the everyday, and at the aggregate the record stands whole: the capital spending, the debt, the order books, and the guarantees are filed, to the dollar, by companies that publish thousands of pages a year. The record is already there. What it waits for is a reader who will say which rung a dollar sits on, how much larger the number has grown, and who is left holding the loss if the wheel stops.

Two things hold at once, and this piece is the space between them. Circular financing is as normal as the day is long. And bubbles burst company by company. The first is why the alarm carries weight; the second is why the alarm, on its own, guides no one. A wheel that turns for the whole market still stops at one address, the address where the cash ran out and the paper stayed. This piece is the framework that tells the two apart: it shows the circle for the ordinary thing it is, and it shows that a bubble, where there is one, arrives one company at a time and not a sector at once. The tool is a ladder, read one rung at a time, with a mark at the one step where the circle closes on itself and nothing stands outside to catch a fall.

The ladder

The ladder begins at the bar and climbs one rung at a time. Each rung adds a little time and one more balance sheet between the sale and its settlement, and the circle around the money widens by that much. Not one of the rungs is new; each predates the companies filling it this year.

The first rung above cash is the slow payment. A supplier ships and waits, and the wait rests on the balance sheet as a payable. Microsoft ended its June year with $26.7 billion of equipment purchases still sitting in accounts payable, against $6.9 billion a year earlier. The rung is trade credit, the first loop in commerce. The number is nearly four times what it was.

One rung along, that same payable can take a third party inside it. In supplier finance, a bank pays the supplier early and the buyer settles with the bank later, so an ordinary payable runs through a financier who now holds the timing. The arrangement carries its own disclosure line, required of buyers since 2023, and Microsoft’s filing shows none: the largest payable book in this group is the plain kind, a wait between two parties, not a loop with a bank set inside it. The rung is real, and on Microsoft’s page it sits empty.

A step up is the customer who pays first. Oracle carries about $75 billion of hardware its customers have prepaid or supplied ahead of the compute they will buy, cash arriving before the service it answers. One step further is the merchant of components: Microsoft holds $27.8 billion of receivables tied to buying server parts on others’ behalf, up from $8.2 billion, its balance sheet standing in the middle of its own supply chain.

Higher still is vendor financing, lending the customer the money to buy the product. Microsoft’s established line for it sits at $3.7 billion and has fallen; the classic rung stayed small while the newer ones swelled. On the same rung, SpaceX carries about $9.0 billion its own prospectus calls a failed sale-leaseback, a financing the accounting declined to read as a sale. Then the stake in the counterparty, the rung the alarm draws first: Microsoft has funded $11.9 billion into OpenAI and books about $24.1 billion a year of revenue from the same relationship, the two legs disclosed in one note. Above that, the contingent promise: Alphabet reports $7.6 billion of financial guarantees and a $43.8 billion ceiling on credit-derivative payments; Meta shows $46.0 billion of maximum exposure to vehicles it carries at $2.9 billion. And at the top, the structures: the joint ventures, the fund-level facilities, the SoftBank loan that gathered 21 new lenders this month.

Walked down, the ladder settles one fact. The money the alarm points at sits in the middle. The great sums are on the middle rungs, in trade credit and prepayments and component orders and stakes, the ordinary trades of any large enterprise carried out at a scale no enterprise has reached before. The vehicles the chart-makers fear, the structures at the top, hold less than the noise around them suggests. Height on the ladder is a measure of complexity, and complexity, on its own, is ordinary.

The pivot

The ladder measures one thing: how far a dollar sits from cash, and how tangled the circle around it has grown. Height is complexity, and complexity is ordinary. What the ladder is built to find is something else, and it comes into view on the way up: the place a circle has closed far enough that a failure inside it has no outside cash to catch it.

The more a loop closes on the same hands, the less stands outside it to break a fall. A trade payable has the whole of commerce behind it; if one supplier goes unpaid, the loss lands on a party who stands apart from the buyer. A circle that funds its own customer and takes its revenue back from that same customer has drawn the catch inside itself. While it turns, nothing shows. When a payment comes up short, the question is what stands outside the circle to absorb it, and the answer is sometimes nothing.

Enron is the structure worth remembering here, and only the structure. Its vehicles were capitalized with Enron’s own stock, so the thing meant to absorb a loss was a piece of the company the loss would fall on. Mark-to-market was the rule of its day, applied as written. The circle closed on one set of hands, and when the business stopped working it had only itself to fall back on. Whether a failure of that kind traces to fraud or to a business that simply stopped is a separate question and a separate craft. This work reads for the earlier thing, the place the cash stops working.

So the cliff is the closure of the loop, not the height of the rung. A structure many rungs up, with an independent party on the far side and a catch that stands outside the circle, is an open loop however intricate. A low, simple circle that funds its own demand and books the return has already pulled the catch inside. Two questions find it, and neither asks about intent: who stands on both sides of the paper, and does the thing that would absorb a loss stand inside the circle or outside it.

The checks

A ladder that only sorted arrangements by shape would be a filing cabinet. What makes it read is that each rung carries its own measure, a single ratio drawn from the filings that says how much of the rung is still cash and how much has turned to paper. Most come down from the quality-of-earnings tradition, the reading Thornton O’glove set out and Howard Schilit carried forward, the practice of asking what a reported number is made of. Applied rung by rung, they become a quality of cash.

Low on the ladder the measures are plain. On the trade-payable rung, the reading is days of payables set against the growth of the revenue they serve: a payable book swelling faster than the sales beneath it is a loop lengthening, cash held longer than the business earns it. On the supplier-finance rung above it, the reading is the program balance a buyer must now disclose, set against the payables it finances, which says how much of an ordinary payable book has a bank threaded through it. On the prepayment rung, prepaid balances against the revenue still owed as performance. On the component rung, the receivable that finances a supply chain against the capital spending it serves. Each is a fraction anyone with the filing can build.

Higher, the measures sharpen. On the stake rung the reading is the revenue a company takes from a firm it has funded, set against the money it has put in. When a stake returns two dollars of revenue for every dollar funded, and the revenue and the funding rest in the same note, the ratio reads the closeness of the loop. On the contingent rung sits the strongest single number the family holds, and it is Rod Dubitsky’s: maximum exposure set beside carrying value. Meta discloses $46.0 billion of maximum exposure to vehicles it carries at $2.9 billion, about fifteen to one, the size of the promise measured against the size of the mark. It reads what a footnote is holding back.

None of these is a verdict. Each reads for the one thing the ladder was built to find, the place the cash stops working. A ratio that drifts the wrong way for a quarter is a question; the same ratio drifting for a year, on a rung high enough that the catch has moved inside the circle, is the ladder pointing at an address. The measures do not accuse. They locate.

Who prints cash

There is one more cut, and it grades the whole ladder from outside it. Two kinds of entity make their own cash. A government issues it. A bank, lending, creates a deposit that did not exist a moment before, which is not a metaphor but the plain mechanics of the modern system, set out by the Bank of England over a decade ago. Every other party does one of two things with cash. It earns it, or it shows something that only looks like it.

The look is convincing at the top of a boom. A chipmaker that funds the buyers of its chips is, in a way, printing its own demand, and demand spends like money for as long as the buyers keep buying. The currency is compute rather than dollars, and the press is the circle: money out to the customer, orders back for the machines, revenue booked, the loop feeding itself. It is real while it turns. It is not cash. The dollars arrive only when a customer outside the circle pays a bill that clears.

Further from cash still is the mark. A fund that carries a private stake at a rising value can show a gain that reads like earnings, a golden egg laid each quarter. The egg is a valuation, not a settlement. It becomes cash on the day the stake is sold, and until that day it can rise and fall with nothing changing hands. A house whose best quarter is a mark on a position it still holds has shown the quality of cash at its thinnest.

This is why the earlier ladder in this series, the one that graded four dollars by what still had to happen to each, put the croupier’s dollar nearest cash and the mark’s dollar furthest. The croupier, the bank that places the paper, is paid first and in the settled article, and owes nothing further; the bank sits near cash whether it is printing a deposit or taking its fee. The mark is paid last, in a price that has to hold. Between them run the register’s dollar, a sale that waits on a solvent buyer, and the operator’s dollar, which a machine must earn across years. The circle does not change which dollar is which. It only makes the distance easier to forget.

Three registers of naming

The circle has been named three times this year, at three volumes, and the piece can stand among them. The loudest naming is the alarm: a bubble, reckless, a house of cards, drawn as one frightening ring. The alarm has the hard part right. Assembling this from the filings is difficult, and the difficulty is exhausting, and at the end of it many readers throw up their hands and shout. The shout has the strain right. It stops one step short of the reading.

The soberest naming came from the Bank for International Settlements, the central banks’ own bank, which in its 2026 report placed circular financing at the top of its list of risks and described it in flat, unfrightened prose: chipmakers and hyperscalers taking stakes in the labs that then commit to buy their chips, data centers built by third parties and leased back on long contracts, terms it called poorly disclosed. The institution is right that some of it is dark, and which part matters.

What is filed is the aggregate: the capital spending, the debt, the order books, the guarantees, printed to the dollar. What sits in shadow is the fine print of the private deals, the pledge order and the bilateral terms. A headline this week put hidden borrowing at more than a trillion and a half; the borrowing is not hidden, it is filed, and it is the terms that stay dark. Keeping those two apart is most of the work, because the alarm blurs them into one, and the blur is the frightening part.

Behind all three namings is the same slow machine. The rules that force a vehicle’s maximum exposure onto the page arrived within two years of the failure that prompted them and took most of a decade to sharpen, and they ask a great deal of any reader who would use them. The disclosure is there. It is written to be turned to rather than read, and turning to it is the whole of the craft.

The close

So the tool is a ladder, and the reading is a walk down it. Which rung does a dollar sit on. How much larger is the number than it was. Who holds the loss if the wheel stops, and does the thing that would hold it stand inside the circle or outside. A measure at every rung to say how much is still cash. None of it forecasts. All of it locates.

Circular financing is as normal as the day is long, and the sentence holds best whole, because half of it is the reassurance and half is the warning, and the two are true together. The circle is ordinary. And if a bubble is here, it will not arrive as a sector falling at once. It will arrive at one company, on one rung, at the moment the cash that was meant to catch a fall turns out to sit inside the circle that is falling. That is a company-by-company event, read one company at a time.

A company found at that rung is, first, only itself. Whether its trouble stays its own, sends ripples through the names around it, or reads as a tell of something larger is a second question, asked after the first and held apart from it. The one company matters to itself before it means anything for the rest. The bubble, if the word carries weight, is what a company might signal, not what it is.

The register was the first circle, and it is still the plainest one, because the drawer rings and the cash is counted in the same breath. Everything above it trades a little of that immediacy for a little more reach, and the trade is old and mostly sound. The work is only to keep counting: to walk the ladder to the rung where the counting stops, and to say so plainly, before the fall names the company.

Standing disclosure: Anthropic is the developer of Claude, which is used in preparing this research. That nearness cannot be fully checked away, which is why no claim here rests on trust in the tool: every figure carries a public source, and the record grades the rest. Among the companies named, OpenAI is a competitor of Anthropic; Alphabet holds a large position in Anthropic; CoreWeave and SpaceX are compute counterparties to it; and NVIDIA supplies the infrastructure that serves it. Companies not named here may hold positions or supply relationships that bear on the filers discussed, and that possibility is part of why every piece is re-checked for bias, ground facts, and filings rather than read against a fixed list. Figures are quoted from the filers without characterization, and the same standard of reading is applied to every filer named.

NOTES

The trade-payable rung. Microsoft Corporation Form 10-K, fiscal year ended June 30, 2026, accession 0001193125-26-323660, Note 6: purchases of property and equipment remaining in accounts payable $26.7 billion at June 30, 2026, against $6.9 billion a year earlier. The Microsoft figures throughout are from this filing.

Supplier finance. The buyer-side disclosure requirement is FASB Accounting Standards Update 2022-04, supplier-finance-program obligations, effective for fiscal years beginning after December 15, 2022. A full-text search of the Microsoft 10-K returns no supplier-finance program.

The prepayment rung. Oracle Corporation, fourth-quarter fiscal 2026 results, Exhibit 99.1 to Form 8-K, accession 0001193125-26-265848: customer-prepaid and customer-supplied hardware of about $75 billion.

The component rung. Microsoft 10-K: $27.8 billion of other receivables related to facilitating the purchase of server components, up from $8.2 billion.

Vendor financing. Microsoft 10-K: customer software financing receivables $3.7 billion, down year over year. SpaceX: Form S-1 / final prospectus, accession 0001628280-26-042639, Note 17: sale-leaseback obligations deemed a failed sale-leaseback, $1.1 billion current and $7.9 billion non-current, about $9.0 billion, carried as related-party debt. The filed term is “failed sale-leaseback”; the reading of the arrangement as a financing in substance is Rod Dubitsky’s.

The stake rung. Microsoft 10-K, related-party disclosure under ASC 850: $11.9 billion funded into OpenAI and about $24.1 billion a year of revenue from the same relationship, the two legs in one note.

