Three numbers sit on the cover of SpaceX’s June prospectus: $135.00, $0.90, and $134.10 (1). The first is what the public paid for a share. The last is what SpaceX received. The number in the middle is the underwriting discount, and it never traveled anywhere. No invoice was issued, no wire was sent. The buyers paid $135.00, the seller collected $134.10, and the ninety cents stayed with the banks that ran the sale. Across 555,555,555 shares, ninety cents is $500 million.

That is a croupier’s arithmetic: the take is lifted while the money is in motion, so the money never has to change hands at all. The filings print the mechanics in three stacked rows on page one: price to public, underwriting discounts, proceeds before expenses. The third row is the first row minus the second. Everything in this piece lives in the middle row.

The middle row, six times

Six registered offerings this year printed that row in full, and mostly the same banks ran the books.

NVIDIA sold $25.0 billion of notes in June across seven maturities, 2028 to 2056. The filed discount runs from 0.080% on the two-year notes to 0.400% on the thirty-year notes, and the cover prints each tranche’s discount in dollars; the seven filed rows sum to $47,350,000, or 18.9 basis points of the money moved (2). The representatives were Goldman Sachs, J.P. Morgan, and Morgan Stanley, and Goldman holds the largest allocation in every tranche of the filed table.

Oracle sold $25.0 billion in February: $75.0 million, 30.0 basis points (3). Broadcom sold $4.5 billion in January: $15.25 million, 33.9 basis points (4). Amazon sold $25.0 billion on July 8: a filed total of $56.1 million, 22.4 basis points (5). Alphabet sold $20.0 billion of dollar notes in February for $72.0 million, 36.0 basis points, and £5.5 billion of sterling notes the next day for £22.5 million (6). In June, Alphabet sold $18.0 billion of its own stock; the filed discount was $175.2 million, 97.3 basis points (13).

And SpaceX’s initial public offering raised $75.0 billion at the $500 million on its cover, 66.7 basis points. The option to sell $11.25 billion more carries a filed commission of zero. The same document notes the regulatory ceiling on underwriting compensation: five percent of gross proceeds. The filed rate is two thirds of one percent.

The rows line up like a rate card, and the card orders by the paper more than by the name. NVIDIA paid 18.9 basis points per dollar placed, Amazon 22.4, Oracle 30.0, Broadcom 33.9, Alphabet 36.0 in dollars. Alphabet is among the strongest credits on this page and sits at the bottom of the card; Oracle, which a July piece in this series found at the last notch of investment grade, paid six basis points less. What separates them is not the name. Alphabet’s curve runs decades longer, and the cost of years is the next section.

The card also holds an inversion worth a sentence. The type of deal that is priced as the expensive one came in below the routine one: the largest initial public offering on record cost 66.7 basis points per dollar, while a seasoned issuer’s stock sale the same month cost 97.3. At $75 billion, the filed rate followed the deal’s size rather than its type. That is a reading, not any filing’s sentence; the two rates are both filed.

One line against one quarter

Goldman Sachs reported its second quarter on July 14: revenue 39% higher than a year ago, earnings of $20.98 per share (7). One line inside the report reads equity underwriting, $985 million, up 130% from a year ago. The furnished explanation: “significantly higher net revenues from secondary and initial public offerings.”

The SpaceX cover splits its $500 million by name. Goldman’s 111,111,111 shares, at the filed ninety cents, is $100.0 million. One deal is 10.2% of that line.

Morgan Stanley reported the next morning, and its furnished release reaches for the word record for the firm, for pre-tax income, and for the division that did the placing. Equity underwriting: $851 million against $500 million a year ago, “on higher IPOs, follow-on offerings and convertibles” (8). Its SpaceX allocation matches Goldman’s, so the same $100.0 million is 11.8% of its line. The debt lines moved the same direction: Goldman’s debt underwriting up 75%, Morgan Stanley’s fixed income underwriting up 48%.

The calendar deserves its own care here. Of the tables quoted in this piece, Broadcom’s, Oracle’s, and Alphabet’s February rows belong to the banks’ first quarter; SpaceX’s, NVIDIA’s, and Alphabet’s May and June rows to the second; Amazon’s July row to the third. Every comparison in this section stays inside the second quarter, where the deal and the line share a calendar.

The placements closed in June. The banks reported in July. NVIDIA reports the quarter of its raise in late August. SpaceX’s first quarterly report as a public company is expected in mid-August. CoreWeave has not named a date. The banks’ compensation did not need to wait for any of that, because it was subtracted before the proceeds arrived. The first claim paid, on money that has not yet done anything.

What the years cost to place

Inside every discount table, one gradient repeats: the longer the money, the more it costs to place. The exhibit above plots every filed tranche discount against its years to maturity, and it carries the prior section’s finding on its face: at any single maturity the issuers sit in a narrow band, while across maturities the cost of placement rises 8.75 times, from 0.080% to 0.700%.

NVIDIA’s 2028 notes cost 8 basis points; its 2056 notes cost 40. Amazon runs 10 to 40. Oracle runs 17.5 to 45 on a curve that reaches 2066. Alphabet’s dollar curve runs 15 to 60. At the far end of everything sits a tranche from Alphabet’s February sterling deal: £1.0 billion of 6.125% notes due 2126, one hundred years out, at a filed discount of 0.700% (6). Placing two-year money for the strongest names costs 8 basis points. Placing hundred-year money costs 70. An earlier piece in this series found that the price of the seat is in the tenor. The price of placing the seat is in the tenor too, and it is filed to the decimal.

Where the middle row does not exist

Everything above comes from registered offerings, where the discount prints on the cover. The same months moved at least as much paper through private placements, where it does not, and the difference is the rulebook, not anyone’s choice. A sale under Rule 144A to qualified institutional buyers carries no public fee table, because the rule does not require one; each disclosure below is exactly what the regulations ask of its path.

SpaceX followed its IPO with $25.0 billion of senior notes in late June, sold under Rule 144A to institutional buyers (9). The closing filing names BofA, Citigroup, Goldman Sachs, J.P. Morgan, and Morgan Stanley as representatives of the initial purchasers. What no SpaceX filing contains is their compensation. The names are filed; the fee is not. There is a clock on this: SpaceX agreed to a registered exchange offer that will eventually move the notes into the filed record. The one company in this ledger whose clock has already run is Broadcom, which filed exchange offers in June 2026 for notes it sold privately in September 2021 and April 2022 (10). More than four years, in the one case the record allows a measurement.

CoreWeave’s June notes sit entirely outside the fee record, and its filings follow the same rulebook. The $1.25 billion and €2.0 billion of 9.625% and 8.500% senior notes closed June 18 under Rule 144A, and the filing names no bank in any selling capacity: no initial purchaser, no representative, no registration rights, only the trustee and its agents (11). The highest coupons in this ledger belong to the one placement whose filed record shows neither who sold it nor what the selling cost, and nothing required it to show either. CoreWeave’s secured borrowing reads differently because credit agreements attach as exhibits: MUFG and Morgan Stanley lead both facilities, with Goldman Sachs and JPMorgan Chase among the arrangers (12). The fee letters those agreements reference stay private, as fee letters do. Press coverage says J.P. Morgan led the June notes; the reader can weigh reporting as reporting.

What the record will eventually show has a shape, and it is a netting rather than a name. CoreWeave’s balance sheet already carries the aggregate: total debt principal of $25,149 million at March 31, less $290 million of unamortized discount and issuance costs (11). The August report will grow that line by roughly what the quarter’s raises cost to sell, the June notes and the May facility together, banks and lawyers and printers unseparated, less the quarter’s amortization running the other way. Because the notes carry no registration rights, that aggregate is as much light as the paper’s own terms ever call for; no future filing has to name who placed them.

Two names paid nothing at all, for opposite reasons. Berkshire Hathaway put $10 billion into Alphabet through a private placement disclosed in Alphabet’s own pricing release; a direct sale has no underwriter, so no discount exists (13). Microsoft, the strongest credit in this ledger by the agencies’ published grades, filed no offering documents in 2026. One paid nothing because its buyer needed no intermediary. One paid nothing because it did not borrow.

A billion dollars of friction

None of this is hidden. That deserves saying plainly, because this series usually works where disclosure thins out, and here disclosure is total: every registered discount is printed to the penny, tranche by tranche, on the first page of a public document. What the disclosure describes is friction.

The banks’ side of the middle row also deserves stating at full strength, because real work and real exposure sit behind it. In a firm-commitment offering, the underwriters buy the paper themselves: between pricing and settlement they own it, they bear the price risk if the market moves against them, they stabilize the aftermarket, and they carry statutory liability under the Securities Act for the registration statement they sign onto, a liability whose defense is the diligence they perform. The work shows up in the filed record itself: SpaceX’s docket carries seven free writing prospectuses across eight days of early June, including interview transcripts and country-specific materials for the United Kingdom, European, and Japanese legs of the sale (1). Their names print on the cover directly beneath the issuer’s, and the cover is where a failed deal would carry them. The discount pays for the insurance, the distribution, and the name; it is negotiated by the most sophisticated issuers alive; and by the regulator’s own yardstick it is set low, two thirds of one percent on the SpaceX cover against a ceiling of five. The friction in this piece is finely priced, and it is still friction. The premium survives undiminished at the exact end of the table where the insured risk is smallest: eight basis points to hold two-year paper for one of the strongest names on this page, for the few days between pricing and settlement.

The dollar tables quoted here sum to $884.8 million across six deals. Amazon’s July table brings the running figure to $940.9 million. The sterling table adds £22.5 million, and three more currencies of Alphabet notes, two series of Alphabet convertible preferred, and a filed commission of up to 0.5% on a $40 billion at-the-market program are still uncounted. The registered record alone shows roughly a billion dollars in six months, paid by six issuers to a rotating cast of the same five or six banks, for the act of placement.

A basis point reads like dust, and the tables are denominated to feel that way. The restoration is one multiplication: on a $25 billion deal, a basis point is $2.5 million. NVIDIA’s 18.9 quiet basis points are $47.35 million, roughly what it separately budgeted for every other cost of the offering combined, the lawyers and accountants and printers together.

The friction also stays in the paper after the cash is gone. NVIDIA’s cover prints all three columns to the dollar, tranche by tranche; summed, they read $24,964,130,000 paid by the public, $47,350,000 of discounts, $24,916,780,000 of proceeds before expenses, against $25.0 billion of principal to be repaid at maturity (2). The wedge between principal and proceeds is $83.22 million, and the cover splits it: $35.87 million is the market’s pricing of the notes, and $47.35 million is placement. The company will repay, at maturity, money it never received. Every discount in this piece has that shape: out of the system as cash on day one, still in the system as principal for two years, or thirty, or one hundred.

And it stands first in line. Before any of these companies earns back the price of its machines, and before it earns the interest and dividends that are the price of its money, its output must cover the cost of having been handed the money at all. Twenty-seven years of machine life, an earlier piece calculated, to pay for the machines and the money that bought them. The placement was paid before that clock started.

One more grading, in this series’ own currency. The Quality of Cash shelf ranks dollars by what must still happen for them to become real: NVIDIA’s receivable needs about 45 days and a solvent buyer, CoreWeave’s machines need years of earning, a mark needs a price to hold. The middle row’s dollar needs nothing further. It was collected at settlement rather than billed, it owes no future performance, and once settled it is indifferent to whether the paper that produced it performs at all. By the shelf’s measure it may be the highest-quality dollar this series has graded. Two bounds sit on the grade. The dollar renews only with the flow: both banks’ own prior-year comparators show how far the line sits below itself when the window narrows. And the same institutions sit on the lender side of the same names’ borrowing, where the exposure this fee avoids lives on: MUFG and Morgan Stanley are the named arrangers and agents of CoreWeave’s credit agreements, whose lender schedules the filed exhibits reference without reproducing. Even here, who holds the risk is referenced rather than filed. The fee dollar is pristine; the franchise around it is not unexposed.

Reading the next prints

Amazon’s $56.1 million priced on July 8, three days into the banks’ third quarter, so the next prints already hold their first entry. SpaceX’s exchange offer will someday put a registered document around $25 billion of notes, and Broadcom’s more than four years suggest the wait is measured in years. CoreWeave’s issuance-cost netting will grow in August by roughly what the quarter cost to place. Both banks described their backlogs in July as higher.

The billion is disclosed, collected, and closed. The output that must repay it has not been reported by anyone yet. Three claims now stand in line ahead of the first dollar these placements produce: the placing, the money, the machines. The first claim has been paid in full, part of it to parties no filing is required to name. Which filing, and in which August, shows the first dollar arriving?

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 above 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) SpaceX Form 424B4, filed June 12, 2026, accession 0001628280-26-042639. Cover page fee box and underwriting section, including the underwriter table and the over-allotment footnote. Roadshow record: free writing prospectuses of June 4 through June 11, 2026, accessions 0001628280-26-040610, -040874, -041013, -041150, -041365, -041761, and -042466.

(2) NVIDIA prospectus supplement, Form 424B5, June 17, 2026, accession 0001193125-26-273139. Cover fee box prints per-tranche dollar rows for price to public, underwriting discounts, and proceeds before expenses; the sums across the seven filed rows are $24,964,130,000, $47,350,000, and $24,916,780,000. Discount ladder and underwriter table from the underwriting section.

(3) Oracle Form 424B2, accession 0001193125-26-035603. Total computed from filed rates and filed tranche sizes.

(4) Broadcom Form 424B2, January 8, 2026, accession 0001193125-26-007683. Per-tranche discounts and totals as filed.

(5) Amazon Form 424B5, July 8, 2026, accession 0001104659-26-081786. Fee table totals as filed.

(6) Alphabet Forms 424B2, February 11 and February 12, 2026, accessions 0001193125-26-046846 and 0001193125-26-048730. Dollar and sterling fee tables as filed, including the 2126 tranche.

(7) Goldman Sachs Form 8-K, July 14, 2026, accession 0000886982-26-000294. Furnished under Item 2.02. Segment revenue lines and quoted language from the furnished release.

(8) Morgan Stanley Form 8-K, July 15, 2026, accession 0000895421-26-000207. Furnished under Item 2.02. Segment revenue lines and quoted language from the furnished release.

(9) SpaceX Forms 8-K of June 23 and June 26, 2026, accessions 0001628280-26-044955 and 0001628280-26-045763. Tranche schedule, Rule 144A language, and the registration rights agreement naming the initial purchasers’ representatives.

(10) Broadcom Forms 424B3, June 17, 2026, accessions 0001193125-26-274187 and 0001193125-26-274169. Exchange offers for notes issued September 30, 2021 and April 14, 2022.