The contingent rung. Alphabet Inc. Form 10-Q, Notes 3, 5, and 10: $7.6 billion of financial guarantees and a $43.8 billion maximum on credit-derivative payments. Meta Platforms, Inc. Form 10-Q for the quarter ended June 30, 2026, accession 0001628280-26-050705: maximum exposure to loss of $46.0 billion against a carrying value of $2.9 billion for a data-center venture it does not consolidate. The maximum-exposure-to-carrying reading is Rod Dubitsky’s.

The structures rung. Reported: the SoftBank facility that added 21 lenders this month; NVIDIA’s investment offer to OpenAI, up to $100 billion, and an order of about $300 billion of compute; and, on the lead exhibit, SoftBank’s sale of its entire NVIDIA stake, about $5.83 billion, redeployed toward its OpenAI position in November 2025.

The quality-of-earnings tradition. Thornton O’glove, Quality of Earnings (1987); Howard Schilit, Financial Shenanigans.

Who prints cash. Bank of England, “Money Creation in the Modern Economy,” Quarterly Bulletin 2014 Q1 (Michael McLeay, Amar Radia, Ryland Thomas).

The institution. Bank for International Settlements, Annual Economic Report 2026, Chapter I, “Progress and peril.”

Enron. Used for structure only. The variable-interest-entity consolidation rules arrived with FIN 46 and FIN 46(R) in 2003 and were sharpened by FAS 167 in 2009.

On the shelf. The four-dollar ladder is The Quality of Cash: What Has to Happen Next, the shelf anchor; the croupier’s dollar is The Croupier Counts First; the aggregate-is-filed reading is Big Tech, Adding It Up: The Trillion That Was Never Hidden; the disclosure-gap companion is Apple, Adding It Up: The Immaterial Toll Booth. All on the Quality of Cash shelf.

Analysis: Cape Fear Advisors.

On July 22, Alphabet, the parent of Google, reported the biggest profit in the company’s history, and, as far as the public record shows, the biggest any listed company has ever reported: net income of $112.2 billion for the three months ended June 30, up nearly threefold from a year earlier. The size can wait a moment; it is the least interesting thing about it.

Because in the same quarter the cash went the other way. Alphabet spent more building than its operations brought in: $39.1 billion of cash from operations against $44.9 billion of capital expenditure, so free cash flow was negative, about $5.9 billion, and the full-year capital plan was raised again, to a range of $195 to $205 billion. The headlines went there, to the negative cash flow, and they were right to. That was the real event.

The record was not the event, because the profit was mostly not cash. Of the quarter’s income, $99.0 billion came from a single line, “gain on equity securities, net.” That line is a mark: the recognition of a change in the value of what Alphabet owns of other companies. No sale occurred, and no dollar was collected. Take it out and the operating quarter reads normally; diluted earnings without the gain were about $2.85 a share, against roughly $2.87 expected. The record was the mark.

It survives even the tax. Pretax income was $138.8 billion. Against it the company booked a $26.6 billion tax provision, but $20.6 billion of that was deferred: a tax recorded, not paid, the accounting shadow of a gain recorded and not collected. One statement over, the two undo each other. The cash flow page removes the $99.0 billion gain and adds the deferred tax back, and the $112.2 billion record resolves to $39.1 billion of cash from operations. Set $44.9 billion of building against that, and there is the minus sign the headlines led with. The reconciliation is Alphabet’s own, printed one page after the record.

Underneath all of it was an excellent quarter, reported plainly and then passed over. Strip the mark entirely and Alphabet still grew revenue 24 percent to $119.8 billion, lifted operating income 30 percent to $40.8 billion, and widened operating margin to 34 percent. Google Cloud grew 82 percent and more than tripled its operating income, to $8.8 billion. The business underneath is large and generative: it produced $39.1 billion of operating cash in the quarter, roughly $185.7 billion over the trailing year, and it very nearly paid for this quarter’s build. Capital expenditure of $44.9 billion, extraordinary on its own, was met all but about $5.9 billion by the quarter’s own operating cash. That is the shape of a formidable company, and it is the part the release reported without dwelling on, before the call moved quickly to the capital plan. A house like this does not lean on a mark to make its number, which is the quiet point beneath the loud one: the mark did not make the quarter, and the quarter stood without it.

The mark itself came from two places, and the difference between them is the whole story.

The first is SpaceX. Alphabet put $1 billion into SpaceX in January 2015, when the company was valued near $6.7 billion. SpaceX went public this June, and once a public price exists the accounting stops being a matter of judgment: the holding is carried at that price, and the change runs through income. Alphabet’s stake was worth about $94 billion at quarter end. This mark is legible because a market makes it, and the doubt around it is legible too, because a market says it out loud: the shares have slipped back below their $135 offering price since the debut, and Alphabet cannot sell into the slide, roughly $80 billion of the position under the standard post-offering lockup and the rest restricted into 2027. A bet the market can price, in public, carrying every question a public price carries.

The second is Anthropic, and here the page goes dark. Alphabet has invested a reported $13.3 billion in Anthropic, a private company whose valuation is reported to have moved from about $350 billion to about $965 billion over the period. When a private holding reprices in a new round, the holder marks it too. So there is an Anthropic mark inside that $99 billion. How large, no one outside the company can say, because Alphabet did not break it out. The $99 billion is a single aggregated line. On the earnings call the finance chief described it only as “primarily due to unrealized gains in our equity securities portfolio.” Neither SpaceX nor Anthropic was named. No analyst asked.

So the two halves of one number sit in two different worlds. One is a bet the market prices, and prices downward, in the open. The other is a bet with no price at all, held by one company that has not broken it out and issued by another that files nothing. A reporter estimating the Anthropic figure is not guessing idly; there is nothing filed to read. The number survives only in other people’s arithmetic: take out the visible half, or borrow the same round’s mark from Amazon, which does name its Anthropic stake, and triangulate toward it.

That is the thing never discussed, and it is the real subject. The largest figures on these statements are often the ones no one is required to explain. Alphabet folds a $99 billion gain into a line and moves on. One filer over, Apple books roughly $20 billion a year from Google into Services and calls the arrangement “quantitatively and qualitatively immaterial,” the figures withheld from its own letters to the SEC. Anthropic, not yet public, discloses nothing at all. Three ordinary and permitted ways a very large figure stays off the front of the page: folded into an aggregate, called immaterial, or not yet filed. Each is available only to a company large enough to write its own disclosure.

What each does with a mark it cannot avoid booking is a tell. Amazon named Anthropic plainly last quarter and let a $12 billion mark carry much of its profit, on real operating cash. Microsoft carries a dedicated line to remove its OpenAI gains from its own adjusted results, asking to be judged without them. Alphabet booked the biggest mark of all and gave it a sentence. Name it, set it aside, or pass over it: three postures toward one forced entry. A house does not rest its number on a mark, because marks move, and the direction is not the house’s to choose.

None of this is abstract this week, because the next several days test it in the open. Microsoft reports July 29; a year ago its OpenAI gains were lifted by a one-time event that has now lapsed. The line to watch is whether the mark now runs the other way, and whether the company still sets it aside. Amazon reports July 30, and its named Anthropic figure is the closest thing anyone has to a key for the number Alphabet would not print. Alphabet’s own quarterly filing follows in early August, and it will either name the two companies and split the gain or keep the aggregate. Anthropic is reported to be heading for an offering as soon as October, the first moment its own numbers would answer for themselves.

Positioned correctly, the record is a distraction, not a headline. At $112.2 billion it is, as far as the record shows, the largest quarterly profit any listed company has ever reported, more than double the prior mark, Saudi Aramco’s roughly $48 billion in 2022. Those are not the same kind of record. Aramco’s was oil, collected in cash. Alphabet’s is a mark, the kind Berkshire Hathaway’s owners are told every year to disregard, because unrealized gains make the profit line move on prices rather than on performance. Alphabet’s own finance chief said the word in passing: unrealized. The company did not sell the record. And yet the ledger will carry the line and drop the note. The record will read “largest quarter ever,” and it will not carry the footnote that the number was a mark on two bets, one the company cannot sell and one no one can price.

The engine nearly covered this quarter; the build ahead is another matter. Paying for it has turned Alphabet from a company that returns capital into one that raises it. In early June it launched an $80 billion equity program, reported as its first equity raise since the 2004 offering, and set its buyback to zero. In the quarter itself it took in about $74 billion of new debt and equity, and ended June holding about $242 billion in cash and marketable securities, roughly double the $127 billion of three months earlier. The rise is itself a mark. At the IPO the SpaceX stake moved into marketable securities, so about $94 billion of that $242 billion is the same stake seen earlier, marketable because it now trades and still locked. Setting it aside leaves about $148 billion available, and cash alone still rose, to $55.9 billion, after $44.9 billion of building.

Those raises, at close to the highest ratings and out to a hundred years, this series has followed before, in The Croupier Counts First and Three Houses, Three Placements. One buyer is worth naming. The same Berkshire Hathaway whose accounts are the standing example of why a reader disregards an unrealized mark took $10 billion of the offering, named in its own prospectus: $5 billion in Class A at $351.81 and $5 billion in Class C at $348.20. The investor most associated with ignoring the mark put cash into the build. Cash is the one thing a mark is not.

None of this is a change of mood, and none of it is hidden. It can look like a conjuring trick, but nothing is conjured: the profit is real, the cash is real, and they sit on different pages. The headlines had the cash flow right; what stayed closed was the profit, one line that holds two marks and a tax that is also not cash. Set the income statement beside the cash flow, the holder beside the issuer, the line beside its footnote, and it resolves into plain arithmetic. Anyone can do the adding, and once it is done, the largest number in the history of corporate profit turns out to be the one that mattered least, in a quarter with better numbers to talk about.

Related, on the Quality of Cash shelf: The Quality of Cash: What Has to Happen Next (the anchor); Apple, The Immaterial House; The Croupier Counts First; Three Houses, Three Placements.

NOTES

Alphabet Q2 2026 results, Form 8-K exhibit 99.1, accession 0001652044-26-000066 (quarter ended June 30, 2026), sec.gov: https://www.sec.gov/Archives/edgar/data/1652044/000165204426000066/googexhibit991q22026.htm. Revenues $119,796M (+24%); operating income $40,770M (+30% from $31,271M); operating margin 34% (from 32%); income before income taxes $138,753M; provision for income taxes $26,560M (effective ~19.1%); net income $112,193M; diluted EPS $9.11. Google Services revenue $94,540M, operating income $39,544M. Google Cloud revenue $24,768M (+82% from $13,624M), operating income $8,814M (from $2,826M). Google Search & other $63,271M; YouTube ads $11,055M. Gain on equity securities, net $99,031M; other income (expense), net $97,983M. Cash flow reconciliation: net income $112,193M, less gain on equity securities $98,999M, plus deferred income taxes $20,618M, plus other operating items ~$5,257M, equals net cash from operations $39,069M (about $185.7B trailing twelve months per the call); purchases of property and equipment $44,924M; free cash flow negative $5,855M (operating cash covered ~87% of quarterly capex). Full-year 2026 capex guidance $195-205B (raised from $180-190B). CFO remark and the absence of an analyst question on the equity gain are from the Q2 2026 earnings call, July 22, 2026 (transcript).

EPS excluding the investment gains (about $2.85) versus consensus (about $2.87): reported earnings coverage, July 22-23, 2026.

Q2 2026 financing activities (same 8-K). Proceeds from issuance of debt, net $24,847M; repayments of debt $3,776M; proceeds from issuance of common stock, net $30,499M; proceeds from issuance of mandatory convertible preferred, net $19,063M; repurchases of stock $0; dividends paid $2,689M; net payments related to stock-based award activity and other financing items about $6,701M; net cash provided by financing $61,243M. New debt-and-equity proceeds this quarter about $74.4B. Cash and marketable securities: $126,840M at March 31, 2026 ($38,063M cash + $88,777M marketable) to $242,474M at June 30, 2026 ($55,911M cash + $186,563M marketable). Marketable securities rose from $88,777M to $186,563M, most of it the SpaceX stake entering the line at its June IPO. Q1 2026 results (March 31 balances), accession 0001652044-26-000043, sec.gov: https://www.sec.gov/Archives/edgar/data/1652044/000165204426000043/googexhibit991q12026.htm. Q1 2026 capex $35,674M; Q2 2026 capex $44,924M.

Alphabet equity raise: about $80B program, free writing prospectus accession 0001193125-26-251733 (press release June 1, 2026), sec.gov: https://www.sec.gov/Archives/edgar/data/1652044/000119312526251733/d160205dfwp.htm. Structure: about $30B concurrent underwritten offerings ($15B mandatory convertible preferred + $15B Class A/C common), about $40B at-the-market program (beginning Q3 2026), and a $10B private placement to Berkshire Hathaway named in the prospectus ($5B Class A at $351.81, $5B Class C at $348.20). The FWP states Alphabet has raised over $85B of debt over the last year. The first-equity-raise-since-2004 characterization and the 100-year (century) bond in this year’s debt program are reported, and are covered in The Croupier Counts First and Three Houses, Three Placements.

SpaceX stake about $94.1B at June 30; $1B invested January 2015 at about $6.7B valuation; about 4.9% ownership; about $80B under the standard post-IPO lockup, $14.1B restricted through Q3 2027; the $135 offering price and its decline are reported. The SpaceX-Anthropic compute agreement (about $1.25B a month) is in the SpaceX S-1, read in The Ninety-Day Annuity.

Anthropic: about $13.3B invested by Alphabet; valuation about $350B to about $965B; offering reported as soon as October 2026.

Amazon Anthropic mark $12.3B, named: 10-Q accession 0001018724-26-000014 (quarter ended March 31, 2026). Microsoft non-GAAP line removing OpenAI gains and losses: 10-Q accession 0001193125-26-191507. Apple, about $20B a year from Google in Services, “quantitatively and qualitatively immaterial,” figures withheld: CORRESP accession 0000320193-24-000061.

Saudi Aramco, about $48.4B, Q2 2022, prior largest quarterly profit of any listed company (Al Jazeera, August 2022): https://www.aljazeera.com/economy/2022/8/15/saudi-aramco-posts-biggest-quarterly-profit-of-any-listed-company.

Standing disclosure

Anthropic is the developer of Claude, which is used in preparing this research, and Anthropic is also one of the two holdings whose marks this piece reads. That nearness cannot be fully checked away, which is why no claim here rests on trust in the tool: every figure carries a public source, and the record grades the rest. Others named in this piece have ties to Anthropic: Alphabet, the subject, carries a reported $13.3 billion position; Amazon, cited here for its own named mark, holds a stake as well; and SpaceX, the other of the two holdings marked here, counts Anthropic as a major compute customer. Companies not named here may hold positions or supply relationships that bear on the filers discussed, and that possibility is part of why every piece is re-checked for bias, ground facts, and filings rather than read against a fixed list. Figures are quoted from the filers without characterization, and the same standard of reading is applied to every filer named.

Analysis: Cape Fear Advisors.

That difference needs a reconciliation the moment has outgrown. The buildout is widely called the most expensive capital outlay in modern history, and it is financed on a circle: the same names invest in one another, buy from one another, and stand behind one another’s paper, through arrangements that run from ordinary terms to the most elaborate special-purpose vehicles. An income statement cannot see the circle; a cash flow statement sees one company’s half of it. This piece reads with a third reconciliation, CBITDA, which does to the cash flow statement what the cash flow statement does to the income statement: it looks past the page, into the footnotes and into other companies’ filings, where the circle nets to zero. It reads the ecosystem whole, and each player’s role inside it, credit and equity on one ledger.

A season of reading those filings has graded dollars end to end on that axis. Sorted by what must still happen, four kinds appear, from the dollar already done to the dollar that may never arrive. Each has a piece behind it.

The croupier’s dollar

The house that places the paper is paid first, and paid in a dollar that needs nothing further. The underwriting discount is collected at settlement, not billed; it owes no future performance; and once settled it is indifferent to whether the paper it placed ever performs at all. By the shelf’s measure it is the highest-quality dollar the series has graded, and it always will be. It is why the house runs the casino, and it is earned: the diligence, the distribution, the brand put behind the paper, the balance-sheet risk carried while it is placed. The work is finished by the time the dollar settles, which is what makes it the best dollar at the table. The banks reported this quarter first, as they are paid first, and that harvest is already printed and read. (1)

The filings price it to the basis point wherever the deal is registered. Across one cohort, the same lead banks throughout, the placement cost ran from NVIDIA’s 18.9 basis points on investment-grade notes to Alphabet’s 97.3 on common stock, with the SpaceX IPO at 66.7 and a full tenor ladder in between, the far end of a curve costing as much as five times the near end. FINRA’s rule contemplates a ceiling of five percent; the SpaceX table printed two-thirds of one percent. The gap between what the rule permits and what the table prints is pricing power, and it runs toward the issuer. (1)

Two bounds sit on the grade. It renews only with the flow, falling back toward its own prior-year comparators when the issuance window narrows. And the same houses that step around the paper’s risk for a fee sit on the lender side of the same borrowers, where the exposure the fee avoids still lives; who holds that risk is referenced in the filed exhibits rather than reproduced in them. The fee dollar is pristine; the franchise around it is not unexposed.

The register’s dollar

The next dollar is booked at the counter and waits to be collected. A sale is recognized; the cash arrives later, on the order of forty-five days, if the buyer is solvent when it comes due. What has to happen next is small and ordinary, and it is not nothing: someone has to pay.

The filed instance is NVIDIA’s receivable, as the fourth house set out. Three unnamed customers held sixty-four percent of a $40.7 billion balance; the balance stood at about forty-five days of the quarter’s sales; the allowance against it priced the worry at roughly one basis point. The lines to watch are whether forty-five days lengthens, whether the concentration shifts, whether one basis point moves. The mirror is already filed on a buyer’s ledger, where vendor payables supplied $960 million of CoreWeave’s first-quarter cash: a receivable that lengthens at one register is the same days, booked as a source, at another. (2)

The operator’s dollar

The third dollar has to be earned by a machine, over years, and this is where the distances open. The operator buys equipment on borrowed money and must run it long enough to repay the borrowing and then some. Asked of CoreWeave as arithmetic, how many years of the machine earning at its filed rate would return the machine’s cost and the cost of the money, the answer ran to fourteen with an amortizing lender it never had, twenty-seven as the stack is filed, and never on the strategy the record shows, against equipment the company depreciates on a six-year schedule and a debt stack of $25.1 billion, fifty-nine percent of it at effective rates of ten percent or higher, none of it reaching past 2032. (3)

This is the dollar with the most still to happen before it is real, and the most that can go wrong in the waiting. The croupier’s dollar was done at settlement; the operator’s dollar is a wager that a depreciating machine outearns its financing across a horizon longer than the machine’s own booked life. Everything downstream of it, the marks in the next rung included, rests on whether this dollar shows up.

The mark’s dollar

The fourth dollar was never collected and never billed. It is a mark: a change in the value of a stake that the rules compel the holder to recognize once the position is material. The number is not a choice. Mark to market does its work when the standard requires it, and the gain lands on the page. What each filer chooses is where on the page it lands and how much a reader is shown, and four of the largest holders chose differently this year.

Amazon named it. It recognized $12.3 billion of upward adjustment on private-company equity, “primarily from our nonvoting preferred stock in Anthropic,” stated the counterparty and the amount, carried the fair value from $14.8 billion to $32.0 billion across two dates, and put the figure on the front page. (4)

Alphabet reported the largest mark of all and named no one. Its gain on equity securities was $36.9 billion for the quarter, against $9.8 billion a year earlier, folded into a single line the filing attributes to “net unrealized gains on non-marketable equity securities.” No portfolio company is named anywhere in the document. (5)

Microsoft removed it. Its OpenAI interest is carried by the equity method, a different instrument, and its result runs through other income; the company then carries a non-GAAP line, “net (gains) losses from investments in OpenAI,” that strips the impact out, so a nine-month gain of $5.9 billion, most of it a one-time dilution gain from the October 2025 recapitalization, is subtracted back out of adjusted earnings. (6)

NVIDIA recognized $16 billion of it in income, $13.4 billion on publicly held stakes and $2.6 billion on private ones, in a quarter its operations produced $50 billion of cash. The operating income does not depend on the marks. Some of the revenue arrives from companies NVIDIA has itself put money into, and the marks are the upside on those same stakes, which either arrives or does not. On the far side of the same ledger, in the investing section, are two numbers: $17.9 billion of new money into private stakes in a single quarter, and $27 billion of investment commitments still open. Whether that number keeps climbing is a line the next prints draw. (7)

A mark becomes cash only when the stake is sold, or it is held until it is worth nothing, and in between it is re-marked up and down with no cash moving at all. A gain this quarter can be a loss next quarter and not one dollar has changed hands. That is why it belongs below the operating line, and why what a company does with it there is the whole of the judgment.

What is standing in for the cash

Behind that order runs a second one. The croupier’s dollar is cash. The register’s dollar is a sale that becomes cash on collection. The operator’s dollar is cash only if the machine earns it across the years. The mark’s dollar is not cash at all until a sale that may never come. Down the rungs, the reported result leans less on cash and more on something standing in for it.

Turned on a single company’s headline number, CBITDA sets its EBITDA against the cash behind it, in the footnotes and the counterparties’ filings, and the two come apart. EBITDA is read as a proxy for the cash an operation throws off; the distance between the proxy and the cash is the finding.

Oracle is the test, a business model in transition, and it prints its own answer. Its headline non-GAAP earnings per share for fiscal 2026 was $7.63; the company states that without the investment gains left inside it the figure is $6.83, so about a tenth of the number is gains rather than operations. Behind the earnings sits a backlog of $638 billion, of which the filing expects twelve percent to become revenue in the next twelve months, and a free cash flow of negative $23.7 billion for the year on $55.7 billion of capital spending. The headline is a promise, and an earnings number that includes the investment gains; the cash ran the other way. None of that reading is ours to prove. Oracle filed each number, the subtraction included; the lens only set them beside one another. (8)

This is not a verdict on Oracle. It is a company netting under extraordinary complexity, and it wrote the uncertainty into the same filing: the returns on its buildout, Oracle says, are “dependent on customer demand and the ability of our key customers to meet their contractual obligations,” and a change in either “may adversely affect operating margins, cash flows” and force a test of “the recoverability of related long-lived assets.” Most of the backlog lands beyond the first year, and the build is funded in the meantime by tens of billions of new debt and equity. The bull reads a contracted backlog no rival can match; the skeptic reads a promise financed ahead of the cash. Oracle wrote both into its 10-K; the next reports answer which governs, not this one. (8)

The reading assumes only what it must, that the filings are accurate and presented fairly in all material respects, and it carries its own inferences in the open, to be tested beside them. What it yields is less a single figure than a set of questions, some direct to the answer and some derivative of it, and much of the work of reading them is still ahead.

Most companies keep the mark out of the number they offer as cash, and the refusal is the tell in the other direction. Nebius kept a $780.6 million revaluation gain, the item that made its net income look remarkable, entirely out of its adjusted EBITDA. The four largest holders of these marks report no EBITDA at all; every mark sits below their operating line, and Microsoft subtracts its OpenAI gain out of adjusted earnings besides. CoreWeave’s gap comes from timing and set-aside interest rather than a mark, and the lens reads it the same way. (9)(10)

One definition leaves the mark inside, and no print has yet turned it into a gain. Two builders on the bitcoin side, TeraWulf and Core Scientific, define adjusted EBITDA to add back their warrant marks but not the remeasurement of the bitcoin they hold, so that line sits inside the metric. This quarter it was a small loss and lowered the figure. In a quarter the price rises, the same line becomes an unrealized gain that flows into the metric with nothing removed. The definition is filed; the quarter that turns it into a gain has not printed. (11)

The names here are the ones the market already talks about, and that is the only reason they are here; the examples could have been others. The reconciliation does not depend on the choice. Applied to any filer in the buildout, the reading holds, because it reconciles to the very numbers already being pointed at, and now and then points to one that no one is circling yet.

What the lens asks next

This is a skeptic’s lens, and it should say so: it asks future value to prove itself as cash, and it holds the dollar in hand above the dollar still promised. Turned the other way it is the bull’s lens too, because a bull’s case is won on the same page a skeptic’s doubt is, in whether a promise becomes cash, on time, from a payer still standing. The lens does not settle that argument. It points to where the argument settles, and names the filing that settles it.

The questions it leaves run past the one company shown here. SpaceX’s coming reports answer the distance between the cash a build of that scale requires and the output imputed to it in the valuation. CoreWeave’s next balance sheet answers which way the customer-liabilities caption resolves, toward revenue earned or toward something that services like debt. Microsoft’s disclosures answer how it carries its OpenAI obligations as the one-time gains roll off. OpenAI’s own numbers, if and when they are filed, answer the largest, how it meets the cash obligations it has taken on, and what they are. And Anthropic’s S-1 answers the last, the presentation and materiality judgments it makes on its first public page.

Some of these the market already circles, OpenAI’s obligations most of all. Others sit in a caption or a threshold few are reading yet. The lens does not answer them; it says which carry the weight.

The market quotes each of these companies on one figure and a multiple, and the figure does not carry its own grade. What has to happen next for it to become cash is printed in the filings, one rung and one reconciliation at a time. The grade is in the gap between the headline and the cash, and that gap is what CBITDA reads.

DISCLOSURE, standing: Anthropic is the developer of Claude, which is used in preparing this research. CoreWeave announced a multi-year agreement with Anthropic in April 2026; Amazon and Alphabet, named in this piece, hold large positions in Anthropic; and NVIDIA supplies the infrastructure providers that serve Anthropic. Figures are quoted from the filers without characterization, and the same standard of reading is applied to every filer named.

Figures are verified against the primary filings; documents are cited by accession number. Analysis: Cape Fear Advisors.

NOTES

(1) The fee ladder and the FINRA comparison are set out in this series’ croupier work: SpaceX Form 424B4, accession 0001628280-26-042639; NVIDIA Form 424B5, accession 0001193125-26-273139; Alphabet common-stock and note fee boxes as filed (June 2026). Cross-references: The Price of the Seat and The Croupier Counts First.