(11) CoreWeave Form 8-K, June 18, 2026, accession 0001769628-26-000291. The debt-footnote netting: CoreWeave Form 10-Q for the quarter ended March 31, 2026, accession 0001769628-26-000222, total principal of debt $25,149 million less unamortized discount and issuance costs of $290 million.

(12) CoreWeave Forms 8-K, March 31 and May 18, 2026, accessions 0001769628-26-000129 and 0001769628-26-000236, and the credit agreements filed as exhibits thereto.

(13) Alphabet free writing prospectus, June 3, 2026, accession 0001193125-26-254490, and Forms 424B5 of June 2 and June 4, 2026, accessions 0001193125-26-252439, 0001193125-26-257702, 0001193125-26-257690, and 0001193125-26-256375. Common stock fee box as filed; ATM commission cap as filed; the private placement as described in Alphabet’s pricing release.

(14) Pieces in this series relied on above, cited directly: “The Price of the Seat” (July 9, 2026), https://capefearadvisors.substack.com/p/the-price-of-the-seat, for the tenor finding and the Oracle rating report it carries; “CoreWeave, Twenty-Seven Years” (July 12, 2026), https://capefearadvisors.substack.com/p/coreweave-twenty-seven-years, for the machine-life arithmetic; “NVIDIA, The Fourth House” (July 14, 2026), https://capefearadvisors.substack.com/p/nvidia-the-fourth-house, for the receivable and the treatment of marks; and the Quality of Cash shelf, https://capefearadvisors.substack.com/s/quality-of-cash, for the grading framework.

The fourth card

Days ago this series read three filings from one quarter and found one asset in three forms. Microsoft held its stake through an equity method that adds the gain to net income and excludes it from the measure the market is asked to watch. Amazon held convertible notes that release the gain through other income as conversions occur. Alphabet held an equity derivative carried at approximately nothing. One asset, three forms, three placements, and the closing line of that piece: the form decides when the gain shows.

The fourth house completes the set, and its card reads like the others. NVIDIA holds equity stakes in the labs and operators that buy its compute, and in its most recent quarter those stakes produced roughly $13.4 billion of unrealized gains on publicly held positions, flowing through other income into a $58.3 billion quarterly profit. (1) The asset class is the same one the three houses chose. The form is direct equity; the placement is the income statement; the gain shows as the marks move. On those terms NVIDIA is a house like the others, the fourth to be dealt in.

Four houses, four answers, no two alike, and the fourth answer is doubled: NVIDIA is marked by force of listing, its largest pure buyer being public, and its own adjusted measures exclude gains and losses from equity securities. (1) The mark is compulsory. The setting-aside is chosen.

The resemblance ends at the register. The first three houses placed cash into the asset class and wait, each in its chosen box, for the machines to make the placements good. The fourth house also sells the machines. That difference changes the house’s relationship to the table, and the rest of this piece prices it.

The house has four sides

Each side is filed, and three of the four are filed by the counterparty.

The first side is the register itself. CoreWeave’s annual report states the arrangement in one sentence: “as a result of our obligations in our current customer contracts, all of the GPUs used in our infrastructure today are NVIDIA GPUs.” (2) NVIDIA’s position is not CoreWeave’s procurement preference; CoreWeave’s customers wrote it into their contracts. The demand side specified the supplier, which means the one price in this system that cannot be shopped is the price of the machine. What that price contains is also filed: for the fiscal year ended January 25, 2026, NVIDIA reported $215.9 billion of revenue and $153.5 billion of gross profit, a 71 percent gross margin; in the quarter ended April 26, 2026, the margin was 75 percent. (1) Of every dollar that reaches NVIDIA, about 29 cents leaves as cost. The rest is the take, booked at sale and collected on paper the filings size at about 45 days. (6)

The second side is the equity. On January 23, 2026, CoreWeave sold NVIDIA 22,935,780 shares of Class A common stock at $87.20 per share, $2 billion in cash. (3) The transaction reconciles across three documents: the $2 billion announced in the first-quarter press release, the price and share count in the proxy statement and the annual report’s subsequent-events note, and the $1,985 million, net of issuance costs, in the audited cash flow statement. Three documents, one leg, and the $15 million between gross and net is the issuance cost the cash flow statement’s own caption names.

The third side is the order forms. Under a Master Services Agreement dated April 2023, NVIDIA is CoreWeave’s customer: CoreWeave provides “infrastructure and platform services through fulfillment of order forms submitted to us by NVIDIA,” and NVIDIA paid approximately $326.3 million under the agreement in 2025 and $59.6 million in the first quarter of 2026. (3) The seller of the machines rents back their output. The $59.6 million paid in the first quarter is about 3 percent of CoreWeave’s revenue, in a quarter whose operating result was a loss of $144 million; whatever margin the order forms carry, the quarter had no operating profit for them to exceed. (5) NVIDIA sits behind CoreWeave as a holder, across from it as a supplier, and in front of it as a customer, and the exhibit index carries the master agreement for each position.

The fourth side is not a position but a fact about the other three: no single document shows them together. NVIDIA appears nowhere in CoreWeave’s related-party footnote, because the accounting definition of a related party turns on voting influence, and NVIDIA’s stake sits below that line. The same relationships fill more than a page of the proxy statement, because the securities-law definition turns on five percent ownership, and NVIDIA owns more than five percent of its customer. (3) One relationship, two disclosure regimes, visible in one and absent from the other. The reader who checks only the financial statements meets a supplier. The reader who checks the proxy meets the house.

One proportion, filed: NVIDIA accounted for 17 percent of CoreWeave’s total purchases from suppliers in 2025, with the balance of the fleet arriving through intermediaries, and every GPU in the fleet is NVIDIA’s. (3) The direct line item understates the position; the contracts state it in full.

Two sides of one counter

This series has already priced what the machines earn. On CoreWeave’s own filed numbers, a dollar of machines returns between 77 and 99 cents over CoreWeave’s own depreciation schedule, before the cost of any money, across every construction the filings permit. (4) The machine loses on its own schedule at every corner of that grid; the player sat down at 99 cents on the dollar at its most favorable corner, short before the first card was dealt.

NVIDIA’s filed number belongs beside it. The machine that returns at most 99 cents to its owner carried a margin of roughly 71 cents to its seller on the day it was sold. Read together, the two filings describe a system that works: silicon that costs about 29 cents to produce, per dollar of sale price, goes on to generate 77 to 99 cents of operating cash for its owner. They also describe where it goes, and on what clock: nearly all of it reaches NVIDIA on receivable paper of about 45 days, and what remains reaches CoreWeave over six years, spread thinner than the machine’s own price. Days against years is the asymmetry, and both sides of it are filed. The machine costs exactly its price; the money that pays the price costs more. Add the interest on what CoreWeave raised, and the outlay passes the price while the returns stop short of it; the last piece put the roads out at fourteen years, twenty-seven, and never. The surplus and the shortfall are the same number, seen from opposite sides of the counter.

One guard travels with the comparison. NVIDIA’s margin is a company-wide figure; CoreWeave’s equipment dollars flow partly through intermediaries, as the 17 percent direct-purchase figure shows; and no dollar-for-dollar mapping between the two filers is claimed here or available in the filings. The two numbers are poles, each filed by its own filer, and the space between them is where the other participants live. What other buyers pay NVIDIA is not filed, and this piece prices only the pair that files both sides of one relationship. One more filed fact: the participant that knows both the machine’s price and its cost took the order form over owning a fleet. The choice is filed. The reason is not.

A chain follows; it is this piece’s reading, not any filing’s sentence. At filed prices, NVIDIA’s margin and CoreWeave’s shortfall are one number. A machine that returns less than a dollar cannot fund a dollar of its own replacement. So the next machine is bought from the buyers’ raises, and the raises are priced in public. The last piece measured how long the raises must continue; the tape on CoreWeave’s unsecured notes prices what continuing costs, every trading day. Read that way, the durability of NVIDIA’s order book is quoted in its customers’ cost of debt. And the same tape prices the stake: the gain NVIDIA still holds in its customers pays only if the raises keep arriving, so the order book and the equity line wait on the same quoted rate. NVIDIA’s own taxonomy bounds the chain: half of data center revenue is attributed to hyperscalers, who file no per-machine economics, and the chain reaches only the class the recast category no longer separates.

The circuit, priced

The public argument of the moment is what to call a supplier that invests in its customers. The legs are filed, and pricing them is shorter than naming them. The counter transaction is the simple one, cash on the barrel: the machine crosses one way, the cash crosses the other, and the register closes. Every leg below is that transaction stretched, in time or into another form, and the quality of cash is the measure of the stretch.

Out: CoreWeave paid its suppliers for a fleet whose in-service cost stood at $31.4 billion in March, money raised in public at $25.1 billion of principal, 59 percent of it at effective rates of 10 percent or higher, beside the equity placements and the customers’ advances. (5) Back: $2 billion of equity in January, at $87.20, and $386 million of compute rental across five quarters. What returned to CoreWeave’s side is a fraction of what crossed to NVIDIA’s, and the largest returning leg came back not as revenue or rebate but as ownership. The legs interlock in one sentence, every clause filed: the money that bought the machines included NVIDIA’s equity, the revenue that services the machines includes NVIDIA’s rent, and NVIDIA’s income includes both the machines’ price and the marks on CoreWeave’s shares. Past that sentence, the circularity is not the finding. The proportions are, and they rank NVIDIA’s interests without guesswork.

The $2 billion stake amounts to about a quarter of a single quarter’s equipment purchases by CoreWeave and about three days of NVIDIA’s quarterly net income, and it can halve or double quietly inside a sixteen-billion-dollar other-income line. The purchases are the profit: the machines’ being bought, at the margin NVIDIA files, is the income statement. If the machines repay their owners, NVIDIA’s equity compounds. If the machines merely keep being bought, NVIDIA collects again. And the two roads are not independent; NVIDIA holds the dial between them, and this is a reading, not any filing’s sentence, though both of its ends are filed. Each architecture NVIDIA ramps, in its own MD&A’s language, shortens the earning life of the fleets its stake is a claim on. The fleets’ owner files the other end: a risk factor that describes “cycling out older components of our infrastructure and replacing them with the latest technology available.” (2) The faster the machines are replaced, the more often the ante is collected, and the less the residual behind the equity is worth. The ramp is filed every quarter. Which take it favors is arithmetic. The filings describe no outcome in which the fourth house has not already been paid.

The form of the support carries its own information, all of it filed. A price cut would have printed at the register, in the margin, immediately. A loan would have printed as paper with a maturity and a rate, able to sour in public, on schedule. The equity printed as neither: no coupon, no maturity, nothing that ever comes due, an asset marked each quarter at a price anyone can check. A reader may ask what separates support with no clock from a discount with no date. The filings use neither word, and the piece prices the difference at the scale the filings can see: a discount broad enough to matter would have shown in the margin, and the margin held at 75 percent, in the quarter of the placement and in the quarter after it, each figure from filed statements. (9) A discount narrow enough to sit beneath a company-wide margin’s resolution is exactly what the equity’s form makes unnecessary.

The form also prices NVIDIA’s two kinds of dollar against each other. A revenue dollar leaves 65.6 cents of operating income by the quarter’s own statement, on 45-day paper. (1) An investment dollar carries no coupon and no maturity, returns a mark, and NVIDIA’s adjusted measures exclude the marks, sixteen billion dollars of them in the quarter just filed. The exclusion is a statement of relative worth in the filer’s own hand: revenue dollars enter every measure NVIDIA publishes; mark dollars enter only the compulsory ones. The placement spent the currency NVIDIA sets aside to defend the currency it is measured on.

Market price is market price, and the concession is made in full: a seller prices where the demand side will pay, and this demand side wrote the seller into its contracts. What the filings add to the concession is participation. NVIDIA’s money stands in the financing that met its price, by purchase agreement and by order form, each used once, each filing when used; the terms its receivable carries are where a third instrument would print, and none has. Whether either is used again, or the third appears, is not a forecast this piece makes. It is a line the next filings print on their own.

The rake, and who is paid before the hand

The three houses’ rakes are all paper until events, each waiting in the box its holder chose. In the language of the table those are secondary rakes, a named cut of the winnings, paid if the game ends with winnings to cut. Each depends on the asset class making good. The fourth house rakes the ante. The house prices the seat: the cost of sitting at this table is the machines, the margin is the rake’s take, and the register collects it in days, against players who recover in years, where they recover at all. In January the house placed a stake as well, and the stake is a secondary rake held on the terms the other three hold theirs: a cut that pays at the end, if the end pays. What separates the fourth house’s cut from the others is the sequence the last section priced. The ante was collected first, the stake crossed back to the player afterward, and the stake stands at a fraction of what the ante had already banked. The ante, not the stake, sets this house apart: it is the only take at this table collected whether or not there is ever a pot, and the only one already in the bank. This series has noted once before that the dealers who arranged CoreWeave’s two rated hands keep a ledger of their own; that ledger is still a later piece. This one stops at the house that stands on all four sides and was paid first.

The seat pricing runs in both directions, and June filed the other one. NVIDIA went to the market its customers use: $25 billion of senior unsecured notes, seven tranches reaching 2056, at yields from 4.27 to 5.63 percent, 20 to 65 basis points over Treasuries, with a use of proceeds one line long: general corporate purposes, including the repayment and refinancing of outstanding notes. (10) The notes outstanding total about $8.5 billion, so the named purpose bounds a third of the raise; the filings do not itemize the rest. This series priced that season’s seats when it priced the operator’s. Inside six months, the same $25 billion cost one issuer a covenant on its own future borrowing, cost another a pledged cash floor with a quarterly test, and cost the fourth house nothing pledged at all. Four issuers drew the same number, and the market wrote four different indentures. The fifth participant at this table never drew it and could not: CoreWeave’s entire debt stack, $25.1 billion of principal, sits within a rounding of the number, 59 percent of it at effective rates of 10 percent or higher, none of it reaching past 2032, assembled raise by raise against machines on a six-year schedule. (5) Each of the four drew in one stroke what the fifth owes in total. None of that stack is filed as the seller’s: the proxy’s related-party pages carry equity, order forms, and purchases, with no debt instrument among them. The lenders hold the years, and the seller holds the days and the upside. The fourth house’s row completes the table: a $58 billion quarterly profit beside a $25 billion raise, nothing pledged, from the same market that funds its customers’ next machines.