(2) NVIDIA receivable and concentration as set out in NVIDIA, The Fourth House; the balance is the quarter reported there and is refreshed at the next NVIDIA print. CoreWeave vendor-payables figure: 10-Q, accession 0001769628-26-000222.

(3) CoreWeave, Inc. Form 10-Q for the quarter ended March 31, 2026, accession 0001769628-26-000222, and CoreWeave, Twenty-Seven Years for the convergence arithmetic in full.

(4) Amazon.com, Inc. Form 10-Q, accession 0001018724-26-000014 (quarter ended March 31, 2026): upward adjustment on private-company equity $12,328 million, “primarily from our nonvoting preferred stock in Anthropic”; fair value of the nonvoting preferred approximately $14.8 billion at December 31, 2025 and $32.0 billion at March 31, 2026. A further $4.5 billion reclassification on converted convertible notes brings the two Anthropic lines to about $16.8 billion; total other income, net was $15.6 billion.

(5) Alphabet Inc. Form 10-Q, accession 0001652044-26-000048 (quarter ended March 31, 2026): gain on equity securities, net $36,915 million, against $9,758 million a year earlier; no portfolio company named.

(6) Microsoft Corporation Form 10-Q, accession 0001193125-26-191507 (quarter ended March 31, 2026): OpenAI carried under the equity method; other income (expense), net included a $5.9 billion net gain from OpenAI for the nine months then ended, primarily the dilution gain from the October 2025 recapitalization; non-GAAP measures exclude the OpenAI impact via the line “net (gains) losses from investments in OpenAI.”

(7) NVIDIA Corporation Form 10-Q, accession 0001045810-26-000052 (quarter ended April 26, 2026): net unrealized gains on publicly held equity securities $13.4 billion and on non-marketable equity $2.6 billion, recognized in other income (expense), net; net income $58.3 billion; net cash from operations $50.3 billion; net additions to non-marketable equity $17.9 billion in the quarter; $27 billion of investment commitments outstanding.

(8) Oracle Corporation Form 10-K, accession 0001193125-26-277521 (fiscal year ended May 31, 2026), and Q4 earnings release, Exhibit 99.1 to Form 8-K, accession 0001193125-26-265848: non-GAAP diluted EPS $7.63, which the company states would be $6.83 excluding investment gains; remaining performance obligations $638 billion, of which approximately twelve percent is expected to be recognized as revenue over the next twelve months; net cash from operations $32.0 billion; capital expenditures $55.7 billion; free cash flow negative $23.7 billion. The quoted risk language (”the economic returns on these investments are dependent on customer demand and the ability of our key customers to meet their contractual obligations ...”) is from the management’s discussion and analysis in the same 10-K. Forward funding, from the financing section of the same 10-K: $46,093 million of proceeds from issuances of senior notes, term loans and other borrowings, and $4,954 million of 6.50% Series D Mandatory Convertible Preferred Stock issued in fiscal 2026 (none a year earlier).

(9) CoreWeave Form 10-Q, accession 0001769628-26-000222, and the adjusted-EBITDA reconciliation in the Q1 2026 earnings release, Exhibit 99.1 to Form 8-K, accession 0001769628-26-000220: adjusted EBITDA $1,157 million, which sets aside $536 million of interest expense; net cash from operations $2,984 million, of which $2,016 million is working-capital movement; cash from operations before working capital $968 million; purchases of property and equipment $7,695 million; the $1.3 billion reclassification from deferred revenue to customer liabilities.

(10) Nebius Group Form 6-K, accession 0001104659-26-064092: a $780.6 million revaluation gain excluded from adjusted EBITDA of $129.5 million. Hyperscaler placement of marks below operating income and the absence of any EBITDA metric: Amazon accession 0001018724-26-000014; Alphabet 0001652044-26-000048; Microsoft 0001193125-26-191507; NVIDIA 0001045810-26-000052.

(11) TeraWulf Inc. Form 10-Q, accession 0001083301-26-000092, and Core Scientific, Inc. Form 10-Q, accession 0001628280-26-031396: each adds back warrant or contingent-value-rights marks in its adjusted-EBITDA definition and does not add back the fair-value remeasurement of digital assets, which therefore remains inside the metric; the remeasurement was a small loss in the quarter reported.

Nine days ago, against this same annual report, we set a second test beside the quality of earnings and called it the quality of cash: of a number presented as a present fact, how much is cash, and how certain is the cash still to come. (1) We ran it against the largest figure in the filing, the $638 billion backlog, and stopped there. This piece runs it through the rest of the document, through the cash itself.

Oracle’s statements carry an unqualified opinion from Ernst & Young, its auditor since 2002, and internal control over financial reporting was audited and found effective. (2) We take the statements as they stand and depend on them. They consolidate two very different businesses into one column because consolidation is what the standards require of a single company, and customers appear in aggregate because the aggregate is what the rules permit. What is vague in what follows is vague because the rules allow it to be. The same requirements that produce the one column leave its largest question open.

The numbers the company reports

For the fiscal year ended May 31, 2026: total revenues of $67.4 billion, up 17 percent. Net income of $17.1 billion, which includes the $2.7 billion one-time gain on the Ampere sale we read in June. (3) Net cash provided by operating activities of $32.0 billion, up 54 percent. Cash used for capital expenditures of $55.7 billion, up from $21.2 billion a year earlier, an increase the company attributes primarily to the expansion of its data centers, with the statement that “we expect this upward trend to continue during fiscal 2027 and in the following fiscal years.” (4)

Oracle publishes the difference itself. In its own liquidity discussion, in its own table: operating cash of $31,977 million, capital expenditures of $(55,663) million, free cash flow of $(23,686) million, stated in the same table as negative 139 percent of net income, against negative 3 percent the year before. (5) The company files the figure.

The whole year’s cash, both directions, sorted. Read gross from the statement of cash flows, $93.8 billion came in during fiscal 2026 and $93.8 billion went out, an identity once the addition to cash counts as a use. The sorting is the point. (6) Of the sources: $32.0 billion was operating cash, all of it; $54.4 billion was capital raised; and $7.4 billion came from asset sales, stock plans and other. Of the uses: $55.7 billion was capital expenditures; $17.6 billion was debt repaid, dividends and other; and $20.5 billion was added to cash on hand, which is how a company that produced $32 billion and spent $56 billion on the buildout ends the year with $20 billion more cash than it started with.

Two more figures from the year the income statement carries: restructuring expense of $1.8 billion, against $374 million a year earlier, under a plan sized at up to $2.1 billion whose filed rationale includes “the adoption and integration of AI technologies across certain functions.” (7) And total borrowings of $129.5 billion at year end, up from $92.6 billion. (8)

The annuity and the buildout

The statements present one company, and one column. The disclosures inside it describe two businesses, told apart by the thing this test measures: what discharges the cash. One column, then, with a core: the annuity inside, the buildout rising around it, and no filed line between them.

The first is the older of the two. Software support produced $19.8 billion of revenue in fiscal 2026, $19.5 billion the year before, $19.6 billion the year before that: flat across the three years the filing carries, discharged by the renewal of software already installed and already running. (9) Call it the annuity.

The second is the one the capital is for. Cloud infrastructure revenue grew 77 percent; cloud revenues crossed to 51 percent of the total. Fiscal 2026 is the year the new business became the majority of Oracle by revenue. (10) Its obligations are discharged differently: by delivering capacity the company is still building, in an expansion the filing states will continue, to customers the filings carry in aggregate. (4)

The cash carries no line between them. Even the sentence the MD&A uses to name the source of the money runs the two together: “Our largest source of operating cash flows is cash collections from our customers following the purchase and renewal of their cloud and software support agreements.” (11) Cloud and support, one phrase. The source is named, and the split between the two businesses stays inside it.

The word on the segment line has traveled further than the segment structure. When cloud entered these filings it named an environment; (12) by fiscal 2013 it named a subscription line beside licenses; (13) in fiscal 2026 it carries the majority of revenue and the buildout this piece prices. Each redefinition is filed, the walk deserves its own piece, and it is enough here that the cloud of the current filing is not the cloud of 2013.

The filing also states where today’s dividing lines sit, and why. Oracle reports three businesses: cloud and software, hardware, and services. The annuity and the buildout both live inside the first, so even the filing’s own three-way division stops short of the line this piece is reading for. Its segment disclosures, in the filing’s words, “align to how our chief operating decision makers (CODMs), which are our Chief Executive Officers and Chief Technology Officer, view our operating results and allocate resources.” (14) The segment standard follows management’s own lens: the filing divides where management divides, and carries together what management runs together. The one column is itself a disclosure. Oracle manages the annuity and the buildout as a single business, and files that fact.

The segment note prices the second company’s arrival inside the first. Direct expenses of the combined cloud-and-software segment ran $8.8 billion, then $10.8 billion, then $16.9 billion, up 56 percent in fiscal 2026 against revenue growth of 19 percent, and the segment’s margin moved from 64 to 63 to 59 percent. The margin nets one more line: the segment’s own sales and marketing, $7.2 billion, which fell 3 percent in the year delivery costs rose 56. The buildout is arriving in the cost of delivering, not in the cost of selling. (15) The disaggregation stops where management’s lens stops. Everything this piece measures from here, the blended margin, the aggregated customers, the unallocated capacity, is the rules working as written.

The promise, and who owes it

The naming pattern we read in June holds: the accounting compels a name on an investment and permits an aggregate on a customer, and the filing follows the accounting. (16) Underneath it sits an anchor of Oracle’s own filing, and a search we ran.

On June 30, 2025, Oracle filed a Form 8-K carrying remarks its chief executive would deliver that day: the company had signed multiple large cloud services agreements, “including one that is expected to contribute more than $30 billion in annual revenue starting in FY28.” (17) The filing stated the size and left the name to the aggregate. A year of filings later, the annual report attributes the $500 billion increase in remaining performance obligations to “certain significant cloud contracts that were entered into during the period,” and discloses, accurately, under the test the rules prescribe, which measures revenue and measures the past, that no customer accounted for 10 percent or more of total revenues in fiscal 2026, 2025, or 2024. (18) The past is diversified. The promise is $638 billion, and the promise carries no names.

We went looking for the tie, and here is the record as we found it. The parties themselves have spoken, in their own names: Oracle and OpenAI have jointly announced data-center capacity partnerships under the Stargate program, statements on the record from both sides, carrying the relationship and stopping there. (19) The reporting goes further: a widely carried figure attributes roughly $300 billion of the backlog to that counterparty. We searched for that figure, or any allocation of the backlog to any named customer, in Oracle’s filings and in both parties’ own communications, and it appears in neither; when the question was put to Oracle’s executives directly, after the reports circulated, no further detail entered the record. (20) The size is filed. The relationship is stated. The tie between them exists, on the record, nowhere.

To be plain about what that does and does not mean: the assumption that ties them is widely shared, and it is a reasonable one; the joint announcements are much of why. Nothing here questions it. The point is the category. An attribution can be well-founded and still be an estimate, and the disclosure that would make it more than an estimate has not been made, by anyone. Well-founded is one thing; filed is another. The number lives in the first, and we carry it there.

The asymmetry belongs to the rules. Revenue concentration looks backward, and the rules require it: the 10 percent test is mandatory, and Oracle passes it, disclosed. Backlog composition looks forward, and the rules leave it to the aggregate, so any allocation of the $638 billion is, necessarily, an estimate made outside the filings. One has to be disclosed; the other has to be estimated. The most consequential unknown in the filing’s largest number sits on the estimated side of that line, and the filing, lawfully, leaves it there.

What the filing does state, beside the aggregate, is its own description of who is in it: “In certain OCI offerings, we are more concentrated among a number of large customers, which could increase these risks.” (21) The shape of the exposure is filed; the names stay with the aggregate.

Paid ahead

In June we set the two disclosures of the backlog side by side, the release marking $75 billion of it “prepaid and customer supplied,” the audited statements carrying the figure whole, and read the difference as necessity: what a document must carry, against what a description wants. (22) The annual filing now supplies the audited shadow of that characterization, and it is a new line in the company’s books.

During fiscal 2026, Oracle’s operating cash included $4.6 billion from “customer prepayments with significant financing component”: payments so far ahead of delivery that the accounting treats them, in part, as borrowing, on which Oracle recognizes interest expense. The comparable figure in fiscal 2025 was zero. In fiscal 2024, zero. (23) On the balance sheet, deferred revenues rose from $10.7 billion to $15.4 billion, and the non-current portion, cash collected for delivery beyond a year out, quadrupled from $1.3 billion to $5.5 billion. (24) Customers have always paid Oracle in advance; support runs that way. Fiscal 2026 is the first year the money arrived far enough ahead, at scale enough, for the accounting to classify part of it as borrowing, and the line that carries it is new to the statements.

The furnished layer, in the same quarter, acquired a new instrument. The earnings release introduced a measure called “net cash outlay for capital expenditures,” which is capital expenditures net of short-term financing and net of customer prepayments, a lens that presents the capital spend after the customers’ contributions to it. (25) It is disclosed, defined, and reconciled, as the rules for such measures require. It is also new, and it arrived in the quarter the GAAP table printed free cash flow of negative $23.7 billion. The new measure nets; the filed statement records: $55.7 billion out for the buildout, $4.6 billion in from prepayments that carry a financing component, and the gap. Both are Oracle’s numbers.

The price of the next dollar

The gap was raised, and the filings state what the raising cost, at every rung.

The year’s gross raising was $54.4 billion: $46.1 billion of senior notes, term loans and other borrowings, net of issuance costs, including fourteen senior-note tranches totaling $43.0 billion at par; $5.0 billion of 6.50% Series D Mandatory Convertible Preferred Stock, issued February 5, 2026; and $3.3 billion of short-term financing tied to the capital expenditures themselves. (26) An at-the-market equity program of up to $20 billion was entered on February 2, 2026, and stood unused at year end. (27) The fourth-quarter call added the forward figure: roughly $40 billion of additional financing expected in fiscal 2027, against a net cash outlay for capital expenditures of roughly $70 billion, excluding $20 to $25 billion of expected customer prepayments. (28)

The fourteen senior-note tranches, as the filing lists them (29):

The maturities run to 2066. Oracle borrowed money for forty years, February 2026 to February 2066, both dates in the debt footnote, in a document whose risk factors state that committed hardware purchases carry “increased excess and obsolescence risk” the company “expect[s]... will continue,” that component lives “may differ from our expectations,” and that data-center lease terms “typically do not align with the duration and pricing of customer contracts.” (30) The instruments and the sentences share a filing.

The table carries its own reading, and the debt footnote dates it: the year’s notes came in two offerings, six tranches issued September 2025, $18.0 billion, and eight issued February 2026, $25.0 billion. (31) Set the long tenors side by side. Twenty years: 5.875 percent in September, 6.55 in February. Thirty: 5.95, then 6.70. Forty: 6.10, then 6.85. The same company, the same seniority, the same tenor, five months apart, at 67 to 75 basis points more. A coupon is a base rate plus a credit spread, and the filing leaves them undecomposed, so we do too; the company pays the coupon whichever component moved. The February notes, $25.0 billion, plus the preferred issued the same month, $5.0 billion, sum to the $30 billion the third-quarter release reported raising “on a record order book that was substantially oversubscribed.” The furnished layer reported the demand. The filed layer reports the price.

Interest expense for the year was $4.6 billion, up a billion. (32) Of the $130.1 billion of principal outstanding, $90.3 billion matures beyond fiscal 2031. (33) The schedule of the money is filed beside the depreciation schedules of the equipment.

What the auditor marked as hard

Ernst & Young’s report names one Critical Audit Matter for fiscal 2026, the designation reserved for the accounts or disclosures that involved especially challenging, subjective, or complex judgment. The matter named is uncertain tax positions: the judgment in intercompany transfer pricing. (34)

Three weeks ago we read another auditor’s report on another company in this business. Deloitte & Touche named two Critical Audit Matters at CoreWeave, and the first was the compute contracts themselves: whether the arrangements are services or leases, the judgment on which the presentation of the economics turns. The second was whether a data-center joint venture must be consolidated. (35) CoreWeave’s remaining performance obligations are $98.8 billion; Oracle’s are $638 billion, entered into substantially in one year, in the same category of business. At the smaller company, the auditor recorded that the contracts were among the hardest calls in the audit. At the larger, the hardest call recorded was taxes. Two firms, two judgments about where the difficulty lives, under one standard. The reader weighs them.

The question

The question is the one our CoreWeave piece ended on: whether a unit of compute is sold for more than it costs to deliver. There, the filings hold every input, and even there the question resists a single answer. The calculation turns on how the disclosures are read and the order in which they are assembled, and the two judgments it turns on most are whether the contracts are services or leases, and whether the financing vehicle is consolidated. The first is the judgment the auditor marked as the hardest call in the audit. (35)(36) A reader can run the calculation. Different readings run to different results. The inputs are filed; the answer is a judgment.

Here the inputs are not even separable. The segment stops at cloud-and-software combined; depreciation stays unallocated in the consolidated totals; the cost of delivering an OCI dollar sits inside the column, undivided, exactly as the standards require of two businesses run as one. The only answer anywhere on the record is furnished: management said on the fourth-quarter call that margins in the new structures are “either at or better than the prior contracts.” (37) The conclusion, without the inputs. Each posture is what the rules make available to each structure. The pure-play filed the inputs and left the judgment open; the diversified company furnished the conclusion and consolidated the inputs.

We would like to answer it. At the pure-play, the honest answer is plural: one result per reading of judgments the auditor itself marked hard. Here, the disclosure that would start the arithmetic is a segment line the rules do not require and the filings do not contain. The absence is structural: the column is built as one because the business is run as one, and the question stays open by construction for everyone outside the CODMs’ own view. The lenders stood outside it too. The only outside prices the filing prints are the capital’s, the coupons in the table above and the 6.50 percent on the preferred, and they price the column whole, never the split. That price may well be right. The filing shows only what it was set against. So the question stands where the rules leave it, in the reader’s hands, and only one thing moves it: the next annual report divides the column further, or carries it whole. The asking begins there.

NOTES

(1) Cape Fear Advisors, “Adding It Up: The Quality of Cash,” June 23, 2026, reading Oracle Corp.’s Form 10-K for the fiscal year ended May 31, 2026.

(2) Oracle Corp., Form 10-K for the fiscal year ended May 31, 2026 (filed June 22, 2026), Reports of Independent Registered Public Accounting Firm (Ernst & Young LLP), and Item 9A, Controls and Procedures.

(3) Form 10-K. Total revenues were $67,357 million, up 17%. Net income was $17,087 million, including a $2.7 billion realized gain on the Ampere transaction, read in the June 23 piece.

(4) Form 10-K, MD&A, Liquidity and Capital Resources. Quoted verbatim on the expected trend in capital expenditures.

(5) Form 10-K, MD&A. Oracle’s non-GAAP free cash flow table: net cash provided by operating activities $31,977; capital expenditures $(55,663); free cash flow $(23,686); free cash flow as a percentage of net income, (139)%, versus (3)% in fiscal 2025.

(6) Form 10-K, Consolidated Statements of Cash Flows, fiscal 2026, gross basis. Sources: operating cash $31,977; proceeds from senior notes, term loan credit agreements and other borrowings, net of issuance costs, $46,093; mandatory convertible preferred stock, net, $4,954; short-term financing related to capital expenditures, net, $3,345; issuances of common stock $1,449; sales and maturities of marketable securities and other investments $5,848; effect of exchange rates $96. Total, $93,762. Uses: capital expenditures $55,663; repayments of senior notes, term loans and other borrowings $6,942; repayments of commercial paper, net, $2,285; dividends $5,787; purchases of marketable securities and other investments and acquisitions $2,039; repurchases of common stock and shares withheld for taxes $206; other financing, net, $337; net increase in cash $20,503. Total, $93,762. Sources and uses each total $93,762 million; the equality is an identity once the increase in cash is carried as a use, and the chart’s value is the sorting. In the chart, “capital raised” combines the notes and term-loan line, the preferred, and the short-term capital-expenditure financing; “debt repaid, dividends & other” combines the remaining outflows other than capital expenditures and the addition to cash.

(7) Form 10-K, restructuring note. The 2026 Restructuring Plan, approved and initiated in fiscal 2026, with total estimated costs of up to $2.1 billion; $1.8 billion recorded in fiscal 2026. Rationale quoted verbatim in fragment.

(8) Form 10-K, notes payable and other borrowings: $129.5 billion total at May 31, 2026, versus $92.6 billion a year earlier; principal outstanding of $130.1 billion.

(9) Form 10-K, revenue disclosures: software support revenues of $19,804 million (fiscal 2026), $19,523 million (fiscal 2025), $19,609 million (fiscal 2024).

(10) Form 10-K. Cloud services revenue (SaaS and IaaS) of $33,989 million; cloud infrastructure (IaaS) growth of 77%.

(11) Form 10-K, MD&A, quoted verbatim and in full: “Our largest source of operating cash flows is cash collections from our customers following the purchase and renewal of their cloud and software support agreements.” The sentence names cloud agreements and software support agreements together; the filing does not apportion operating cash between them.

(12) Oracle Corp., Form 10-K for the fiscal year ended May 31, 2011, Item 1: “Cloud computing environments provide on demand access to a shared pool of computing resources in a self-service, dynamically scalable manner,” and “Cloud computing has evolved from technologies and services that Oracle has provided for many years.” In that filing the word describes an environment; it appears in no revenue segment name. The services business of the period included “managed cloud services.”

(13) Oracle Corp., Form 10-K for the fiscal year ended May 31, 2013: the software business’s first operating segment is “new software licenses and cloud software subscriptions,” which “includes database, middleware and applications software licenses, as well as cloud software subscriptions that provide access to a broad range of our software offerings.” The fiscal 2012 Form 10-K’s corresponding segment was “new software licenses,” with cloud appearing in a segment name for the first time in the fiscal 2013 report.

(14) Form 10-K (fiscal 2026), quoted verbatim: segment disclosures “align to how our chief operating decision makers (CODMs), which are our Chief Executive Officers and Chief Technology Officer, view our operating results and allocate resources.” The three businesses, cloud and software, hardware, and services, per Item 1 and Note 13. Segment reporting under ASC 280 follows the management approach: the reportable segments are the components the CODMs review.

(15) Form 10-K, MD&A segment results, Cloud and Software Business. Revenues of $58,530 million (fiscal 2026), $49,230 million (2025), $44,464 million (2024). Cloud and software expenses of $16,850 million, $10,827 million, and $8,783 million; sales and marketing expenses of $7,212 million (fiscal 2026) versus $7,473 million (2025), a 3% decline. Total margin of $34,468 million on $58,530 million (59%), versus $30,930 million on $49,230 million (63%) and $28,514 million on $44,464 million (64%). The margin is stated after both expense lines.

(16) The June 23 piece, on the pattern: names where the accounting compels them (Ampere, TikTok USDS Joint Venture LLC), the aggregate where it permits the group.

(17) Oracle Corp., Form 8-K, June 30, 2025 (Regulation FD). Quoted verbatim in fragment.

(18) Form 10-K, Note 1 (attribution of the RPO increase, quoted verbatim in fragment) and the concentration disclosure: “No single customer accounted for 10% or more of our total revenues in fiscal 2026, 2025 or 2024.”

(19) OpenAI and Oracle joint announcements regarding Stargate data-center capacity, including the July 2025 expansion. Statements by the parties, not filings. The announcements state the partnership; they state no dollar amount and draw no connection to the contract disclosed in the June 30, 2025 Form 8-K.

(20) The ~$300 billion figure originates in press reporting (The Wall Street Journal, September 2025) and subsequent coverage. We searched Oracle’s SEC filings and both companies’ own communications for the figure, or for any allocation of remaining performance obligations to a named customer, and found none. On Oracle’s September 2025 earnings call, following the reports, the question was put to executives and no counterparty details were provided. The counterparty, whichever company it is, files no financial statements from which the other side of the contract could be read. As with the distinction in note (22), none of this is raised to suggest the attribution is doubtful; it is raised to place the attribution in its category.

(21) Form 10-K, Item 1A, Risk Factors, quoted verbatim. The adjacent language on leveraged customers and multi-year commitments was quoted in the June 23 piece and is incorporated by reference rather than repetition.

(22) The June 23 piece, note (7) there: the $75 billion appears in the earnings release furnished June 10, 2026 (Form 8-K, Exhibit 99.1, Item 2.02), and does not appear as a line in the audited statements. The necessity reading is set out there.

(23) Form 10-K, Consolidated Statements of Cash Flows: “Increase in deferred revenues from customer prepayments with significant financing component,” $4,592 million in fiscal 2026, with no comparable amount in fiscal 2025 or 2024. The revenue note discloses the associated recognition of interest expense on the financing component.

(24) Form 10-K, balance sheet and Note 7: deferred revenues of $15,395 million ($9,916 current, $5,479 non-current) at May 31, 2026, versus $10,733 million ($9,387 current, $1,346 non-current) a year earlier.

(25) Oracle Corp., Form 8-K (Exhibit 99.1), furnished June 10, 2026, and the fiscal Q4 2026 earnings call: “net cash outlay for capital expenditures,” defined as capital expenditures less short-term financing related to capital expenditures and less customer prepayments. Furnished, on the same Item 2.02 posture as note (22). Management, on the call: “We did introduce this quarter, this net cash outlay for capital expenditures.” As with note (22), the observation concerns which document carries the measure, and nothing further.

(26) Form 10-K, Consolidated Statements of Cash Flows and MD&A. The statement line of $46,093 million includes term loan credit agreements and other borrowings and is net of issuance costs; the $43.0 billion is the stated par value of the fiscal 2026 senior notes. Mandatory Convertible Preferred Stock: 6.50% Series D, issued February 5, 2026, cash proceeds of $5.0 billion net of issuance costs. Short-term financing related to capital expenditures, net: $3,345 million. The gross capital-raised total of $54.4 billion is per note (6).