Reading the register’s print

NVIDIA reports first. Its fiscal quarter closes in late July and, on the cadence of its prior filings, prints in late August, within days of CoreWeave’s own second-quarter report. (7) August grades both sides of the counter, and the two prints answer different halves of the same question.

From CoreWeave, the watch list is inherited from the last piece: the ratio of EBITDA to in-service equipment, the direction the reclassified customer liabilities take, and the rate on any raise that follows. The chain above rereads that rate as NVIDIA’s own line: the durability of the order book, quoted daily in the customers’ cost of debt. One line joins the list from this piece: in the quarter of the placement, NVIDIA’s $1,985 million was the largest of the three lines that closed the gap between CoreWeave’s purchases and its operating cash flow, and whether any next gap closes with NVIDIA’s money again is a line that files itself. (5)

From NVIDIA, the print is read with this series’ oldest instrument, the quality of cash, and five lines carry it.

The receivable. Three unnamed customers hold 64 percent of a $40.7 billion balance; the balance stands at about 45 days of the quarter’s sales, and the $4 million allowance prices the worry at about one basis point. (6) The baselines to watch: whether 45 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 the register is the same days, booked as a source, somewhere else.

The portfolio. The full position is filed: publicly held equity at $39.1 billion of fair value, non-marketable positions at $42.3 billion after $17.9 billion of net additions in the quarter, $1.0 billion in equity-method funds, and $27 billion more committed for the fiscal year. (9) The whole stands near $82 billion, roughly one and a half quarters of net income; each addition is money moving from NVIDIA back into the system that buys from it, and the cadence of additions is the circuit’s pulse. One mark is computable in advance, to the dollar: 22,935,780 CoreWeave shares, at each quarter’s closing price. The baseline: marks on public positions ran 23 cents of every dollar of net income last quarter, a floor. And the two marks the filer names, $13.4 billion on public positions and $2.6 billion on stakes that have no ticker, sum past the entire $15.9 billion other-income line even at the boundaries of their rounding; everything else in the line netted below zero, and interest sits in its own captions. (1)(9) How much income arrives as marks, and how much at the counter, are two lines the same page separates.

The categories. Between the annual report filed in February and the quarterly report filed in May, one word left NVIDIA’s taxonomy. The indirect-customer sentence is otherwise unchanged: “CSPs, Neocloud builders, AI model makers, enterprises, and public sector entities” became “CSPs, AI Clouds, AI model makers, enterprises, and public sector entities.” The quarterly report adds that revenue by market platform “and the comparable periods have been recast.” (8) A named class of buyer became part of a wider one, retroactively, in a single filing. The dependence on the category is unchanged; the category has been redefined, and what the taxonomy no longer separates, only the customer-concentration disclosures still can.

The margin. Seventy-one for the year, seventy-five for the quarter. The margin is NVIDIA’s price holding: the number the demand side takes as given, and the first place any change in demand would show.

The placements, from CoreWeave’s side: each sale of its equity files its price and its buyer. NVIDIA’s January purchase is the one this piece prices; any next one prints its own comparison, and CoreWeave’s next proxy extends the order-form series.

The fourth house was paid in days, at a margin it files, before the grading of years began. When its print arrives in August, the questions are the series’ oldest ones, asked of the strongest hand at the table: how much of the quarter’s income arrived as cash from customers, how much arrived as marks on the customers’ shares, and how much of the customers’ cash began the quarter as NVIDIA’s own?

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 prior pieces in this series, hold large positions in Anthropic; 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) NVIDIA Corporation: Form 10-K for the fiscal year ended January 25, 2026, accession 0001045810-26-000021: revenue $215.9 billion, cost of revenue $62.5 billion, gross profit $153.5 billion (71.1%). Form 10-Q for the quarter ended April 26, 2026, accession 0001045810-26-000052: revenue $81.6 billion, gross profit $61.2 billion (75.0%); unrealized gains on publicly held equity positions of approximately $13.4 billion within $15.9 billion of other income; net income $58.3 billion. The investment portfolio and its placement were read in “NVIDIA, Adding It Up: The Quality of Cash” in this series. Other income (expense), net of $15,929 million and net income of $58,321 million per the Q1 FY2027 results; the CFO commentary attributes the swing to “unrealized gains in publicly-held and non-marketable equity securities,” and the company’s non-GAAP measures “exclude acquisition-related and other costs, other, gains/losses from equity securities, net, certain other income and expense” (Form 8-K furnished May 20, 2026, accession 0001045810-26-000051, press release and CFO commentary; furnished, not filed, and labeled accordingly). The Q1 FY2027 10-Q’s MD&A states operating income at 65.6% of revenue in its percentage-of-revenue table, and its segment discussion attributes the Data Center increase to growth “driven by the ramp of our Blackwell systems.” The income statement reports interest income of $540 million and interest expense of $102 million in their own captions, outside other income (expense), net, of $15,929 million; total other income, net, of $16,367 million sums the three. The two named unrealized-gain figures, $13.4 billion on publicly-held positions and $2,603 million on non-marketable positions, sum to between $15,953 million and $16,052 million at the boundaries of the publicly-held figure’s rounding, above the $15,929 million line at every point of the band; the remainder of the line therefore netted between $(24) million and $(123) million.

(2) CoreWeave, Inc. FY2025 Form 10-K, accession 0001769628-26-000104: “as a result of our obligations in our current customer contracts, all of the GPUs used in our infrastructure today are NVIDIA GPUs,” and “our current customers have contractually specified our use of NVIDIA GPUs.” From the same filing’s risk factors: “Part of this process entails cycling out older components of our infrastructure and replacing them with the latest technology available,” as quoted in “CoreWeave, Twenty-Seven Years” in this series.

(3) CoreWeave, Inc. Definitive Proxy Statement, filed April 22, 2026, accession 0001769628-26-000191, “NVIDIA Related Party Transactions”: the Master Services Agreement dated April 2023 and the order-form language verbatim; payments from NVIDIA of approximately $326.3 million (fiscal 2025) and $59.6 million (three months ended March 31, 2026); NVIDIA at 17% of total purchases from suppliers for fiscal 2025; the Securities Purchase Agreement of January 23, 2026, 22,935,780 Class A shares at $87.20 per share, $2 billion aggregate. The subsequent-events confirmation is the FY2025 10-K, Note 16. The net figure is the Q1 2026 Form 10-Q, accession 0001769628-26-000222, financing activities: common stock private placement, net of issuance costs, $1,985 million. NVIDIA is absent from the financial-statement related-party note (ASC 850 turns on voting influence and principal ownership) and present in the proxy (Item 404 turns on 5% beneficial ownership); the difference is definitional, and both documents are cited here.

(4) “CoreWeave, Twenty-Seven Years” (July 12, 2026) in this series, note 5: the full grid of constructions from the Q1 2026 Form 10-Q, accession 0001769628-26-000222, and the FY2025 Form 10-K. Six-year pre-financing return per dollar of machines: 77 cents as reported, 89 with stock compensation added back, 99 at the most favorable corner.

(5) Same sources: in-service gross equipment $31,352 million at March 31, 2026; total principal $25,149 million with 59% at effective rates of 10% or higher; revenue $2,078 million and operating loss $(144) million for the quarter ended March 31, 2026; financing detail per that piece’s notes 1 and 3. The first-quarter funding gap: purchases of property and equipment $7,695 million against operating cash flow of $2,984 million, a gap of $4,711 million, closed to within the statement’s smaller items by net new debt of $1,955 million, an $810 million draw on the company’s own cash, and the private placement, $1,985 million net, the largest of the three; gross property and equipment grew from $33,941 million to $40,933 million in the quarter. The placement carries no coupon and no maturity; the debt’s rates are the principal figures above.

(6) NVIDIA Form 10-Q for the quarter ended April 26, 2026, accession 0001045810-26-000052: accounts receivable $40.7 billion; three customers representing 64% of the balance; allowance for doubtful accounts $4 million; as read in “NVIDIA, Adding It Up: The Quality of Cash.” The receivable stands at approximately 45 days of quarterly sales: $40.7 billion against $81.6 billion of revenue in a 91-day quarter. The $4 million allowance is approximately one basis point of the balance.

(7) NVIDIA’s fiscal second quarter ends in late July; the prior-year second quarter was reported in late August, per the Form 8-K furnishing those results. CoreWeave’s second-quarter timing runs on its own prior cadence, per note 15 of “CoreWeave, Twenty-Seven Years” (July 12, 2026). Neither company had announced a reporting date as of this writing; both cadences are the filed record of prior years.

(8) The taxonomy comparison, verbatim: Form 10-K, accession 0001045810-26-000021: “indirect customers include CSPs, Neocloud builders, AI model makers, enterprises, and public sector entities.” Form 10-Q, accession 0001045810-26-000052: “indirect customers include CSPs, AI Clouds, AI model makers, enterprises, and public sector entities”; same filing: revenue by market platform “and the comparable periods have been recast,” and the hyperscaler and diversification sentences as quoted.

(9) NVIDIA Form 10-Q, accession 0001045810-26-000052, fair value and non-marketable securities notes: publicly-held equity securities at fair value of $39.1 billion at April 26, 2026 ($21.0 billion reported in marketable equity securities, $8.9 billion in other assets under lock-up through December 2027, and $9.2 billion of Level 2 warrants and preferred stock convertible to common stock in public companies), of which $27.4 billion is subject to short-term lock-up restrictions; non-marketable equity securities carried at $42.3 billion, on the filing’s own walk: balance at beginning of period $22,251 million, plus net additions of $17,899 million, plus unrealized gains of $2,603 million (recognized, per the filing’s footnote, in other income (expense), net), less reclassifications of $389 million (primarily to marketable securities following public market trading), less impairments and unrealized losses of $28 million, equals $42,336 million; equity-method investments of $1.0 billion in infrastructure funds, with maximum loss exposure including committed amounts of $2.3 billion; total investment commitments of $27 billion, subject to certain contingencies, expected to be made through the remainder of fiscal year 2027; net unrealized gains on publicly-held equity securities held at period end of $13.4 billion for the quarter. The $82 billion total is the sum of the publicly-held fair value, the non-marketable carrying value, and the equity-method balance; it excludes the $27 billion of commitments. The margin held across the placement, quarter against quarter: fiscal 2026 revenue of $215,938 million and gross profit of $153,463 million (Form 10-K, accession 0001045810-26-000021) less nine-month revenue of $147,811 million and gross profit of $102,370 million (Form 10-Q for the quarter ended October 26, 2025, accession 0001045810-25-000230) give fourth-quarter revenue of $68,127 million and gross profit of $51,093 million, a 75.0% margin for the quarter ended January 25, 2026, the quarter containing the January 23 placement; the quarter ended April 26, 2026 filed 75.0%.

(10) NVIDIA senior unsecured notes, June 2026: pricing term sheet (free writing prospectus) filed June 15, 2026, accession 0001193125-26-271326, and prospectus supplement, accession 0001193125-26-273139, registration statement 333-287619: $25,000,000,000 across seven tranches ($3.5 billion of 4.250% notes due 2028; $3.5 billion of 4.350% notes due 2029; $4.0 billion of 4.500% notes due 2031; $3.5 billion of 4.750% notes due 2033; $4.0 billion of 4.950% notes due 2036; $3.0 billion of 5.550% notes due 2046; $3.5 billion of 5.625% notes due 2056); yields to maturity 4.273% to 5.628%; spreads of 20 to 65 basis points over benchmark Treasuries; net proceeds before expenses $24,916,780,000; settled June 18, 2026 (closing Form 8-K, accession 0001193125-26-275783). Use of proceeds, per the prospectus supplement: “general corporate purposes, including the repayment and refinancing of outstanding notes.” Long-term debt outstanding at April 26, 2026: $8,470 million (Form 10-Q, accession 0001045810-26-000052). The cohort comparison is “The Price of the Seat” (this series): Oracle’s $25 billion (February 2026) carried a covenant limiting further issuance; SpaceX’s $25 billion (June 2026) carried a pledged minimum cash position and a quarterly leverage covenant; Amazon’s $25 billion priced July 8, 2026; the operator’s unsecured notes bear 9.625% due July 2032.

The quarter, in one currency

In the quarter ended March 31, 2026, CoreWeave reported revenue of $2,078 million, up 112 percent from a year earlier, and an operating loss of $144 million. With $1,147 million of depreciation and amortization added back, the quarter produced $1,003 million of EBITDA, a 48 percent margin. Reported operating cash flow was $2,984 million, but $2,016 million of that was working capital: receivables, payables, and deferred revenue moving through the quarter. Cash from operations before working capital was $968 million. Purchases of property and equipment were $7,695 million. Free cash flow was negative $4,711 million. Capital expenditure ran 7.7 times EBITDA. (1)

Depreciation is the accounting shadow of cash that already left. Lengthen the schedule and the shadow thins; the cash is unchanged. CoreWeave has pulled that lever once already: effective January 1, 2023, it extended the estimated life of its computing equipment from five years to six, and every quarter since has carried a thinner shadow per dollar of fleet than the old schedule would have cast. (6) The machines did not change when the estimate did, and neither did the cash. $7,695 million left this company in one quarter to buy equipment. The test that matters is whether the asset can return the cash and pay for the money, inside the time the asset exists.

The convergence test

What the money costs: CoreWeave’s interest expense was $427 million in the quarter, plus $97 million of interest capitalized into the asset base rather than expensed. Annualized against average debt of $23,116 million, the cost of the borrowed money is 9.1 percent. That figure is consistent with the stack’s own terms at quarter end: $25,149 million of principal across a dozen instruments, 59 percent of it at effective rates of 10 percent or higher, the highest at 15 percent, and the debt “bears interest at variable rates, the majority of which is unhedged.” On its unsecured name, in June, CoreWeave paid 9.625 percent in dollars and 8.500 percent in euros. (3)

What the assets produce: gross property and equipment stood at $40,933 million at quarter end, of which $9,581 million was construction in progress, not yet in service, earning nothing. The arithmetic excludes it. Annualized EBITDA of $4,012 million against $31,352 million of in-service gross equipment is a cash yield of 12.8 percent. The base is mostly the machines themselves: $26,627 million of it is technology equipment on the six-year schedule, and the rest, data center equipment, leasehold improvements, and software, carries filed lives from three to twelve years; note 5 prices the blend. (4)

For a depreciating asset to pay its own interest and return its principal within its life, the required annual cash yield is the cost of money plus the principal spread across the years. At 9.1 percent money:

A five-year life requires a 29.1 percent yield. A six-year life, the company’s own schedule, requires 25.7. A ten-year life, 19.1; twenty years, 14.1. The equipment produces 12.8, and the lines cross at roughly twenty-seven years. Twenty-six point eight, on the unrounded figures. (5)

The rate is not the cause. Over its scheduled life, at its current yield, a dollar of machines returns about seventy-seven cents, before the cost of any money. The shortfall is native to the machine; the cost of money only sizes it. The twenty-seven years, and the residual below, are that multiplication plus a rate.