(27) Form 10-K, MD&A: equity distribution agreement entered February 2, 2026, for an at-the-market program of up to $20 billion; no shares sold under it as of May 31, 2026.

(28) Oracle fiscal Q4 2026 earnings call (furnished): approximately $40 billion of expected debt and equity financing in fiscal 2027, including the announced $20 billion at-the-market program, with no additional debt anticipated in calendar 2026; net cash outlay for capital expenditures of approximately $70 billion, excluding $20–25 billion of expected customer prepayments and timing.

(29) Form 10-K, MD&A, Liquidity and Capital Resources: “During fiscal 2026, we issued a total of $43.0 billion par value of senior notes comprising the following,” the fourteen tranches as tabled, in the filing’s own order.

(30) Form 10-K, Item 1A, Risk Factors, quoted verbatim in fragments: hardware excess-and-obsolescence risk expected to continue; useful lives and performance characteristics that may differ from expectations; lease terms that typically do not align with the duration and pricing of customer contracts.

(31) Form 10-K, Note 6, which tabulates a date of issuance and an effective interest rate for each tranche: six tranches issued September 2025 totaling $18.0 billion; eight issued February 2026 totaling $25.0 billion, including the 5.35% notes due May 2033 and the floating-rate notes at Compounded SOFR plus 1.11%. Effective rates at the long tenors: 5.91% (September 2045) against 6.59% (February 2046); 6.05% against 6.74%; 6.17% against 6.89%. Oracle fiscal Q3 2026 earnings release, March 10, 2026 (Form 8-K, Exhibit 99.1, furnished): “Within days of the announcement, Oracle raised $30 billion through a combination of investment grade bonds and mandatory convertible preferred stock, with a record order book that was substantially oversubscribed.”

(32) Form 10-K: interest expense of $4,599 million in fiscal 2026, versus $3,578 million in fiscal 2025.

(33) Form 10-K, maturities of notes payable: fiscal 2027 $7,210 million; 2028 $10,145; 2029 $5,500; 2030 $7,250; 2031 $9,750; thereafter $90,250.

(34) Form 10-K, Report of Independent Registered Public Accounting Firm: a single Critical Audit Matter, uncertain tax positions, centered on intercompany transfer pricing.

(35) Deloitte & Touche LLP, Report of Independent Registered Public Accounting Firm in CoreWeave, Inc.'s Form 10-K (fiscal 2025): two Critical Audit Matters, the first naming among the significant judgments "the determination of whether the contracts with customers are accounted for as a revenue contract for cloud-based services or a lease contract for cloud computing equipment," the second concerning the consolidation accounting for a joint venture formed to acquire and develop a multi-phase data center campus. CoreWeave's remaining performance obligations of $98.8 billion per its Form 10-Q for the quarter ended March 31, 2026. A Critical Audit Matter records where an audit's most difficult judgments fell; it is not a ranking of business risk. The comparison in the text is between the auditors' recorded judgments, and stops there.

(36) Cape Fear Advisors, “CoreWeave, Adding It Up,” June 13, 2026.

(37) Oracle fiscal Q4 2026 earnings call (furnished): management’s statement that margins on the new large contracts are “either at or better than the prior contracts.” A call remark; the Form 10-K reports margin at the cloud-and-software segment level and above.

Analysis: Cape Fear Advisors · capefearadvisors.substack.com

Correction (July 3, 2026): As published, this piece described both of Deloitte's critical audit matters at CoreWeave as concerning the compute contracts, and the second as the consolidation of the financing vehicle. The second matter concerns the consolidation of a data-center joint venture. The text has been corrected; the original wording is preserved in the pinned comment.

The quality of earnings asks whether reported income is confirmed by cash; it looks back. NVIDIA’s first quarter answers it plainly: revenue $81.6 billion, (1) operating cash flow $50.3 billion against $1.8 billion of purchases of property, equipment and intangible assets. (2) That is the backward test, and it is not the one this piece asks.

The quality of cash asks the forward one. A backlog — revenue under contract, not yet delivered — is an estimate of future cash, resting on inputs no one outside the company can observe: whether the contracts convert, at what discount, whether the customer performs. It is a promise, stated as a present fact. Against Oracle, that promise was $638 billion, and the work was to strain it. (3)

As of April 26, 2026, NVIDIA’s remaining performance obligations on contracts greater than one year were $2.6 billion. (4) Against it, the balance sheet carried $40.7 billion of accounts receivable. (5)

The two are different kinds of number. The backlog is an estimate that runs for years and turns on judgments the reader cannot check. A receivable is revenue already earned and already billed, a current asset, carried net of the company’s allowance for amounts that may not pay — an allowance that, against $40.7 billion, stands at $4 million. (6) It requires less judgment, and it allows less. The near-cash number is the large one; the estimate is the small one.

NVIDIA reports both, in both filings, in both periods. The annual report, for the year ended January 25, 2026, carried $38.5 billion of receivables against $2.3 billion of obligations; (7) the quarterly report, three months later, $40.7 billion against $2.6 billion. (5)(4) The line items are required, and the required figures carry the realized claim ahead of the promised one.

The circular financing is real — NVIDIA holds equity in companies that buy its compute, which we read in May — and the mark those stakes produce is where a soft number would sit, if one did. (8) The investments generated the bulk of $15.9 billion in other income for the quarter, lifting net income to $58.3 billion. (9) That is the number to press. It does not reach the cash; the cash flow statement removes it to arrive at operating cash. (2) The larger part is marked to quoted market prices; the rest is carried, in the company’s words, “at cost minus impairment, if any, and... adjusted for observable price changes,” moved only on an observed transaction. (10) The markup is the one mark-to-market requires; the removal from cash is the one the instruction directs; and an impairment, the filing says, would be measured the same observable way. The number is hard not by election but by rule. It moves only on what can be observed, in either direction, and only as non-cash. The one number that could soften the loop is the one the company is given no room to soften.

And the cash is there. At quarter end NVIDIA held $13.2 billion in cash and equivalents and $37.1 billion in marketable debt securities, against $259.5 billion of total assets. (11) It raised no debt in the quarter; financing activity was a net use of $21.3 billion, nearly all of it returned to shareholders, against $7.5 billion of long-term debt outstanding. (12) A separate $13 billion went out during the period for a technology license. (17) The ring such a company runs is funded from what it holds.

So we accept the position the filing shows, and read the second question a realized claim raises. The first — is it cash — the allowance answers. The second is who owes it. Three direct customers accounted for 30%, 18%, and 16% of the receivable at quarter end, two-thirds of it in three names; a quarter earlier the figures were 25%, 18%, and 13%. (13) Which of the three sit among the companies NVIDIA funds is not separately disclosed under the segment structure introduced this quarter. (8) The same small promise carries no soft backlog, and no forward floor either: almost nothing is under contract beyond the quarter. What stays open is whether the realized claim stays realized as that funded demand matures and the cadence that earns it turns. The company carries the cadence in its own words. It brings new architectures “on a one-year product cadence,” and notes that “customers may postpone purchasing new architectures or may adopt new technologies more gradually than anticipated.” (14) The same cadence runs against NVIDIA’s own books: inventory stood at $25.8 billion and accrued obligations for excess inventory purchases at $3.1 billion, the line that produced a $4.5 billion charge a year earlier when one market closed. (15) The receivable says how much has been earned. The concentration says by whom. The risk factor says the pace that earns it is the company’s own. All are disclosed. All are true. The work is to hold them at once.

Against Oracle, the near-cash portion was the smaller share of a long promise, and the company supplied the figure that sized it. (16) Against NVIDIA, it is the larger share, carried in the required filings. The same two kinds of number fall in opposite proportion.

The ring is the same in both. A company sells; the buyer it helped fund pays; the cash returns and funds the next sale. Within the accounting structure and the disclosures, that ring can be completed. It works. Nothing requires a company to complete it, and the filing does not say whether one has. Both rings run on cash; what differs is its quality — how much is earned and how much promised, how near it is, how much judgment stands between the number and the money. The next filing will move the proportion or hold it, and the asking begins there.

---NOTES

(1) NVIDIA Corp., Form 10-Q for the quarter ended April 26, 2026, and the fourth-quarter earnings release. Revenue was $81,615 million.

(2) Form 10-Q, Condensed Consolidated Statements of Cash Flows. Net cash provided by operating activities was $50,344 million; purchases of property, equipment and intangible assets were $1,757 million.

(3) Cape Fear Advisors, “Adding It Up: The Quality of Cash,” June 23, 2026, reading Oracle Corp.’s Form 10-K for the fiscal year ended May 31, 2026.

(4) Form 10-Q, revenue note. Remaining performance obligations on contracts greater than one year in length were $2.6 billion, of which approximately 40% is expected to be recognized over the next twelve months.

(5) Form 10-Q, Condensed Consolidated Balance Sheets. Accounts receivable, net, was $40,710 million.

(6) Form 10-K, Schedule II, Valuation and Qualifying Accounts. The allowance for doubtful accounts was $4 million, against accounts receivable of $40,710 million. Accounts receivable is a current asset; the company separately discloses, among its risk factors, that it may grant extended payment terms to some customers, which would lengthen collection on the affected balances.

(7) Form 10-K for the fiscal year ended January 25, 2026, Balance Sheets and revenue note. Accounts receivable, net, was $38,466 million; remaining performance obligations greater than one year were $2.3 billion.

(8) Cape Fear Advisors, “NVIDIA Q1 FY27,” May 21, 2026, which read the company’s equity investments in customers of its compute. The holdings and their measurement are disclosed in the Form 10-Q and Form 10-K investment notes. The share of revenue attributable to those customers became less separable under the segment structure introduced this quarter.

(9) Form 10-Q, Condensed Consolidated Statements of Income. Other income (expense), net, was $15,929 million; net income was $58,321 million. Net unrealized gains on publicly-held equity securities held at period end were $13.4 billion.

(10) Form 10-K, Note 1. Non-marketable equity securities are “measured at cost minus impairment, if any, and are adjusted for observable price changes in orderly transactions for a similar investment in the same issuer (the measurement alternative)”; impairment is assessed quarterly on observable factors, “including the investee’s operating performance and market trends.” Publicly-held equity securities are measured at fair value (Level 1 quoted prices or Level 2 observable inputs), with changes recognized in income as the standard for equity securities with readily determinable fair values requires; unrealized changes are non-cash and are removed in the statement of cash flows. The filing discloses that a hypothetical 10% decline in the publicly-held balance would reduce its fair value by $1.8 billion — the same measurement in the opposite direction.

(11) Form 10-Q, Balance Sheets. Cash and cash equivalents $13,237 million; marketable debt securities $37,098 million; total assets $259,474 million. Marketable equity securities, separately, were $30,237 million.

(12) Form 10-Q, Statements of Cash Flows and Balance Sheets. Net cash used in financing activities was $21,283 million, including $19,312 million of share repurchases; the statement reports no proceeds from the issuance of debt. Long-term debt outstanding was $7,470 million.

(13) Form 10-Q, revenue note, concentration of revenue and accounts receivable. Three direct customers accounted for 30%, 18%, and 16% of the accounts receivable balance as of April 26, 2026, and 25%, 18%, and 13% as of January 25, 2026. Separately, three direct customers each represented 21%, 17%, and 16% of total revenue for the quarter.

(14) Form 10-K, business description and risk factors. NVIDIA brings “new advanced architectures on a one-year product cadence, including our Rubin platform,” and notes that “customers may postpone purchasing new architectures or may adopt new technologies more gradually than anticipated.”

(15) Form 10-Q, Balance Sheets and accrued liabilities note. Inventories were $25,797 million; accrued obligations for excess inventory purchases were $3,121 million. NVIDIA recorded a $4.5 billion charge in the first quarter of the prior fiscal year, associated with H20 excess inventory and purchase obligations, after a license requirement closed its China data-center compute market.

(16) The comparison is to the character of the number, not the accounting line. Oracle’s “$75 billion” prepaid and customer-supplied hardware figure is a funded portion within its remaining performance obligations, supplied in the fourth-quarter earnings release furnished on Form 8-K (June 10, 2026). NVIDIA’s accounts receivable is a separate realized line on the balance sheet, beside its remaining performance obligations rather than within them. The two are alike in where they sit on the path from promise to cash — the near, realized end — and unlike in accounting category. The readings are directionally drawn.

(17) The $3.96 billion of accrued purchase consideration in current liabilities is the unpaid remainder of a non-exclusive license NVIDIA entered in December 2025 for the language-processing-unit technology of Groq, Inc. — under which it also hired certain Groq employees, while acquiring no equity interest, customer contracts, or existing products. The full arrangement is an outbound technology license of roughly $13 billion, recorded in investing activities, and is separate from NVIDIA’s equity-investment portfolio. Groq, Inc. is an inference-chip company, distinct from similarly named entities. We note it for size; at thirteen billion dollars it is, in this filing, a rounding entry.

When we read Oracle’s fourth-quarter results earlier this month, we left one question open: whether the annual report, when it came, would name the customer behind the backlog. It came. It does not.