Twenty-seven is the flat arithmetic: interest on the full dollar for the full life, and no credit for the surplus in the meantime. Shorter constructions exist, and the piece prices each of them. An amortizing lender, retiring principal dollar by dollar at the money’s own 9.1 percent, moves the crossing to about fourteen years. With the quarter’s $153 million of stock compensation added back, shares not being cash, it is eleven. On the fleet’s average size across the quarter, every assumption taken at once, the shortest crossing these filings permit is about nine years. (5) Every one of them ends past the six-year schedule, and every one leaves the machine still owing at year six: between 44 and 78 cents of the original dollar, depending on the construction. But the constructions are not equally on file. No amortizing lender appears in this stack. The instruments are bullets, 80 percent of principal due within five years, nothing reaching past 2032, so fourteen is the number for a loan structure CoreWeave has never filed. Under a bullet, the surplus must wait somewhere until the principal falls due, and the filings show where it waits: it was spent, $7,695 million of new machines in the quarter against $968 million of operating cash, every surplus dollar and roughly seven raised beside it. That deployment is itself a rate. At the company’s own schedule, a new machine returns roughly minus 7 percent a year before any financing, minus 3.4 with the stock compensation added back, so the surplus, reinvested the way the record shows, compounds downward, and no year on the calendar clears the fleet. The range is fourteen, twenty-seven, never: the lender the company never had, the stack it has, and the strategy it is executing at these yields. Twenty-seven is the moderate reading. And the flat method’s zero, the assumption that the surplus earns nothing while it waits, is the strongest corner’s own measured rate: there, a reinvested dollar returns 99 cents over its scheduled life, about minus 0.2 percent a year. The bull case’s best arithmetic prices the waiting at zero. One guard travels with all of this. The minus 7 holds at the company’s stated life; if the machines outlive the schedule, the reinvestment rescues itself, and that is the residual question arriving through another door. Every road out runs through year six.

The company’s schedule for the same machines is six years, and six is itself the product of the extension. The company’s words for the change: “reflecting continuous advancements in hardware performance, software optimization, and data center design improvements.” The same annual report’s risk factors describe the fleet’s future differently: “Part of this process entails cycling out older components of our infrastructure and replacing them with the latest technology available.” The pace of advancement lengthens the life of the asset in one note and requires the asset’s replacement in another. Both statements are filed. Neither is incorrect. They are the same fact, entered in two boxes. (6)

The residual, or who has said what the machines are worth

The same arithmetic runs backward. Held to the company’s own six-year schedule, over six years at a 12.8 percent yield, the equipment generates about 77 cents per dollar of cost in cash; the interest on the money consumes about 54 of those cents. For the lender to be made whole at year six, the machines must then be sold, or refinanced against, at roughly seventy-eight cents on the original dollar.

The company’s own depreciation schedule carries them at zero.

The question “what is a six-year-old GPU worth” is, in this record, a filed requirement. The capital structure needs the answer to be seventy-eight cents on the dollar. The books carry zero. The gap between those answers is the value of the fleet, and the tenors on file sit beside one of them. (7)

CoreWeave raised money six times in the first half of 2026: a non-recourse GPU facility in March, convertible notes and senior notes in April, a syndicated facility in May, dollar and euro senior notes in June. Every instrument matures by October 2032. The average tenor is about 5.9 years, against a stated equipment life of 6. Nothing in the debt stack reaches past 2032. Six-year money, against a six-year schedule and a twenty-seven-year requirement. (8)

Alphabet borrowed to 2126 this year, Amazon and Oracle to 2066, SpaceX to 2056. CoreWeave, whose machines must run twenty-seven years, has issued nothing that reaches past 2032, at any price the record shows. The refinancing arithmetic follows mechanically. A twenty-seven-year breakeven against a 5.9-year average tenor means the money must be raised roughly four and a half times over before the equipment has paid for itself. Each refinancing is a fresh underwriting at whatever rate the market offers that year, and the June paper, the cleanest read on the unsecured rate, came at 9.625 percent. By the second week of July the same notes, priced at par on June 11, were quoted at 96.50, a 10.42 percent yield. (9)

The price of the promise

The objection is that the customer contracts carry the value, not the machines. CoreWeave’s revenue is contracted years forward, $98.8 billion of remaining performance obligations, with investment-grade counterparties reported behind the largest commitments. On that reading, the equipment is a delivery mechanism, and the promise that matters belongs to the customer.

The income statement answers it, and the income statement is CoreWeave’s. The $98.8 billion lives in a footnote, and a remaining performance obligation is not a receivable. At this company the cash can arrive before the compute does, and some already has, in deposits and in invoicing ahead of performance. Money that arrives early sits as an obligation and discharges only as the machines deliver, and the income statement measures the discharge every ninety days. The most recent measurement is the quarter this piece opened with: revenue at the contracted rates, and 12.8 cents of annual EBITDA per dollar of in-service equipment. The contracts are not a second credit standing beside the machines. They are the machines’ output, sold forward, sometimes paid for early, and earned at the fleet’s pace and the fleet’s margins. The conversion rate is the number the twenty-seven years is built on. The objection holds on the balance sheet and in the footnotes. The income statement runs the contracts through the machines each quarter and prints what they earned.

The market priced the objection twice. Both prints are filed.

In March, CoreWeave closed an $8.5 billion facility through a ring-fenced subsidiary, non-recourse to the parent, secured by GPUs and one customer’s contract. Moody’s rated it A3. Investment grade, higher than SpaceX, higher than Oracle. In May, CoreWeave closed a $3.1 billion facility that carries everything the March facility lacked: the parent’s unconditional guarantee, subsidiary guarantees, a pledge of 100 percent of the borrower’s equity, security over substantially all assets. Its proceeds fund infrastructure dedicated to “two large, non-investment grade customers.” Moody’s rated it Ba2. Speculative grade. Same borrower family, same asset class, same banks, seven weeks apart, five notches. None of the mechanics is unusual; ring-fenced paper rating above its sponsor is what ring-fencing is for, and the two facilities differ in their customers as well as their structures. The direction is the measurement. The May facility carries strictly more of CoreWeave’s own credit support, the parent’s unconditional guarantee, the subsidiary guarantees, the full equity pledge, and it rated five notches lower. Everything the parent added did not close the gap the customers opened. The only thing that lifted the March facility to investment grade was the customer’s contract. And if the customer contract is the credit, then everything that is not the customer contract, CoreWeave’s own promise and CoreWeave’s own machines, is being priced separately, and the price is Ba2 and 2032. (10)

Both facilities were arranged by the same two banks, Morgan Stanley and MUFG, structuring on the investment-grade side in March and leading the speculative-grade syndication in May, paid on each. Five notches apart, one dealer. The role has a name at the table, the croupier, and it keeps its own ledger; this piece only notes who dealt both hands. (11)

Then, at the end of March, a counterparty acted on the inversion, and filed it. Applied Digital, CoreWeave’s landlord at Ellendale, amended its leases so that two data halls moved from CoreWeave the parent to the same ring-fenced subsidiary that borrowed at A3, with the parent demoted to a springing guaranty behind the subsidiary, plus a $50 million letter of credit. The landlord’s own 8-K explains why, in one sentence: the refinanced debt “received an investment grade credit rating of A3. These ratings compare favorably to CoreWeave Parent’s credit rating of BB.” And it draws its own conclusion: the swap is “favorable to the holders of its 9.250% notes due 2030.” The landlord told its own bondholders that replacing CoreWeave’s direct promise with a ring-fenced entity’s promise, keeping CoreWeave only as backup, improved their position. It is how one counterparty now structures around CoreWeave’s name. (12)

The same principle printed in the other direction on July 9, when S&P cut Oracle to BBB minus, the last notch of investment grade. The update’s rationale carries a section headed “OpenAI remains a key credit risk”: S&P estimates OpenAI at “roughly half of the $638 billion in RPO,” and writes that if OpenAI could not pay, Oracle “could be left with massive data center leases that it might be unable to exit or have to re-lease to new tenants under less-favorable terms.” At CoreWeave, a customer’s promise lifted ring-fenced paper five notches above the parent. At Oracle, a customer’s promise dragged an investment-grade parent to the edge of speculative grade, and Oracle’s shares rose 2.7 percent on the day of the cut. The counterparty is the credit, in both directions, priced by the agencies inside four months. And the same update, on the industry rather than the name: “the unit economics of the AI infrastructure business remain opaque.” (16)

The bull case, stated at full strength

The bull case is filed in the same 10-Q. It is three propositions, the same claim made at three dates.

First, that the machines are worth more today than the paper says. The revenue side of the argument runs through rent: the fleet earns at contracted rates set when compute was scarce, above what the spot market would pay, and will keep earning them. $9,581 million of construction in progress earns nothing yet and converts into that rent as it enters service. Revenue grew 112 percent year over year; remaining performance obligations are $98.8 billion. On this proposition the six-year schedule is not aggressive but conservative, and the books understate the fleet.

Second, that they will be worth roughly seventy-eight cents on the original dollar at year six. Not more than the schedule’s zero, which is a low bar; seventy-eight cents, which is the bar the arithmetic sets. The first proposition does not imply the second. A machine can out-earn its book today and still be worth little when its successor ships in volume; the same annual report that lengthened the life warns of exactly that cycling.

Third, that the company can keep engaging the capital markets, because the first two propositions mature more slowly than the debt does. A twenty-seven-year breakeven financed on 5.9-year money must be refinanced roughly four and a half times, and each turn reprices at whatever the market then believes about propositions one and two. The current mark is 9.625 percent, unsecured.

Two of the arithmetic’s inputs could improve, the price of money and the operating margin, and the filed numbers price both doors. Cheaper money: the risk-free rate falls to zero from 2032 onward, and CoreWeave rolls everything that matures at the best secured spread it has ever printed, the 2.25 points over the floating benchmark on the March facility. The machines still need roughly thirteen years to return their principal, against a six-year schedule, because the first six years of interest have already been paid at nine. And the spread carries a premise worth noticing: 2.25 points exists, in this record, only on paper secured by an investment-grade customer’s contract. The cheap-money door funds the third proposition with the customer’s credit, the same credit the ratings have already priced.

The margin door has a filed ceiling. Revenue per dollar of in-service equipment, at contracted rates, runs 26.5 cents a year. That is the yield of a costless company, and it clears the six-year requirement by about three months on the flat arithmetic, by about a year with the surplus credited. With selling and administration at zero, the yield is 15.8 percent, crossing at about fifteen years flat, ten credited. What the six-year schedule asks of the margin lever is nearly all of it either way: a margin over cash operating costs between 84 and 97 percent, leaving between 2.9 and 15.7 cents of each revenue dollar for power, rent, and people, against the 40.5 cents those costs take today. (2) The door exists, and it is between 0.8 and 4.2 points wide. All of these are conditional milestones, not forecasts. Each holds the fleet’s current pricing and productivity for the duration, and each is exactly as good as the quarter it is computed from.

Each proposition has a filed gauge. The third grades itself: every raise files its own rate and tenor. The second is graded by the lenders’ tenors already, and by any residual assumption that ever surfaces in a filed structure. And the first has a tell already sitting on the books.

The tell on the first is the fork this publication read in July. In the first quarter, $1.3 billion moved out of deferred revenue, the caption that resolves into revenue, and into a caption called customer liabilities that neither the quarterly nor the annual filing defines. A year earlier, a $230 million customer deposit was recharacterized as debt when the contract acquired a termination right, with a named standard and a stated cause. In that precedent, customer money changed category at the moment the contract’s terms changed. The quality of cash entered this series as a question about money still to come; here it attaches to money already received, the same dollars, revenue-shaped or debt-shaped, with the shape decided after they arrive. If the second quarter’s report shows the $1.3 billion flowing toward revenue as performance occurs, the first proposition is holding. If it hardens further toward something that services like debt, the terms are moving, in the one place the books would show it first. The fork is filed. The next filing picks a tine. (14)

One ratio grades the whole: does EBITDA per dollar of in-service gross equipment rise, quarter over quarter, faster than the useful life of the equipment falls? Both numbers are filed every ninety days. The trend so far runs the wrong way on the broader measure: 12.5 percent to 11.7 to 8.6 to 9.8 on total gross equipment across the four quarters the filings cover, with the fourth quarter derived from the annual report less its nine months. The single upturn arrives in the strongest reported quarter, and the August print says whether it was a turn or a quarter. The in-service measure is the better gauge, and the construction-in-progress conversion is the bull case’s whole first leg. That single line is the hyperscaler argument, reduced to something checkable. (13)

Two questions, answered once

CoreWeave has two questions it must keep answering, and it has answered each once. The first is accounting: matching the stated life of the machines to their actual life. It is a judgment call, and the company has made it once. Effective January 2023, five years became six. Asked once more, and solved for rather than chosen: what would the estimate have to be for the machines to pay for themselves and for the money that bought them? On the current quarter, twenty-seven years. The company’s answer is six. The arithmetic’s is twenty-seven, and never fewer than nine on any construction in the notes. Both are now on the table, and only one of them was ours.

The second is cash: how much, and for how long. That answer is on file: six raises in the first half of 2026 averaging 5.9 years at issue, a debt stack in which nothing reaches past 2032, and 59 percent of quarter-end principal priced at 10 percent or higher. Everything in this piece resolves if that financing carries the company to actual cash profitability, to quarters where operations fund the machines and the machines earn the contracts. Then the maturities are met from operations, the refinancings shrink instead of repeating, and the residual question never needs a sale to answer it. The quarter’s $7,695 million of equipment purchases sets the scale of the meantime: at that filed pace, a year of building adds roughly $30 billion of machines on the same terms, at 9 percent money against a six-year clock.