The filing names plenty. Its competition section names the companies Oracle competes with, in full and by legal name, including Adobe, Amazon, Microsoft, and SAP, among others. (1) It names a transaction partner, too. On November 25, 2025, SoftBank Group Corp. acquired Ampere Computing, and Oracle received $4.3 billion in exchange for its equity, debt and call option interests, on a $2.7 billion realized gain. (2) The sale is disclosed and the buyer is named.

That is the only name from the AI-infrastructure story anywhere in the filing, and it appears on a sale. The pattern is consistent. Oracle names the companies it holds equity in: Ampere, and TikTok USDS Joint Venture LLC, a 15 percent stake that is the substantial majority of its non-marketable investments. (3) It names the buyer when it sells one. Investments and dispositions carry the name, because the accounting compels it. The customers behind $638 billion of contracted future revenue are aggregated, carried as a group.

There is a reason for it. A sale of an investment requires the filing to name the buyer. A revenue contract leaves the customer to the aggregate. The name appears where the accounting calls for it, and the customers, where it allows the group, are carried as a group.

The same logic runs through the rest of the filing. The next time it appears, it governs a number rather than a name.

The number is $638 billion. (4) It is the largest figure in the filing, and the one the market reads as proof of demand. Remaining performance obligations: revenue under contract, not yet delivered, not yet earned. Backlog.

It is a promise of future payment. The total of what customers have agreed to pay, carried forward and stated as a single figure, its value dependent on the contracts converting as written, over horizons that run past five years, at the rate and on the schedule the company expects. An estimate of future cash, presented as a present fact. A mark.

The accounting already has a standard for numbers like that. When a figure on a balance sheet is an estimate built on inputs no one can observe, fair-value measurement calls it Level 3 and asks the company to disclose the unobservable inputs and to show how the number moves when they change. The premise is that an estimate this soft is owed context; the reader is owed what it rests on. A backlog is the same kind of number, an estimate of future cash built on unobservable inputs, conversion, discount, and the customer’s performance among them. It is disclosed under the revenue standard instead, which asks for the aggregate and the timing. The same figure, under the fair-value standard, would carry the inputs and the sensitivity. The sensitivity belongs to that standard, and the backlog is filed under the other one. (5)

The discipline for reading a number that way is a generation old. It grew up around earnings, which are a record of the past: reported income could be real on the page and soft underneath, and the way to read a company was to get behind the recognized figure to the cash that stood behind it. Call it the quality of earnings. The income statement made a claim; the cash flow statement settled it. The work could be done inside one company’s filing, and it still can. That test has not retired.

The backlog adds a second number that needs the same kind of testing, and it is a different kind of number. Earnings look back; the backlog looks forward. It is a claim on cash not yet paid, owed by customers against contracts that run for years, and a claim on the future answers to a different test than a record of the past. Call it the quality of cash. It joins the quality of earnings rather than replacing it: one number for what has been earned, one for what is still to come. So the question for the backlog is the plain one. How much of the $638 billion is already backed by cash, and how much is an expectation.

Oracle answers it, in the way answers come, across more than one page. The annual report disaggregates the backlog by time. In its words, it expects to recognize “approximately 12% as revenues over the next twelve months, 34% over the subsequent month 13 to month 36, 34% over the subsequent month 37 to month 60 and the remainder thereafter.” (6) It elected an available exemption and gave the schedule anyway. The reader can see when the company expects the revenue to arrive.

That disaggregation runs by time. It stops there. The $638 billion stands as a single figure, the funded portion and the portion still to be raised held together within it.

The split does exist. Twelve days before the annual report was filed, the earnings release reported that “the prepaid and customer supplied hardware portions of our large AI contracts now total $75 billion,” and that this “substantially reduces the amount of capital Oracle must raise.” (7)

The same $638 billion, disclosed twice. The earnings release, furnished June 10, marks the $75 billion. The annual report, filed June 22, carries the figure whole.

The number is true in both places; a company puts in a release only a figure it will stand behind. What changes between the documents is its necessity. The release was describing the quarter, of a buildout that needs less outside capital than its size suggests, and the $75 billion belongs to that description. The audited statements carry what the standard asks, and the standard asks for the aggregate and the timing. The figure appears where the description called for it, and the statements stand complete without it.

So we accept the $638 billion, and the judgments folded into it. A backlog this size carries many: when the contracts convert, at what discount, whether the customer performs. Oracle discloses the schedule and stands behind the figure. We accept it.

The judgments are not all equally visible. The schedule is on the page; the discount is in the method; the amount is in the release. The one that is hardest to see is the certainty of the cash that has not yet come, owed by customers, over years. That certainty is itself a judgment, and it is the one the backlog number leaves unspoken, because a single figure carries its magnitude but not its degree of doubt.

Oracle carries that, too, in a different section. Among the risk factors, the company writes that some of its customers “may be highly leveraged and subject to their own operating and regulatory risks,” and that it “may experience risks of non-payment and non-performance.” Should its key customers prove unable to pay or otherwise perform, the company writes, it “could be locked into multi-year commitments for excess data center space and related capital expenditures, as well as associated financings, without receiving corresponding revenue.” (8)

That is the company stating, in its own words, that the certainty of the forthcoming cash is a judgment. The backlog and that sentence are in the same filing. The number says how much is coming. The risk factor says that whether it comes is a judgment about someone else. Both are disclosed. Both are true. The work is to hold them at once.

That is the test, and it travels. The question it carries into any filing of this kind is the plain one: of the backlog presented as a present fact, how much is cash, and how certain is the cash still to come. The answer to the second half lives partly outside the filing that reports the number, in the party that owes it. A reader can accept the backlog, as we accept Oracle’s, and still hold the question open. Oracle answered it where the standard let it: in the release, and in the risk factors. The question is what the test adds. The filing is where the asking begins.

NOTES

(1) Oracle Corp., Form 10-K for the fiscal year ended May 31, 2026 (filed June 22, 2026), Item 1, “Competition.” The full enumeration: “Adobe Systems Incorporated, Alphabet Inc., Amazon.com, Inc., Cisco Systems, Inc., Intel Corporation, International Business Machines Corporation, Microsoft Corporation, Salesforce, Inc. and SAP SE, as well as other companies like Hewlett-Packard Enterprise and Workday, Inc.”

(2) Form 10-K, Notes to Consolidated Financial Statements. “On November 25, 2025, SoftBank Group Corp. acquired all of the equity interests of Ampere (the Ampere Acquisition). We received cash proceeds of $4.3 billion in exchange for our equity, debt and call option interests in Ampere... We recorded $2.7 billion of realized gain.”

(3) Form 10-K, Notes. “The substantial majority of the non-marketable investments we held as of May 31, 2026 were with TikTok USDS Joint Venture LLC, an equity method investee in which we have an ownership interest of 15%.”

(4) Form 10-K. “Remaining performance obligations were $638 billion and $138 billion as of May 31, 2026 and 2025, respectively.” The same figure is reported in Oracle’s fourth-quarter earnings release, furnished on Form 8-K (Exhibit 99.1) on June 10, 2026.

(5) The comparison here is to the kind of estimate, not the accounting category. A remaining performance obligation is disclosed under ASC 606, Revenue from Contracts with Customers, which calls for the aggregate transaction price allocated to unsatisfied obligations and an explanation of when the entity expects to recognize it (606-10-50-13). It is not a fair-value measurement and is not claimed to be one. The point is only this: an RPO of this kind is an estimate of future cash resting on inputs that cannot be observed from outside the company, which is the same character of estimate that, when it appears as a fair-value measurement, ASC 820 classifies as Level 3 and subjects to disclosure of the significant unobservable inputs and a sensitivity analysis (820-10-50). The two standards govern different things and are not interchangeable. What travels between them is the nature of the number, an unobservable estimate of future cash, and the observation is that the disclosure attaching to that nature differs depending on which statement the number lands in. Oracle itself applies ASC 820 elsewhere in the same filing.

(6) Form 10-K, Note 1. Oracle elected the practical expedient available under ASC 606 that permits omitting the timing of remaining performance obligations, and provided the quantitative schedule regardless.

(7) Oracle Corp., Form 8-K (Exhibit 99.1), furnished June 10, 2026. The $75 billion figure for the prepaid and customer-supplied hardware portion appears in the earnings release. It does not appear as a line in the audited financial statements of the Form 10-K filed June 22, 2026. An 8-K earnings release is furnished rather than filed: under Item 2.02 it is “not deemed filed” for purposes of Section 18 of the Exchange Act and is not incorporated into the registration statement. The 10-K is filed and audited. The distinction is not raised to suggest the figure is doubtful in either place; it is almost certainly accurate in both. It is raised because the funding characterization that bears most directly on the quality of the cash is present in the document that carries the lighter disclosure obligation and absent from the one that carries the heavier.

(8) Form 10-K, Item 1A, Risk Factors. Quoted verbatim; the fragments are drawn from the risk factor addressing large, long-term customer arrangements.

The cash and burn

The audited balance sheet shows OpenAI ended 2025 with just over $50 billion in assets, roughly half of which is cash. Approximately $25 billion of cash against the $20.92 billion 2025 operating loss — approximately one year of runway from the audited position, on operating expense alone. Forward burn is projected at $25 to $27 billion in 2026 and $57 to $63 billion in 2027. From whence the cash is the structural question the audited document poses; the operational answer has come from subsequent funding, notably the $122 billion Series 7 closed in March 2026, led by Amazon, NVIDIA, and SoftBank.

The related party flows

The 2025 audited financials disclose related party transactions with two of OpenAI’s largest equity holders.

OpenAI paid Microsoft $17.2 billion in 2025: $10.59 billion in research and development, $6.05 billion in cost of revenue, $527 million in sales and marketing, $42 million in general and administrative. Microsoft paid OpenAI $303 million.

OpenAI’s $13.07 billion in 2025 revenue includes $303 million from Microsoft and $867 million from SoftBank. Both are equity holders. Approximately nine percent of OpenAI’s 2025 revenue came from two of its largest investors.

Half of OpenAI’s $34 billion 2025 cost stack flowed to Microsoft. The concentration sits in the operating lines: by the reported line items, approximately 81 percent of cost of revenue and 55 percent of research and development were paid to Microsoft, against 9 percent of sales and marketing and 3 percent of general and administrative.

Microsoft, the largest equity holder, receives the largest share of OpenAI’s 2025 spend and also provides revenue back to the company. SoftBank, also an equity holder, provided $867 million of OpenAI’s 2025 revenue. This is a configuration similar to ones we have read at the customer-investee scale in NVIDIA’s Three Pillars and at the supplier-investor scale in the CoreWeave reconstruction.

The accounting choices

The headline figure is $38.5 billion attributable net loss. The pre-attribution net loss is $60.4 billion. The gap between them — $21.82 billion — is two accounting allocations: $17.87 billion to noncontrolling members capital and $3.95 billion to redeemable noncontrolling interests.

Separately, the $60.4 billion pre-attribution figure itself includes a $41.55 billion non-cash fair-value adjustment tied to OpenAI’s 2025 for-profit conversion — a known, explained consequence of the conversion structure, not part of the question here. The 2025 operating loss is $20.92 billion.

Zitron flagged the noncontrolling allocations directly. “It is unclear what this means... I will not speculate further.” The mechanical effect is documented; its meaning, as Zitron said, is not. The classification choice is the same kind that shapes CoreWeave’s lease-versus-service fork.

OpenAI’s S-1 was filed confidentially on June 8. When it lands publicly, the same components will appear in sharper detail.

This piece grounds on the audited 2025 financials surfaced by Ed Zitron at Where’s Your Ed At, independently verified by the Financial Times. The structural reads extend work developed across the SpaceX series, the NVIDIA Three Pillars piece, and the CoreWeave reconstruction.

Serious people are raising questions about what’s happening at CoreWeave, and because it matters to our other work, we paused to take a look. The questions in the air are good ones. Whether the depreciation schedule runs too long. Whether the debt is too heavy. Whether the customer contracts are as solid as the backlog suggests. Whether the whole thing is a bubble. There is an instinct underneath them that something might be wrong. Read the filings for a concealed flaw and they come back clean. What is worth seeing here is in plain sight.

So we read them for something else, the same thing we read the SpaceX filings for. The interesting feature of these companies is the disclosed architecture: what the documents are built to show, and where what they show leads. With SpaceX it led to one question the filings invite and the reader has to answer alone, which was where the cash comes from. With CoreWeave it leads to a calculation the filings supply every input for and leave to the reader to run. This matters now beyond CoreWeave, because the same structure is about to be tested again by Anthropic and by OpenAI when their numbers reach a filing, and CoreWeave is the cleanest case to learn to read it on. It is already public, already audited, already rated.

So our question is narrower than the ones in the air. Set aside whether the schedule is too long, the debt too heavy, the moment a bubble. Ask instead whether a unit of compute earns more than it costs to deliver. Everything below is the company’s own disclosure, arranged until that question comes into view.