What are these machines worth in 2032? The schedule has answered. The tenors have answered. The agencies have answered twice, five notches apart, depending on whose promise stood behind the same silicon. A counterparty has begun structuring around the answer. The only answer not yet on file belongs to the equipment itself, and it reports every ninety days, in one ratio, starting with the print expected in August. (15)

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, and Amazon and Alphabet, named in prior pieces in this series, hold large positions in Anthropic. Figures above 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) CoreWeave, Inc. Form 10-Q for the quarter ended March 31, 2026, accession 0001769628-26-000222 (CIK 1769628): income statement; cash flow statement (operating cash flow $2,984M, of which the working-capital swing is $2,016M: receivables +$1,042M, payables +$960M, deferred revenue +$575M, other -$561M); purchases of property and equipment including software, $7,695M. EBITDA computed as operating loss $(144)M plus depreciation and amortization $1,147M. The receivables line reverses a year-end balance: accounts receivable stood at $1,659M at September 30, 2025, $3,169M at December 31, and $2,120M at March 31; the quarter collected what the fourth quarter had billed. The deferred-revenue line is cash motion only: the caption’s balance fell from $8,185M to $7,523M, driven, per the filing’s own walk, by the $1.3 billion reclassification to customer liabilities (note 14) and $304M of revenue recognized from the opening balance, partially offset by invoicing in advance of performance; the cash flow statement’s +$575M is the net cash the caption took in during the quarter. Financing activities, same statement: proceeds from issuance of debt, net, $3,290M; repayments of debt $1,335M; common stock issued in a private placement, net of issuance costs, $1,985M. Cash, cash equivalents, and restricted cash fell $810M in the quarter, and the working-capital swing above includes $960M of payables.

(2) Same 10-Q. Cost of revenue $716M against revenue $2,078M; the filing’s own definition places rent, utilities including power, and data-center personnel in cost of revenue, and server and equipment depreciation in technology and infrastructure expense.

(3) Same 10-Q, Note 10: interest expense $427M; interest capitalized $97M; total principal $25,149M across the facilities and notes as listed, with effective rates as disclosed; the unhedged variable-rate language verbatim. June unsecured notes: Form 8-K filed June 18, 2026, accession 0001769628-26-000291: $1,250M of 9.625% senior notes and EUR 2,000M of 8.500% senior notes, both maturing July 15, 2032, senior unsecured. Note 10’s own schedule bridges the figures: contractual interest expense of $483M, plus $41M of amortization of debt discounts and issuance costs and accretion of redemption premiums, less $97M capitalized, equals the $427M expensed. The piece’s 9.1% therefore prices interest incurred plus the amortized cost of raising the money, $483M plus $41M, annualized, over average debt. On contractual interest alone the rate is 8.4% and the flat crossing moves to about 22.5 years; the discounts and fees the $41M amortizes were cash at issuance, so the piece carries them in the cost. The income statement’s interest expense, net, $536M, is the broader caption: per the filing it is “reflected net of capitalized interest,” and interest income is reported separately, in other income. The $109M between it and Note 10’s $427M is financing cost beyond the debt schedule, which the filings do not itemize. Cash interest paid in the quarter was $364M; it differs from interest incurred by coupon timing and the capitalized share. The figures differ by caption, not by contradiction.

(4) Same 10-Q: property and equipment, gross, $40,933M; construction in progress $9,581M; the in-service base is the difference. Annualized EBITDA is the quarter’s $1,003M times four.

(5) Required-yield arithmetic, flat method: cost of money 9.1% plus 100 divided by the life in years; breakeven life equals 100 divided by the spread between produced yield (12.8%) and cost of money, roughly 26.8 years. The flat method charges interest on the full original dollar for the full life and credits the surplus with no return in the meantime; it is the arithmetic of the stack as filed, bullet maturities with nothing retired along the way. Charging every machine dollar at the debt rate is itself a choice among the sources. Equity carries no filed coupon, and the nearest filed price of unsecured CoreWeave risk is the June notes’ 10.47% trading yield (note 9); customer money, the one time the filings priced it, priced as debt: the $230 million Magnetar deposit bore $19 million of interest expense in a single quarter (FY2025 10-K; note 14). The sinking-fund alternative credits each year’s surplus against principal at the same 9.1%: the six-year requirement falls from 25.7% to 22.3%, and the breakeven from 26.8 years to about 14.2; no amortizing instrument appears in the stack (note 8). The produced yield bears every operating cost on the income statement, including the quarter’s $153M of stock-based compensation (cash flow statement); added back, as the company’s own adjusted measure does, the yield is 14.7% ($1,156M per quarter, within a million of the company’s reported $1,157M Adjusted EBITDA), for breakevens of 17.6 years flat and 11.0 credited. The denominator is quarter-end in-service equipment; averaging the quarter’s endpoints ($24,565M at December 31, being total gross of $33,941M less $9,376M of construction in progress per the FY2025 10-K property note, and $31,352M at March 31) gives $27,959M, a 14.4% yield, and breakevens of 18.9 and 11.5. The add-back and the average base together: 16.5%, and 13.4 years flat, 9.2 credited, the shortest crossing these filings permit. A third construction follows the cash flow statement: surplus reinvested in new equipment at fleet economics, an unlevered return at the stated life of roughly minus 7% as reported, minus 3.4% with stock compensation added back; at a negative reinvestment rate there is no breakeven year. Over the six-year schedule the fleet returns, per dollar of cost and before financing, about 77 cents as reported, 89 with stock compensation added back, and 99 at the strongest corner; every corner is short of the dollar. The corners meet: the strongest corner’s 99 cents over six years is an internal rate of return of roughly minus 0.2%, so the flat method’s zero on the waiting surplus is the strongest corner’s own measured rate. Unpaid principal at year six runs from 78 cents, as the piece computes it, to 44 at that corner. The base is not uniformly six-year property: of the $31,352M at March 31, technology equipment, $26,627M, carries the six-year life; software, $827M, three to six years; data center equipment and leasehold improvements, $3,878M, twelve; furniture, $20M, three to five (Q1 10-Q property note; FY2025 10-K Note 1). At each component’s own filed life, longest lives granted, the six-year requirement of 25.7% becomes a blended 24.7%, an equivalent uniform life of about 6.4 years, and the gap at the schedule narrows from 13 points to about 12; the 26.8-year crossing is unchanged, because it uses no schedule. The crossing is the least stable number in the piece: near the current spread, each added point of produced yield removes roughly five to six years, and 13.8% instead of 12.8 moves the flat crossing from 27 years to 21. The six-year multiplication is the most stable. The computation excludes construction in progress from the asset base, annualizes the strongest reported quarter, and ignores taxes. Door arithmetic, flat method throughout: cheaper money holds 9.1% through 2032 and 2.25% thereafter (the March facility’s floating spread over a zero benchmark; note 10). The margin ceiling is annualized revenue, $8,312 million, over in-service gross equipment, $31,352 million: 26.5%. Selling and administration free adds their $233 million per quarter back to EBITDA: 15.8%. The margin allowance at a six-year life is the ceiling less the requirement: 0.8 points of yield flat (2.9 cents per revenue dollar) or 4.2 credited (15.7 cents), against cash operating costs other than selling and administration of $842 million per quarter, 40.5 cents per revenue dollar. All inputs from the same 10-Q except the December 31 property detail, from the 10-K.

(6) CoreWeave FY2025 Form 10-K, accession 0001769628-26-000104, Note 1: useful lives (technology equipment six years; software three to six; data center equipment and leasehold improvements the shorter of lease term or up to twelve years), and verbatim: “Effective January 1, 2023, the Company changed its estimate of the useful life for its computing equipment utilized in data centers from five to six years, reflecting continuous advancements in hardware performance, software optimization, and data center design improvements.” The risk factor quoted appears in the same annual report.

(7) Residual arithmetic: over six years the asset yields 12.8% times six, about 77% of cost; interest at 9.1% times six consumes about 54 points; the shortfall to return principal in full is roughly 100 minus the 3.7-point annual spread times six years, about 77.6% of original cost, required as residual value at year six. The depreciation schedule runs the equipment to zero over the same six years.

(8) The 2026 instruments and maturities, all filed: DDTL 4.0, matures March 2032 (8-K March 31, 2026, accession 0001769628-26-000129); April convertible notes due October 1, 2032 and 9.75% senior notes due 2031 (10-Q Note 16); DDTL 5.0, matures November 15, 2031 (8-K May 18, 2026, accession 0001769628-26-000236); June dollar and euro notes due July 15, 2032 (8-K June 18, 2026). Average tenor at issue approximately 5.9 years. Maturity ladder per the FY2025 10-K: 80% of principal within five years, nothing past 2032.

(9) The comparison tenors are filed by the respective issuers and were set out in “The Price of the Seat” (July 9, 2026) in this series. The trading level of the June notes is market data, not a filing, per LCD (John Atkins, PitchBook/LCD, July 9, 2026): “CoreWeave’s par-priced June 11 offering of 9.625% six-year senior notes slumped to 96.50 (10.42%).” The last reported trade before publication, July 10, 2026, was at 96.28, a 10.47% yield, per FINRA TRACE (CUSIP 21873SAK4). Oracle’s share move on the day of its downgrade is per press coverage of the S&P action (Investing.com, July 9, 2026).

(10) DDTL 4.0: 8-K accession 0001769628-26-000129, Exhibit 99.1: “first investment-grade rated GPU-backed financing”; A3 (Moody’s) / A (low) (DBRS); non-recourse; “secured by substantially all assets of CoreWeave Compute Acquisition Co. VIII, LLC” and an associated customer contract; floating tranche SOFR + 2.25%. DDTL 5.0: 8-K accession 0001769628-26-000236, Exhibit 99.1: Ba2 / BB+; “unconditionally guaranteed by the Parent pursuant to a parent guarantee and pledge agreement” and by subsidiaries, secured by substantially all assets and a 100% equity pledge; proceeds supporting “customer contracts with two large, non-investment grade customers”; SOFR + 4.50%. Remaining performance obligations $98.8B per the Q1 2026 10-Q. That the March facility’s customer is investment grade is inference, labeled as such, from the rating and from the May facility’s stated contrast.

(11) Both exhibits name the arrangers: MUFG and Morgan Stanley as co-structuring agents and joint bookrunners (with Goldman Sachs) in March; Morgan Stanley and Mitsubishi UFJ Financial Group as joint lead arrangers and bookrunners in May.

(12) Applied Digital Corporation Form 8-K/A, accession 0001493152-26-014498, filed April 1, 2026 (event March 30), Item 1.01, and Form 10-Q accession 0001144879-26-000030, Note 18: the ELN-02 lease amendment suspending two of four data halls and the simultaneous re-lease of the same halls to CoreWeave Compute Acquisition Co. VIII, LLC on substantially the same terms, conterminous; the assignment of the ELN-03 lease to the same subsidiary with the parent released; the Unconditional Springing Guaranties of Payment and Performance; the $50,000,000 letter of credit; and the quoted sentences on ratings and on the transactions being favorable to the holders of the 9.250% notes due 2030. Note that “BB” is Applied Digital’s characterization of the parent rating on the Fitch scale; Moody’s rates the parent family Ba2.

(13) Same 10-Q and prior filings: EBITDA against total gross property and equipment, annualized: 12.5% (Q2 2025, $578.7M EBITDA on $18,585M of quarter-end gross), 11.7% (Q3 2025, $682.3M on $23,234M), 8.6% (Q4 2025), 9.8% (Q1 2026, $1,003M on $40,933M); directional only, with the construction-in-progress share rising across the period. The fourth quarter is not separately filed and is derived from the FY2025 10-K less the nine months ended September 30, 2025: operating income of $(46)M less $44M gives $(90)M for the quarter; depreciation and amortization of $2,454M less $1,633M gives $821M; EBITDA of $731M, annualized against $33,941M of total gross equipment at December 31.

(14) The $1.3 billion reclassification from deferred revenue to customer liabilities, its gross-versus-net presentation, and the undefined caption were set out in “CoreWeave, Taken as a Whole” (July 4, 2026) in this series; the underlying disclosure is the Q1 2026 10-Q, Contract Balances. The $230 million recharacterization is the FY2025 10-K: the MagAI Capacity Agreement amendment, the refundable deposit “now consider[ed]... in-substance debt (the ‘Magnetar Loan’) under ASC 470, Sale of Future Revenue,” with no services provided as of December 31, 2025.

(15) CoreWeave has not announced its second-quarter reporting date as of this writing. First-quarter results were reported May 7, 2026 (Form 8-K, accession 0001769628-26-000220); on that cadence the second quarter’s report is expected in August. The annualized building pace is the quarter’s $7,695 million of purchases of property and equipment, times four.

(16) S&P Global Ratings, Research Update: “Oracle Corp. Downgraded To ‘BBB-/A-3’ From ‘BBB/A-2’ On Rising Business Risk And Weaker Cash Flow; Outlook Stable,” July 9, 2026 (primary contact Andrew Chang). Verbatim: “OpenAI remains a key credit risk. We estimate that OpenAI makes up roughly half of the $638 billion in RPO... If OpenAI were unable to pay Oracle, we believe Oracle could be left with massive data center leases that it might be unable to exit or have to re-lease to new tenants under less-favorable terms.” And: “Our base case assumes that Oracle’s FOCF deficit will peak in fiscal 2027 and turn positive by 2029 as profitability ramps, but in our view, the unit economics of the AI infrastructure business remain opaque.” A rating action commentary is the agency’s published opinion, cited as such.

One asset, three forms

The asset is a position in a private frontier AI laboratory. Two laboratories matter here, OpenAI and Anthropic, and three public companies carried positions in them through the first calendar quarter of 2026: Microsoft in OpenAI; Amazon in Anthropic and, as of this quarter, OpenAI; Alphabet in a company its filings call “a private company” and “a certain strategic investment.” The three filers are the largest sellers of the compute the laboratories buy, a relationship each discloses in its own way and none of them prices here.

The three forms differ in more than accounting. A stake carries rights a note does not; a note carries seniority the preferred does not; a commitment is not yet a position at all. The reading here is narrower: whatever else a form determines, it also decides where, in each set of financial statements, the position and its gain are permitted to appear. That consequence is the subject.

Everything below is quoted from the filers’ own quarterly reports for the period ended March 31, 2026, and every figure is filed unless marked otherwise. (1)

Microsoft: removed by choice of measure

Microsoft holds approximately 27 percent of OpenAI on an as-converted basis, accounted for under the equity method. The mechanics are unusual enough that the filing explains them: income or loss is computed under the hypothetical liquidation at book value method, “because our liquidation rights and priorities differ from our underlying ownership interest.” Microsoft recognizes income based on the change in what it would receive if OpenAI’s net assets were distributed at book value. The commitment stands at $13 billion, of which $11.8 billion has been funded. (2)

In October 2025, OpenAI recapitalized into a public benefit corporation. Microsoft’s proportionate ownership went down, and a dilution gain went through other income: $5.9 billion of net gains from the OpenAI position in the nine months ended March 31, 2026, lifting net income by $4.5 billion after tax and diluted earnings per share by $0.60. Reported nine-month net income: $97,983 million.