A note before the numbers, because it is the whole posture. This reconstruction relies on the figures. CoreWeave disclosed material weaknesses in internal control over financial reporting, of the kind common for a recently public company, and stated that the deficiencies did not result in a material misstatement. The auditor issued an unqualified opinion. The statements carry no going-concern qualification. We take all of it as given, and depend on it.

The numbers the company reports

For the quarter ended March 31, 2026 (Form 10-Q): revenue of $2,078 million; cost of revenue $716 million, covering facilities, rent, power, and depreciation of power installations; technology and infrastructure of $1,273 million, which the company describes as the depreciation and amortization on its servers, switches, networking equipment and software, together with research-and-development staff; sales and marketing $69 million; general and administrative $164 million; an operating loss of $144 million and a net loss of $740 million. Interest expense, net, was $536 million, with a further $97 million capitalized. Depreciation and amortization was $1.1 billion.

Property and equipment, gross, was $40,933 million, of which technology equipment is $26,627 million and construction in progress, which neither depreciates nor yet earns, is $9,581 million.

At December 31, 2025 (Form 10-K) the company ran 43 data centers with over 850 MW of active power and roughly 3.1 GW of total contracted power to deploy over future periods. It does not disclose how many GPUs it operates. It discloses the architecture: NVIDIA GB200 and GB300 NVL72 systems.

What the contracts are, and what that decides

The first judgment is whether CoreWeave sells a service or leases equipment, and the auditor, Deloitte & Touche, named it a Critical Audit Matter, defined in the opinion as a matter material to the statements that involved “especially challenging, subjective, or complex judgments.” In Deloitte’s words the significant judgments include determining “whether the contracts with customers are accounted for as a revenue contract for cloud-based services or a lease contract for cloud computing equipment, and the identification and treatment of contract terms that may impact the timing and amount” of revenue. The company concluded service, not lease: because “generally either there are no identified assets or customers do not control or direct the use of underlying hardware,” the arrangements are accounted for as service contracts under ASC 606.

That conclusion carries consequences through the statements. It keeps the GPUs on CoreWeave’s books as its own assets, depreciated by CoreWeave, with the depreciation reported in technology and infrastructure, below the gross-margin line, rather than in cost of revenue. The reported gross margin therefore does not carry the depreciation of the equipment that earned the revenue.

It also builds the collateral. Because the contracts are services and the GPUs are CoreWeave’s, both the equipment and the contracted revenue stay on CoreWeave’s side of the ledger, and both are then pledged for the debt that bought the equipment. The company describes its delayed-draw term loan facilities, $11.8 billion outstanding, as “collateralized with the assets underlying the contributed contracts and the pledged contractual cash flows, generally from investment grade counterparties,” amortizing “as contracted cash flows are generated.” The facilities run through named CoreWeave subsidiaries, CoreWeave Compute Acquisition Co. VIII, LLC among them, secured by first-priority pledges of those subsidiaries’ equity and assets, non-recourse. The subsidiaries are consolidated; the assets and debt stay on the balance sheet. So the GPUs and contracts the service judgment leaves on CoreWeave’s books are the same GPUs and contracts that secure the debt. The question this reconstruction reaches, whether a unit’s revenue covers its cost over the asset’s life, is a question about the very assets and revenue pledged as collateral.

Six years, or five

The company depreciates technology equipment over six years. Effective January 1, 2023 it changed that estimate from five years to six, citing “continuous advancements in hardware performance, software optimization, and data center design improvements,” and disclosed the effect: $20 million less expense and $0.10 more earnings per share in 2023. Over the same span the equipment’s supplier, NVIDIA, has described annual architecture generations, Hopper, Blackwell, Rubin, and stated generational efficiency gains of up to 25 times. The accounting life spans six of those generations.

The estimate is the company’s to make, and this reconstruction does not substitute its own. It records one property of it. The estimate changes reported operating income and does not change cash, because depreciation is non-cash. A longer life lowers the reported charge and lifts reported income; the cash is identical. The income statement’s profitability moves with the assumption. The cash does not.

The cash, read whole

Start with what the company reports. Operating cash was positive and rising across three years: $1,833 million in 2023, $2,749 million in 2024, $3,058 million in 2025. This is the figure the growth reading rests on, a business whose operations throw off cash, with capital spending building future capacity.

Inside that operating cash sits the prepayment. Customers pay in advance, and the cash arrives in operating activities as the deferred-revenue line: $1,986 million in 2023, $2,049 million in 2024, $4,174 million in 2025. This is an ordinary part of operating cash, recorded as the accounting requires. It is worth isolating not because it is unusual but because of what it is: cash received for compute not yet delivered, the customer financing the very service whose economics are the question here. Set it beside the operating cash. In 2023 the prepayment inflow was larger than the operating cash; in 2024 it was most of it; in 2025 it was larger again, $4,174 million against $3,058 million. Remove the prepayment from each year and the operating cash reads $(153) million in 2023, $700 million in 2024, and $(1,116) million in 2025. Across the three full years, the reported operating cash is, in net, the cash customers paid ahead of delivery, cash recorded against compute owed later, on fleet in part still to be bought.

The first quarter of 2026 reads differently. Operating cash was $2,984 million; the prepayment inflow that quarter was $575 million, a smaller share. Remove it and operating cash is still positive, about $2,409 million. The quarter rested instead on two other movements: accounts receivable came in, a $1,042 million collection, and accounts payable and accrued expenses rose, a $960 million inflow from paying vendors more slowly. A single quarter turns on whichever working-capital line happens to move; the three-year pattern is the steadier fact, and the prepayment is one working-capital line among several. The full set of them, for the quarter, is in the next section.

Against that operating cash, the fleet. Cash spent on property and equipment was $7,695 million in the quarter ended March 31, 2026, and $10,271 million across 2025. The cash earned and the cash spent on the fleet are the same compute, entered in two sections of the one statement. The investing figure runs growth and replacement together: the capacity not yet earning, and the fleet consumed earning current revenue. The split between them is where the two readings of the business diverge, and the filing carries the combined figure. The reader who wants the split assembles it.

The three years hold one shape. Operating cash positive, and made of prepayment. Investing cash larger and negative: $(3,148) million, $(8,658) million, $(10,271) million. Financing cash filling the gap each year: $1,788 million, $7,464 million, $9,308 million. The cash spent on the fleet runs ahead of the cash earned from delivering compute, and the difference is raised, every year shown, from customers paying in advance and from capital markets.

What has been paid for, and when it comes due

This would be an accounting curiosity but for what the company has already been paid. Customer prepayments, recorded as deferred revenue, were $7.5 billion. Deferred revenue is an obligation to deliver compute, measured at the cash received for it. The cash has been spent, on the fleet, in the investing line above. The obligation to deliver remains.

Remaining performance obligations were $98.8 billion, of which the company discloses 36% converts in months one through 24, 39% in months 25 through 48, and the rest in months 49 through 84. Sixty-four percent lands in year three and beyond: compute contracted, and in part prepaid, now, to be delivered across a horizon that runs well past the current equipment, on fleet not yet acquired, whose acquisition the cash the business generates has not so far funded without external capital.

The obligation is also concentrated. For the quarter, one customer was 45% of revenue and a second was 20%. A year earlier the first customer alone was 72%. The contracted revenue that secures the debt and underwrites the backlog is, by the company’s own disclosure, drawn substantially from one customer, then two.

The change in deferred revenue during the quarter was, in the company’s words, “primarily driven by reclassification of $1.3 billion to customer liabilities.” Total deferred revenue fell from $8.2 billion to $7.5 billion over the quarter; the $1.3 billion left the deferred-revenue line, much of it non-current, and appears in other current liabilities, which rose from $162 million to $1,534 million. No cash moved. What changed is the obligation’s classification, out of deferred revenue, which is discharged by delivering compute, and into a current customer liability, a nearer-dated obligation of a form the filing does not state. The filing does not give the cause, and this piece does not supply one.

In the same quarter, the other side of the customer relationship moved too. Accounts receivable fell from $3,169 million to $2,120 million, a $1,042 million collection that arrives in operating cash as an inflow. So within one quarter the company collected down a billion dollars of what customers owed it, took on $1,748 million more owed to its vendors, with accounts payable rising from $1,623 million to $3,371 million, and reclassified $1.3 billion of what it owed customers from a promise of compute into a current liability. Each of these adds to reported operating cash, and each is the timing of who pays whom and when. The filing records the movements and leaves them there.

On the balance sheet, and beside it

The company’s debt was $25,149 million as of the quarter-end balance sheet, at stated effective interest rates ranging from 2% to 15% across the individual facilities and notes, the lowest being a convertible note and the highest a delayed-draw term loan. Against that it held $2,244 million in cash, plus $777 million restricted. After the quarter closed it raised more, on top of that balance and in the same pattern: $4.0 billion of 1.75% convertible senior notes in April 2026, disclosed as a subsequent event; and in June 2026 it priced $1.25 billion of 9.625% senior notes and €2 billion of 8.500% senior notes, both due 2032, a roughly $3.5 billion raise. These are senior unsecured. The lower-cost facilities, by contrast, are in the company’s description secured by cash flows “generally from investment grade counterparties.” The counterparties are investment grade, not CoreWeave, whose own corporate credit is rated below investment grade by all three major agencies, Moody’s at Ba3, S&P at B+, and Fitch at BB-. The 9.625% coupon on the June unsecured notes is the price at which that credit raised unsecured money.

Some of the structure sits beside the balance sheet rather than on it. The company discloses an unconsolidated joint venture in which a developer holds 85% and CoreWeave 15%, where CoreWeave provides construction, administrative and property management, is the anchor tenant under a 15-year lease, guarantees up to $95 million, and may have to fund the build “to the extent the JV is unable to secure third-party financing.” Whether CoreWeave is the primary beneficiary, and so must consolidate the entity, is the second Critical Audit Matter the auditor named. The company concluded it is not, and does not consolidate.

It is not the only such arrangement. The company also discloses, under DCSP Financing Arrangements, a data center service provider that will “design, purchase, build, and manage” a data center for it, in which CoreWeave is at once the customer, the lender (a senior secured loan of up to $305 million at 13%), and the seller-lessee of $116 million of infrastructure assets it “continues to control,” recorded as a financing obligation at a 15% imputed rate. And it discloses a category of leases with developer-operators it treats as unconsolidated. The same shape recurs from the other side, too: in June 2026 a subsidiary of Prime Data Centers sought roughly $850 million in bonds to fund a build leased entirely to CoreWeave for 15 years, about $2.2 billion in rent, with CoreWeave carrying substantially all operating costs. The arrangement the consolidation judgment leaves off CoreWeave’s balance sheet is, there, recorded on someone else’s.

What the auditor marked as hard

Four judgments shape how the business reads: the contract characterization, the useful life, the classification of fleet spending, and the consolidation of the financing entity. Each is the company’s to make, the company made each one, and the auditor concurred and issued an unqualified opinion. Two of the four the auditor also singled out as Critical Audit Matters, the service-versus-lease characterization and whether to consolidate the joint venture, a designation reserved by the standard for matters that were material to the statements and involved especially challenging, subjective, or complex judgment. The independent auditor agreed with the company on both, and recorded that these two were among the hardest calls in the audit.

Each judgment carries a consequence, on the record. Treating the contracts as services keeps the equipment depreciation below the gross-margin line and keeps the assets and the contracted revenue on CoreWeave’s books, where they can be pledged. Six years rather than five lowers the annual depreciation charge and lifts reported income. Classifying the fleet spending as growth places it in investing rather than in cost of revenue. Concluding CoreWeave is not the primary beneficiary places the financing entity beside the balance sheet rather than on it. This piece sets the four down with their consequences. The reader weighs them.

The question

The company is paid, in part in advance, to deliver compute. Two-thirds of that obligation comes due in year three and beyond, from one and then two customers, on fleet still to be bought. The operating cash the business reports is in net the prepayment for that compute; the cash spent on the fleet runs ahead of the cash earned delivering it; the gap is raised each year from customers paying ahead and from capital markets.

Per-unit cost against revenue, over the life of an asset, is a calculation no company is required to publish, and its absence here is ordinary. It matters for this business because it is the calculation the answer turns on. The filings hold every input. A reader can run it.

So the question is the one the good questions circle. Set aside whether the schedule is too long, the debt too heavy, the moment a bubble. The question underneath them is whether a unit of compute is sold for more than it costs to deliver. The filings disclose everything required of them, and the calculation waits in the inputs for whoever assembles it. The machine is running. This is what it runs on.

Analysis: Cape Fear Advisors · capefearadvisors.substack.com

Sources: CoreWeave, Inc. Form 10-Q for the quarter ended March 31, 2026, and Form 10-K for the year ended December 31, 2025 (financial figures, capacity, useful-life estimate, deferred revenue, remaining performance obligations, customer concentration, debt terms, the joint venture and DCSP arrangements, and the report of independent registered public accounting firm, Deloitte & Touche LLP). April and June 2026 note offerings: CoreWeave press releases, April 14 and June 11, 2026. Corporate credit ratings (Moody’s Ba3, S&P B+, Fitch BB-): the respective agencies’ rating actions. NVIDIA architecture cadence: NVIDIA public statements, GTC 2026.