Then the placement. Microsoft’s own non-GAAP measure, adjusted net income, is defined by a single exclusion: it “exclude[s] net gains and losses from investments in OpenAI.” Adjusted nine-month net income: $93,500 million. The position added four and a half billion dollars to the number GAAP produces, and the number Microsoft promotes removes it.

The measure is symmetric. A year earlier the same line ran the other way: the nine months ended March 31, 2025 carried roughly $2.0 billion of OpenAI losses, and adjusted net income was higher than reported. The measure removed losses then and removes gains now. Microsoft has built a wall, in both directions, between the laboratory and the earnings figure it asks the market to watch: whatever OpenAI does to the income statement, the headline is defined so that it does not move. (3)

Amazon: delivered by choice of instrument

Amazon invested $8.0 billion in Anthropic convertible notes between the third quarter of 2023 and the fourth quarter of 2025. The notes are classified as available-for-sale, reported at fair value as Level 3 assets, and their unrealized gains rest in accumulated other comprehensive income, a balance-sheet account that touches the income statement only when something happens.

Things have been happening. In the first quarters of 2025 and 2026, portions of the notes converted to nonvoting preferred stock, and the conversions reclassified $3.3 billion and $4.5 billion of parked gains into “Other income (expense), net.” In the first quarter of 2026, Amazon also recorded “an upward adjustment of approximately $12.3 billion to our nonvoting preferred stock in ‘Other income (expense), net’ to reflect observable changes in price.” The quarter’s net other-income gain, $15.6 billion, is by the company’s own description primarily Anthropic. (4)

The carrying values, against the basis: the notes stood at approximately $42.2 billion at March 31, 2026, the nonvoting preferred at approximately $32.0 billion. Call it $74.2 billion of carrying value against $8.0 billion invested. The unrealized gain still parked in accumulated other comprehensive income, after the reclassifications: $36.3 billion. That is the reservoir. Each future conversion or liquidity event moves some of it from the parking account into income.

The position is still growing. Subsequent to quarter end, Amazon invested a further $5.0 billion in Anthropic nonvoting preferred, amended its commercial arrangement covering AWS cloud services and the performance of AWS chips, and made available to Anthropic a financing facility of up to $20.0 billion, expiring thirty months after a liquidity event such as an initial public offering, with an option to invest up to $5.0 billion more in future equity financings. (5)

And this quarter the second laboratory arrived on the same balance sheet. Amazon invested $15.0 billion in OpenAI Series C preferred and signed a commitment letter for a further $35.0 billion, an obligation that becomes mandatory upon the earlier of specified milestones or a public listing, and terminates if unspent by the end of 2028. Amazon’s carrying value in equity investments in private companies went from $16.2 billion to $48.1 billion in ninety days. (6)

Amazon’s laboratory gains run through the income statement, twelve billion dollars in a quarter when a price is observed, three to four billion at a time when notes convert, with a thirty-six billion dollar reservoir queued behind them. Available-for-sale accounting and the measurement alternative for private stakes are standard. The instrument chose the placement, and the placement puts the laboratory inside the earnings the market reads.

Alphabet: not material

Alphabet’s first-quarter derivatives table carries a line that did not exist at year end. Equity derivatives, gross notional: zero at December 31, 2025; $30,000 million at March 31, 2026. The note explains: “The notional amounts for equity derivatives represent an agreement for future capital funding in the form of notes receivable or equity to be funded in multiple tranches contingent upon the achievement of specified operational and financial milestones through 2030. The fair value of these equity derivatives was not material as of March 31, 2026.” (7)

The commitments note gives the full shape: “a $40.0 billion investment in a private company consisting of a $10.0 billion capital commitment and $30.0 billion of future capital funding contingent upon the achievement of specified operational and financial milestones through 2030, which is accounted for as an equity derivative. We expect to fund $10.0 billion in the second quarter of 2026 in the form of a non-marketable security.” Alphabet’s total future funding commitments moved from $1.1 billion to $40.7 billion in the same ninety days. (8)

What the filing does not carry is a name. The counterparty is “a private company,” “a future private investment,” “a certain strategic investment.” A full-text search of the quarterly report returns zero instances of Anthropic and zero of OpenAI. The identification of the counterparty as Anthropic is reported and inferred, not filed: press reporting placed Alphabet’s planned Anthropic investment at $40 billion in the same weeks, court documents put Alphabet’s existing ownership near 14 percent, and the filed structure matches the reported one. The inference is strong, and Alphabet’s own paper is the reason it must remain one. (9)

Microsoft’s laboratory position produces income the company then excludes, visibly, with a reconciliation table. Amazon’s produces income the company reports, at billions per quarter. Alphabet’s produces, so far, a notional amount in a table and a fair value described as not material. A $40.0 billion commitment to the most consequential private company in its orbit enters the record as a derivative worth approximately nothing, held against a counterparty with no name. At inception a derivative’s fair value can sit near zero, and the funding has not yet flowed. The visibility is the consequence.

Readers of the backstop work will recognize the move, because it sits one row away in the same table. Alphabet’s data-center lease backstops are accounted for as credit derivatives: $28.4 billion of gross notional at March 31, 2026, carried at $339 million of fair value. Its energy-infrastructure backstops, in the same filing, sit under the financial-guarantees caption, with the “maximum potential amount of future payments” stated: $9.0 billion. Same filer, same quarter, and the characterization decides what a reader can see. (10)

The counterexample is filed by another house entirely. On June 8, Broadcom arranged for an investor partner to take on agreements to purchase AI racks Broadcom designed, and the leases of those racks to a customer, and then backstopped the customer’s lease obligations itself: “a backstop agreement with the investor partner for the customer’s lease obligations over 5-year terms,” rising as racks deploy, falling as the customer pays, “with a maximum exposure of $29 billion.” The remedies are stated: assume the lease, or sell the racks. Neither the customer nor the investor partner is named. The names stayed out of Broadcom’s report too. The number could not. A capped backstop is a different instrument from a credit derivative, and part of the gap between the two disclosures is the difference between the two things. Grant that fully and set the pages side by side anyway. Broadcom stands behind a stated maximum of $29 billion of AI-lease obligations. Alphabet’s data-center lease backstops carry $28.4 billion of notional, a figure its own note defines as the maximum potential amount of future payments. Nearly the same exposure, two forms: one page carries the maximum, the other carries $339 million of fair value. (11)

The fork, transposed

Set the three placements side by side, as of the same ninety days:

Microsoft: the gain is in net income, and the promoted measure removes it. A reader of the adjusted number sees no laboratory at all, by symmetric design.

Amazon: the gain is in net income, $12.3 billion of it in one quarter on an observable price, with $36.3 billion staged in an equity account for future quarters. A reader of the headline sees the laboratory every time something happens, and something has happened three quarters running.

Alphabet: the gain, if there is one, is nowhere. A $40.0 billion commitment is a notional in a footnote at a fair value not material, and the counterparty is unnamed. A reader of the headline sees nothing, and a reader of the footnotes sees a shape without a face.

The instrument decides for Amazon. The measure decides for Microsoft. The characterization decides for Alphabet. What a shareholder in each company knows about the same asset, in the same quarter, from the same class of document, differs by an order of magnitude, and in Alphabet’s case by the presence or absence of the asset itself.

The gains are marks on private shares, and the marks move on observable prices, which for a private company means financings. Amazon’s $12.3 billion adjustment reflects observable changes in price, and after quarter end Amazon invested $5.0 billion more. Alphabet expects to fund $10.0 billion in the second quarter. The same houses that carry the marks are buyers in the rounds where private prices are set. Every one of those facts is filed by the house that carries it.

The question

The laboratories are private. Their prices are set in rounds, and the rounds are funded in large part by the three houses above. Three pieces of paper carry the positions, and none of them was dealt. Each house chose its card when it sat down to play, and each card shows only when the rules written on it require. An equity-method stake shows every quarter, and Microsoft publishes a second number that does not. A note shows when it converts or when a price is observed, so Amazon’s quarter carried sixteen billion. A derivative shows when its fair value turns material, a threshold the funding schedule has not yet crossed.

The quality of cash entered this series as a question about a backlog: how much of a number still to come is already backed by cash, and how much is expectation. The three houses answer it from the other side of the table, because their future cash requires no such test. Alphabet’s $30.0 billion is contingent and unfunded, and it has already bought a characterization. Amazon’s $35.0 billion is a commitment letter, and it is already mandatory upon a milestone. At this end of the ladder the quality of cash is a currency in itself: the forms above were not bought with cash that moved. They were bought with cash that could.

When the same asset, in the same quarter, can add four and a half billion to one company’s earnings and be removed, add sixteen billion to another’s and be kept, and rest at approximately zero on a third’s, what does a reader hold who holds the headline number? And when the next round prices, whose number will it appear in, whose number will it be defined out of, and whose page will it never reach? Each house answered before the gain existed, when it chose its card. The answers are in the paper, and so are the next ones.

DISCLOSURE, standing: Anthropic is the developer of Claude, which is used in preparing this research. Amazon and Alphabet hold large positions in Anthropic. Figures above 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; quarterly reports are cited by accession number. Analysis: Cape Fear Advisors.

NOTES

(1) Microsoft Corporation Form 10-Q for the quarter ended March 31, 2026, accession 0001193125-26-191507 (fiscal third quarter). Amazon.com, Inc. Form 10-Q for the quarter ended March 31, 2026, accession 0001018724-26-000014. Alphabet Inc. Form 10-Q for the quarter ended March 31, 2026, accession 0001652044-26-000048. All figures in this note derive from these three filings except where marked reported or inferred.

(2) Microsoft 10-Q, Note 1, Investments: “We have an investment of approximately 27 percent of OpenAI on an as-converted basis accounted for under the equity method of accounting... We calculate our equity method income or loss using the hypothetical liquidation at book value (’HLBV’) method because our liquidation rights and priorities differ from our underlying ownership interest... We have made total funding commitments of $13 billion, of which $11.8 billion has been funded as of March 31, 2026.” The October 2025 recapitalization and dilution gain per the same note and Note 3.

(3) Microsoft 10-Q, MD&A and Non-GAAP Financial Measures: “These non-GAAP financial measures exclude net gains and losses from investments in OpenAI.” Reconciliation table, nine months ended March 31: net income $97,983 million (2026) and $74,599 million (2025); net (gains) losses from investments in OpenAI, net of tax, $(4,483) million (2026) and $2,045 million (2025); adjusted net income $93,500 million (2026) and $76,644 million (2025). MD&A: nine-month other income included $5.9 billion of net gains from investments in OpenAI, primarily the dilution gain from the OpenAI recapitalization, increasing net income by $4.5 billion and diluted EPS by $0.60.

(4) Amazon 10-Q, Note on investments: “From Q3 2023 to Q4 2025, we invested $8.0 billion in convertible notes from Anthropic, which are classified as available-for-sale and reported at fair value... and as Level 3 assets... a gain of approximately $3.3 billion and $4.5 billion was recorded in ‘Other income (expense), net’ [on the Q1 2025 and Q1 2026 conversions]. In Q1 2026, we also recorded an upward adjustment of approximately $12.3 billion to our nonvoting preferred stock in ‘Other income (expense), net’ to reflect observable changes in price.” MD&A: “The net gain of $15.6 billion in Q1 2026 is primarily from an upward adjustment for observable changes in price relating to our nonvoting preferred stock in Anthropic and the reclassification adjustment for the gains on available-for-sale debt securities.” Derivation of the $74.2 billion: the notes ($42.2 billion) are an available-for-sale debt line; the nonvoting preferred ($32.0 billion) sits within “equity investments in private companies not accounted for under the equity-method” ($48.1 billion at March 31, 2026, which the filing states relates primarily to the Anthropic preferred and the OpenAI preferred, $15.0 billion). Converted paper is recorded in the equity line at conversion-date fair value and simultaneously leaves the notes line, so no portion of the position appears in both lines; $42.2 billion plus $32.0 billion states the whole Anthropic position once.

(5) Amazon 10-Q, subsequent events: the $5.0 billion further investment; the amended commercial arrangement, “primarily for the provision of AWS cloud services,” including “contractual obligations related to the performance of AWS chips”; the financing facility “not to exceed $20.0 billion that will expire 30 months after a liquidity event, as defined, such as an Anthropic initial public offering or direct listing”; and the option to invest up to $5.0 billion in future equity financings.

(6) Amazon 10-Q: “In Q1 2026, we invested $15.0 billion in Series C Preferred Stock of OpenAI, and we also entered into an equity commitment letter agreement... to purchase additional shares of Series C Preferred Stock... with an aggregate purchase price of $35.0 billion... we are obligated to purchase all remaining Commitment Shares upon the earlier to occur of (i) OpenAI meeting specified milestones, and (ii) OpenAI directly or indirectly consummating an initial public offering or direct listing... The parties’ obligations under the Letter Agreement will terminate if we have not invested the Commitment Amount by December 31, 2028.” Private-company equity carrying values of $16.2 billion and $48.1 billion at December 31, 2025 and March 31, 2026, per the same note. The letter agreement is filed as Exhibit 10.1.

(7) Alphabet 10-Q, Note 3, gross notional amounts of outstanding derivative instruments: equity derivatives $0 at December 31, 2025 and $30,000 million at March 31, 2026; the quoted description and fair-value statement follow the table.

(8) Alphabet 10-Q, MD&A contractual obligations and Note 5: the $40.0 billion structure and second-quarter funding expectation as quoted; variable interest entity future funding commitments of $1.1 billion and $40.7 billion at December 31, 2025 and March 31, 2026. One layout observation: the MD&A liquidity bullet describing the $40.0 billion commitment omits the “through 2030” milestone window that appears in Notes 3 and 5 and the contractual-obligations paragraph.

(9) REPORTED AND INFERRED, not filed: the identification of the counterparty as Anthropic. The Information reported Alphabet planned an investment of up to $40 billion; court documents reported via Data Center Dynamics place Alphabet’s existing Anthropic ownership near 14 percent; Bloomberg reported Google backstopping lease payments at five Anthropic-linked data-center sites. Alphabet’s existing stake predates the quarter and appears in its filings only within unnamed non-marketable equity securities.

(10) Alphabet 10-Q, Notes 3, 5, and 10: data-center lease backstops “accounted for as credit derivatives,” gross notional $16,940 million at December 31, 2025 and $28,436 million at March 31, 2026, fair value a $339 million liability; approximately $15.3 billion of additional credit-derivative notional entered into in April 2026; energy-infrastructure backstops under the financial-guarantees caption, “maximum potential amount of future payments under these guarantees was $9.0 billion.” The characterization contrast on the liability side is set out in “Apple just said a lot about Google” (July 8, 2026) in this series.

(11) Broadcom Inc. Form 10-Q for the fiscal quarter ended May 3, 2026, accession 0001730168-26-000054, Note 11, Subsequent Events, verbatim: “On June 8, 2026, we arranged for an investor partner to take on certain agreements to purchase AI racks based on custom AI accelerators designed by us and the related lease agreements with a customer that enable access to compute capacity. In connection with the arrangement, we entered into a backstop agreement with the investor partner for the customer’s lease obligations over 5-year terms. The backstop will increase over time as the AI racks are deployed and decrease as the customer makes payments on its lease obligations, with a maximum exposure of $29 billion. In the event of default by the customer, we have various remedies, including the assumption of the lease or effecting a sale of the AI racks, which would reduce our maximum exposure.” Neither the customer nor the investor partner is named in the note.

On March 31 CoreWeave filed a current report announcing an $8.5 billion loan facility. Moody’s rated it A3 and DBRS rated it A (low), and the company described the transaction, correctly, as the first investment-grade rated financing secured by high-performance computing infrastructure and an associated customer contract. The floating portion priced at 2.25 percentage points over the benchmark short-term lending rate. The facility matures in March 2032. It is non-recourse, meaning the lenders may look to a single subsidiary, CoreWeave Compute Acquisition Co. VIII, LLC, and not to CoreWeave itself. The customer is not named.(1)

Seven weeks later, on May 18, the same company filed another current report announcing another facility, this one $3.1 billion. Same collateral, graphics processors installed in data centers. Same two banks leading the syndicate. Moody’s rated this one Ba2 and Fitch rated it BB+. Both are speculative grade. It priced at 4.50 percentage points over that same benchmark, after tightening by half a point on demand the company called exceptional. The filing states what changed, in its own words. The proceeds support “customer contracts with two large, non-investment grade customers.”(2)

Five notches, seven weeks, and two and a quarter percentage points a year. The chips did not change. The customer did.

A3 sits one notch above the rating Moody’s assigned SpaceX in June and two above the rating it carries on Oracle. A3 is a rating on the facility. Baa1 and Baa2 are ratings on the companies. Set the borrower aside and the arithmetic is plain. Silicon in a warehouse, pledged beside a tenant’s contract that the rating implies is investment grade, borrows better than Oracle borrows. The same silicon, pledged beside two tenants who lack that rating, borrows below investment grade.

The A3 belongs to a ring-fenced entity whose paper carries somebody else’s promise to pay.

That is a seat. The seat is whose promise stands behind the position, and the market prices it in an indenture, every time, without ever saying so out loud.

Both facilities are secured by the same class of equipment. The structures differ. Source: CoreWeave, Inc., Forms 8-K filed March 31 and May 18, 2026, accessions 0001769628-26-000129 and 0001769628-26-000236, Exhibit 99.1 each.

Three times this year a borrower has asked the American bond market for exactly twenty-five billion dollars. Oracle in February, SpaceX in June, Amazon this week. Alphabet asked for fifteen billion in February and took twenty, on more than a hundred billion of orders.(16) CoreWeave’s speculative-grade facility tightened during syndication. Every bond sold. The market wanted to lend at every seat at the table.

What varied was the promise it extracted.

Alphabet pledged nothing to its lenders and borrowed for a century. Its 6.125% notes due 2126, one billion pounds, dated February 13 of this year, come due on February 13 of the year twenty-one twenty-six. No covenant, no cash floor, no stated project. The century tranche drew close to ten times its size in orders.(16) Alphabet may buy the bond back at any time before August 2125, by compensating the holders.(4)

Amazon pledged nothing to its lenders and borrowed to 2066. Eight tranches, twenty-five billion dollars exactly, coupons from 4.600 percent to 6.250 percent. The prospectus says the money is for “general corporate purposes, which may include, but are not limited to, repayment of debt, acquisitions, investments, working capital, investments in our subsidiaries, capital expenditures, and repurchases of outstanding shares of our common stock.”(5) Forty-year money whose stated purpose includes buying back the borrower’s own stock. Amazon was not obliged to justify the loan, and did not.

Oracle pledged its conduct. Moody’s rates it Baa2 with a negative outlook and S&P rates it BBB with a negative outlook, both citing the capital program and the concentration of its cloud revenue in a small number of counterparties. In a free writing prospectus filed February 1, Oracle told the market it expected to raise forty-five to fifty billion dollars during the calendar year. Roughly half would come from selling stock, or instruments that convert into stock. The other half would come from one bond deal, and only one. “Oracle does not expect to issue additional bonds during calendar year 2026 beyond this transaction,” the document says, and “this funding plan reflects Oracle’s commitment to maintaining an investment-grade rating.”(6) A promise about future conduct, filed with the Commission, to protect a letter. The market then lent it twenty-five billion dollars across eight tranches reaching 2066, the same year Amazon reached this week.

SpaceX pledged cash. Moody’s assigned Baa1 with a stable outlook on June 18, and its rating action states that “a key credit consideration is SpaceX’s intention to maintain at least $25 billion of cash and marketable securities, supplemented by full revolver availability,” alongside a consolidated leverage covenant tested quarterly with no step-downs. Days later the company priced twenty-five billion dollars across five tranches, the longest running to 2056, with proceeds named to the repayment of its bridge loan in full.(7) Twenty-five billion dollars held motionless in order to borrow twenty-five billion dollars, and the borrowed money was told where to go.

CoreWeave pledged the hardware. Its debt stack carries a weighted average effective rate of 9.1 percent, and 9.9 percent excluding the convertibles. Fifty-nine percent of principal, nearly fifteen billion dollars, carries an effective rate of ten percent or higher. The most expensive facility runs at fifteen percent. The company discloses that its debt “bears interest at variable rates, the majority of which is unhedged.”(8) Nothing in its capital structure matures after 2032.

CoreWeave borrowed three times this spring. In March it pledged the machines and an investment-grade customer’s contract, and the fixed portion of that money cost about 5.9 percent. In May it pledged the machines and two customers who are not investment grade, and paid 4.50 percentage points over the benchmark. In June it pledged nothing at all, borrowing on its own name, and paid 9.625 percent. Its other unsecured notes pay 9.000, 9.250, and 9.750 percent.(8)

What stands behind those chips is a small number of promises. Committed contracts, which are take-or-pay and typically carry a prepayment, produced 98 percent of CoreWeave’s revenue in the first quarter. Roughly 65 percent of that revenue came from two customers, and no other customer reached ten percent. A year earlier a single customer produced 72 percent. One of the two is OpenAI, under a master services agreement signed in May 2025. Whether either of those customers is a party to either facility is not disclosed.(9)

Banks lent SoftBank forty billion dollars unsecured, against no collateral at all, and that bridge comes due in March 2027. S&P rates the company BB+, below investment grade, and cut the outlook to negative in March. When SoftBank sought ten billion dollars more against its OpenAI shares, the market declined. Reporting has it cutting the target by forty percent when lenders hesitated at collateral that no public market prices daily, watching the talks stall in June, and reviving them only by offering a corporate guarantee, so that lenders could look past the pledged stake to SoftBank itself.(14)

A century, forty years, thirty years, six years, and a margin call. The order books were full at every one of them.

Every one of these borrowings found lenders. The difference is what each was made to promise. Sources in the notes below.

The inversion sits on the page. The buildout’s shortest paper funds the equipment with the shortest life. Alphabet borrows to 2126. CoreWeave borrows to 2032 against equipment its own annual report depreciates over six years, an estimate the company makes and reviews, and every dollar of that stack must be refinanced, on terms nobody can know, before the machines it bought have finished paying for themselves. The company that sells CoreWeave those machines states in its own annual report that it brings new advanced architectures “on a one-year product cadence.”(13) The refinancing clock spins fastest exactly where the assets melt.

A house is legible from the other side of the table.

Microsoft approached the table and walked away without money. Its most recent notes carry Moody’s highest rating, and the agency’s own action explains that the company “will not receive any cash proceeds from the issuance of the new notes in the exchange offers,” because the old notes were surrendered, retired, and cancelled.(3) The paperwork moved. The money did not.

It watches from a seat it does not have to fund. It holds approximately 27 percent of OpenAI on an as-converted basis, accounted for under the equity method. Its own filing describes Microsoft as “a major investor in OpenAI” which “will continue to receive revenue-sharing payments,” and which holds “rights to OpenAI’s intellectual property, including models and infrastructure.” Under the restructuring announced in April, Microsoft’s revenue-share payments to OpenAI end, while OpenAI’s payments to Microsoft continue through 2030 against a cap. And when Microsoft reports adjusted net income, its own preferred measure, that measure excludes net gains and losses from its investments in OpenAI. Across the nine months ended March 31, the difference between the two figures is roughly four and a half billion dollars, and the difference is the position in the buyer.(10)

Amazon is the same shape, doubled.

It borrowed to 2066 for purposes that include, in the prospectus language, investments. Between 2023 and 2025 it invested eight billion dollars in Anthropic convertible notes. In the first quarter Amazon put fifteen billion dollars into OpenAI Series C preferred stock and signed a letter agreement committing thirty-five billion dollars beyond that, which it may exercise at its sole discretion and must exercise upon the earlier of specified milestones or an OpenAI public listing. In that same quarter AWS and OpenAI expanded an existing thirty-eight billion dollar commercial arrangement by one hundred billion dollars over eight years, an arrangement that “includes contractual obligations related to the performance of AWS chips.” The Anthropic arrangement is also primarily for the provision of AWS cloud services, and it also includes the use of AWS chips.(11)

Equity in both laboratories. Compute sold to both. The seller’s own chips written into both contracts. Forty-year money available for all of it.

NVIDIA sells CoreWeave the graphics processors that secure CoreWeave’s loans. CoreWeave spent $7.7 billion on property and equipment in the first quarter. In January, NVIDIA bought approximately 23 million shares of CoreWeave’s Class A common stock at $87.20 per share, two billion dollars, in a private placement.(12) The seller of the collateral holds equity in the borrower. NVIDIA’s own annual report observes that “access to capital can be particularly constrained for less-capitalized companies, which may face difficulties securing financing for large-scale infrastructure projects.”(13)

The commitments run among the same names. Microsoft funds OpenAI and sells OpenAI compute. Amazon funds both laboratories and sells both laboratories compute, with its own chips written into the contracts. NVIDIA sells the machines and holds equity in the company that buys the machines. Google sells the chips, and states that in connection with certain of those agreements it provides credit backstops to support third-party data centers. It carries an equity derivative in a private company it does not name.(15) What the houses commit to one another, they commit against balance sheets the market lends to for a century.

None of that is irregular. Vendors have financed their customers since the first equipment lease. Special purpose vehicles, take-or-pay contracts, credit backstops, prepayments: this is the ordinary furniture of capital-intensive industry, and in these filings each one appears correctly characterized and correctly disclosed.

Several of these filings also rest on judgment. A useful life is an estimate. A materiality threshold is a determination. The characterization of an instrument is a choice among permitted treatments, and so is the measure a company asks the market to judge it by. In every case the judgment belongs to the filer, and in every case the filer made it and disclosed it. The circle is not the finding, and neither is the judgment.

What is not evenly distributed is the rake. Taking a margin on the flow, holding a position in whoever pays the margin, and pricing that position against a balance sheet the position cannot move. The third condition is what permits the first two. A player can borrow money. It owns a balance sheet, and its own bets move it. What it can do is rent a balance sheet that its bets cannot move, for one facility, at a stated price. That price is on the page. About five point nine percent when CoreWeave borrowed behind an investment-grade customer’s promise, and 9.625 percent when it borrowed behind its own. Three and three-quarters percentage points separate the two. Some of that gap is the collateral, some is the structure, and some is whose promise stood behind it, and the filings do not separate them. Whatever share of it is rent, the rent is payable one facility at a time. What a player cannot do is own one.

The March facility is a secured term loan. The June notes are unsecured corporate bonds. The May facility priced at a floating rate; March and June are fixed coupons. Source: CoreWeave, Inc., Forms 8-K of March 31, May 18, and June 18, 2026.

The bond market does not argue with any of it. It prices it. A century of duration for one credit and nothing past 2032 for another is a view, stated in the only language an indenture has. So is twenty-five billion dollars of cash held still. So is two hundred twenty-five basis points for the difference between one customer’s promise and another’s. The market does not raise its voice. It writes the number into the term, and the term into the paper, and then the paper trades.

Which returns the question to the seat, and to what the seat is made of.

Alphabet accounts for its data-center backstops as credit derivatives, and its most recent quarterly states both numbers: the most it could ever be called on to pay, $28.4 billion, and what it carries the promise as worth, a liability of $339 million, which the company describes as not material.(15) That is what a ledger of lent credit looks like on a house’s balance sheet. An investment-grade rating on a warehouse of graphics processors is what the invoice looks like when an agency prints one.

Somewhere behind that facility is a customer whose promise turned pledged silicon into investment-grade paper. CoreWeave did not name it. TeraWulf, disclosing a twenty-year Anthropic lease at Hawesville, Kentucky, worth roughly nineteen billion dollars of contracted revenue, said only that the payments are “expected to be supported by an investment-grade credit,” and did not name that either.(15) No rule required either of them to name a counterparty, and nothing here identifies either one. The tenants are named. The credits go unnamed.

Every bond at this table sold. The one borrowing that did not clear was SoftBank’s, against a stake in a private company, and it moved only when SoftBank offered its own guarantee on top of the collateral. What separated these borrowers was not whether the money was there, and not the size of the ask, and not the quality of the collateral, which in CoreWeave’s case was identical across five notches.

What separated them was the cash standing behind the paper. Whose it was. How it was earned. Whether the promise to pay it could survive the position going wrong. A promise from a balance sheet the bet cannot move is worth a century. A promise from two customers who are not investment grade is worth five and a half years. The price of the seat is the quality of the cash behind it.

Between March and May the promise behind CoreWeave’s paper changed. The filings name neither customer. They state the credit quality of one pair, and the rating of the other, and they leave everything else alone. What they do state is the price of the difference: five notches, and two hundred twenty-five basis points.

Who owns the house, which tables are open, who is sitting at which tables, and who wants a seat?

The cost is in the tenor.

NOTES

(1) CoreWeave, Inc., Form 8-K filed March 31, 2026, accession 0001769628-26-000129, Items 1.01 and 2.03, Exhibit 99.1. States the A3 and A (low) ratings, the non-recourse structure, the floating tranche at SOFR plus 2.25% and a fixed tranche at approximately 5.9%, the March 2032 maturity, and the securing entity. Item 1.01 filings are required, not elective. FILED.

(2) CoreWeave, Inc., Form 8-K filed May 18, 2026, accession 0001769628-26-000236, Items 1.01 and 2.03, Exhibit 99.1. States the Ba2 and BB+ ratings, final pricing at SOFR plus 4.50% after a fifty basis point tightening, an approximately 5.5 year maturity, issuance through CoreWeave Financing DDTL V, LLC, and the credit quality of the two customers. That the March facility’s unnamed customer is investment grade is an INFERENCE drawn from the rating and from the contrast with this filing. It is not stated. Neither customer is named in either filing. The two facilities also differ in structure: the March facility is non-recourse through a separate entity and was privately arranged; the May facility was publicly syndicated through a different entity. Both are secured by high-performance computing infrastructure.

(3) Moody’s Ratings, rating action assigning Aaa to Microsoft’s new notes, June 2026; quoted language verbatim. Microsoft and Johnson & Johnson are the only two U.S. public companies rated AAA by S&P and Aaa by Moody’s. REPORTED.

(4) Alphabet Inc., Form 8-K filed February 13, 2026, accession 0001193125-26-051423, Exhibit 4.6, the global note for £1,000,000,000 of 6.125% Notes due 2126, principal due February 13, 2126, redeemable at the company’s option on a make-whole basis prior to August 13, 2125. Alphabet is rated AA+, not AAA. FILED.

(5) Amazon.com, Inc., Form 424B5 filed July 8, 2026, accession 0001104659-26-081786, Registration No. 333-293246. This is the priced prospectus supplement; the preliminary supplement of July 7, accession 0001104659-26-080950, states the tranches without amounts or coupons. Eight tranches totaling $25.0 billion, maturities 2029 through 2066, coupons 4.600% to 6.250%. The use of proceeds language is verbatim. FILED.

(6) Oracle Corporation, free writing prospectus filed February 1, 2026, pursuant to Rule 163, accession 0001193125-26-032650, Registration No. 333-277990. All quoted language is verbatim. Tranche detail from Oracle’s Form 424B2, accession 0001193125-26-035603, Registration No. 333-277990: $25,000,000,000 across eight tranches, the longest maturing in 2066. The Moody’s and S&P ratings and outlooks are REPORTED. FILED.

(7) Moody’s Ratings, rating action assigning Baa1 to Space Exploration Technologies Corp., June 18, 2026; quoted language verbatim. Fitch assigned BBB+ and S&P assigned BBB, both stable. Tranche detail from the company’s pricing announcement of its inaugural $25 billion issuance, settled June 26, 2026. Baa1 is three notches above Baa3, the lowest investment grade rating at Moody’s. REPORTED and FURNISHED.

(8) CoreWeave, Inc., Form 10-Q for the quarter ended March 31, 2026, accession 0001769628-26-000222, Note 10, Debt, for the weighted average rate, the distribution of rates, and the maturities. Rates in that note are effective rates, which include issuance costs and discount, and are not coupons.

The three spring borrowings, each from a filed current report: Form 8-K of March 31, 2026, accession 0001769628-26-000129, Exhibit 99.1, a facility secured by high-performance computing infrastructure and an associated customer contract, non-recourse to CoreWeave, floating tranche at 2.25 points over the benchmark short-term rate and a fixed tranche at approximately 5.9%, maturing March 2032; Form 8-K of May 18, 2026, accession 0001769628-26-000236, Exhibit 99.1, secured, priced at 4.50 points over that benchmark, supporting contracts with “two large, non-investment grade customers”; Form 8-K of June 18, 2026, accession 0001769628-26-000291, Item 1.01, $1,250 million of 9.625% dollar-denominated senior notes and EUR 2,000 million of 8.500% euro-denominated senior notes, senior unsecured, guaranteed by subsidiaries, maturing July 15, 2032. Existing unsecured coupons of 9.000%, 9.250% and 9.750% appear in CoreWeave’s April 2026 current reports.

The March money was raised through a separate, non-recourse subsidiary and the June money on the parent’s own credit. The collateral, the customer’s contract, and the structure all sit inside the difference between the two rates. The May facility priced at a floating rate; the March and June figures are fixed coupons. FILED.

(9) The same CoreWeave 10-Q. Committed contracts are described as take-or-pay and as typically including a customer prepayment; they produced 98% of revenue in each of the three months ended March 31, 2026 and 2025. The top two customers produced approximately 65% of revenue in the quarter, with no other customer at 10% or more; the top customer alone produced approximately 72% a year earlier. The master services agreement with OpenAI OpCo, LLC dates to May 2025, with an order form in September 2025. CoreWeave’s revenue concentration and the customers of its financing facilities are separate disclosures. No filing states whether the two sets overlap, and nothing here suggests that they do. FILED.

(10) Microsoft Corporation, Form 10-Q for the quarter ended March 31, 2026, accession 0001193125-26-191507. The 27 percent as-converted equity method investment, the hypothetical liquidation at book value method, and the $13 billion of total funding commitments of which $11.8 billion was funded, are stated in the filing, as is the quoted partnership language. The filing states that the partnership was extended in October 2025 and April 2026. Adjusted net income is a non-GAAP measure that, in the company’s words, excludes net gains and losses from investments in OpenAI; nine month reported net income was $97,983 million against $93,500 million adjusted. The difference reflects net gains, not losses. FILED. The direction of the revenue share after the April restructuring, and the 2030 term and cap, are REPORTED, and were read at length in “Microsoft just said a lot about SpaceX” (Cape Fear Advisors, April 27, 2026).

(11) Amazon.com, Inc., Form 10-Q for the quarter ended March 31, 2026, accession 0001018724-26-000014. States $8.0 billion of Anthropic convertible notes invested from Q3 2023 to Q4 2025, with conversion of portions to nonvoting preferred stock in Q1 2025 and Q1 2026. The filing also states a $15.0 billion Q1 2026 investment in OpenAI Series C Preferred Stock and a letter agreement for $35.0 billion of additional Commitment Shares, purchasable at Amazon’s sole discretion and mandatory upon the earlier of specified milestones or a public listing; the AWS expansion of an existing $38.0 billion commitment by $100.0 billion over eight years, with the quoted chip language; and, as a subsequent event, a further $5.0 billion invested in Anthropic nonvoting preferred stock. FILED. Anthropic is the developer of Claude, the model used in preparing this analysis. The figures above are quoted from Amazon’s filing without characterization.

(12) CoreWeave, Inc., Form 10-Q for the quarter ended March 31, 2026, accession 0001769628-26-000222: “In January 2026, we entered into a securities purchase agreement with NVIDIA Corporation for a private placement of approximately 23 million shares of our Class A common stock at a purchase price of $87.20 per share, for aggregate gross proceeds of $2.0 billion.” Purchases of property and equipment, including capitalized internal-use software, were $7,695 million in the quarter. FILED. What portion of CoreWeave’s equipment purchases is paid to NVIDIA is not disclosed.

(13) NVIDIA Corporation, Form 10-K for the fiscal year ended January 25, 2026, accession 0001045810-26-000021. Verbatim: “We continue to execute Data Center compute product introductions, bringing new advanced architectures on a one-year product cadence, including our Rubin platform.” The observation regarding access to capital appears in the same passage, verbatim. FILED.

(14) S&P Global Ratings outlook revision, March 2026. The margin loan sequence, the forty percent reduction in the target, the stall, and the corporate guarantee are REPORTED, sourced to Bloomberg and others. SoftBank has publicly stated that the $40 billion unsecured bridge, maturing March 2027, will be repaid through existing assets and other financing measures. SoftBank is not an SEC registrant for these obligations; nothing in this note is drawn from a filing with the Commission. REPORTED.

(15) Alphabet Inc., Form 10-Q for the quarter ended March 31, 2026, accession 0001652044-26-000048. The $28.4 billion gross notional and the $339 million fair value liability, described as not material, are read at length in “Apple just said a lot about Google” (Cape Fear Advisors, July 8, 2026). Alphabet is not named in, and is not stated to be a party to, the CoreWeave facility discussed above. FILED.

Separately, and cited for a different proposition: TeraWulf, Inc., Form 8-K filed July 6, 2026, accession 0001104659-26-080583, Exhibit 99.1, for the Hawesville lease and the quoted phrase. That filing does not name the investment-grade credit it refers to. Neither does this piece. No connection between that credit and any filer named above appears in the record. FILED.

(16) The order books for Alphabet’s February offerings, the upsizing of the dollar tranche from fifteen to twenty billion dollars, and the roughly tenfold oversubscription of the sterling century tranche, are REPORTED, sourced to Bloomberg and others.

This morning Apple published a press release. The headline is American manufacturing: a new multiyear commitment with Broadcom, expected to exceed $30 billion, more than 15 billion chips made in the United States, a $1.5 billion expansion of Broadcom’s plant in Fort Collins, Colorado.(1) Tim Cook is quoted. The administration is thanked. The number is large, it is stated plainly, and it appears in a press release, which no rule compelled. Apple chose to say it. That is the first card of the week to turn face up, and it confirms a card Broadcom turned two days earlier.

On July 6 Broadcom filed a current report with the SEC. It says Broadcom and Apple agreed to extend their collaboration through 2031, with Broadcom developing and supplying custom silicon for multiple generations of Apple products. It names Apple. It states no figure. It attaches no agreement. And it is filed under Item 8.01, Other Events, the box a company uses for something the rules did not require it to say. Broadcom was not compelled to file this. A supply relationship in the ordinary course of business carries no such duty, and the generous reading, likely the correct one, is that Broadcom chose to announce ordinary-course business. The choice stands either way, and the box it chose carries a name and nothing more.(2)

One agreement, two elective doors. Broadcom volunteered the name and withheld the figure. Apple volunteered the figure and filed nothing. A $30 billion, five-year commitment, and not a line of it was forced onto a page. Each party turned one half face up and kept the other down.

A number a company announces is a chip. It is the size of a position laid on the table, agreed to be worth its face for the length of the hand, and whether it redeems for cash is a separate question the announcement leaves alone. The $30 billion measures a vault, the composite of owned chips, some disclosed and some not. By naming Broadcom and putting a figure to the relationship, Apple vouched for a durable book of business, silicon it will buy through the end of the decade. A vault the counterparty confirms stops being a claim. It sits on the table as a fact.

The same record carries a second Broadcom chip, larger, at another table. Its June quarterly report, in the subsequent-events note, states a guarantee: Broadcom backstops a customer’s roughly $35 billion chip-financing vehicle, and states the maximum it could be called to pay, $29 billion.(3) So on one side of the week’s ledger sits a $29 billion backstop, and on the other, confirmed this morning by Apple, $30 billion of committed Apple silicon through 2031.

The note calls the beneficiary “a customer” and the party beside it “an investor partner.” Broadcom’s April announcement had already named the three principals: Broadcom, Google, and Anthropic.(4) The June note withholds the names; the April filing supplies them. Stacked, the two Broadcom disclosures name Google as the investor partner behind the $29 billion backstop.

The guarantee standard governs that note, and it leads with the ceiling: the maximum Broadcom could be called to pay, $29 billion, is the headline obligation. The name it abstracts to “a customer.” So the figure is on the page and the party is stacked in from April.

Google leads with a different number. It backstops the same kind of buildout, standing behind the data centers that house it, and accounts for its data-center backstops as credit derivatives, carried not at what could be lost but at what the position is worth to hold. Its filing states both figures. The notional, the amount the backstops could be called to cover, reached $28.4 billion at the end of March, up from $16.9 billion three months earlier, with roughly $15.3 billion of further agreements added in April. The fair value of those same derivatives is a liability of $339 million, which Google calls not material.(5)

The distance between those two numbers is the finding. Twenty-eight billion of exposure is not a small sum by any measure a person carries in their head. It is carried at three hundred thirty-nine million and labeled immaterial. The label is earned, and only scale earns it: a balance sheet large enough to cover a bet this size without feeling it is permitted to price the bet near zero and be stating it correctly. The number is enormous. The word Google is allowed to use for it is immaterial. Both hold at once, and only the size of the safe reconciles them.

Google’s own filing sets the two treatments side by side. On one page, its backstops of energy infrastructure are financial guarantees, stated the way Broadcom stated its, a maximum potential amount of future payments, $9.0 billion.(6) On the next, its backstops of data centers are credit derivatives, stated at fair value, immaterial. Same filer, same quarter, two backstops, one word apart, and the word decides which figure leads. The guarantee box leads with what could be lost. The derivative box leads with what the position is worth, and files what could be lost as one line among the derivatives. Both boxes abstract the name, Broadcom’s to “a customer,” Google’s to “certain third parties.”

So Anthropic’s financing is legible only stacked. Broadcom leads with the ceiling and calls the customer “a customer.” Google leads with the value, files the ceiling as a detail, and calls the counterparties “certain third parties.” Each box leads with the number that serves it and abstracts the name, and no single document holds the whole: the size from Broadcom, the characterization from Google, the name from an April announcement and the counterparty filings.

At one table, that is a large position, correctly priced low against a strong counterparty and an enormous balance sheet.

The house does not appear at one table. Across more than one, the same box repeats. The roughly $40 billion Google committed to a private company, ten billion now and thirty billion contingent through 2030, is carried as an equity derivative; the private company is not named, and the commitment matches the one reported for Anthropic.(7) Google’s cloud arm, in its own words, supplies gigawatts of chips to customers and, in connection with those agreements, provides credit backstops to support third-party data centers.(8) One filing shows a hedge. The stack shows a house, taking the cut on the flow, holding the position, and pricing the bet against a balance sheet the bet cannot move. The rake and the bet are the same hand. Google sells the chips, backstops the buildings they run in, and holds equity in the buyer.

The treatment is honest only for that hand. A player backstopping a weak counterparty could not carry the chip at an immaterial fair value, because the value would have to reflect a real chance of being called and no reserve behind it, and the chip would land material, because the default would be plausible. Only a balance sheet that covers the systemic bet without noticing gets to price it at almost nothing and state it correctly. The box is not a disguise. It is a declaration of role. A guarantee at face value is a player’s box. A credit derivative at immaterial fair value is a house’s box, honest precisely because the house is a house.

The paradox stays on the page, unresolved: the position that most holds up the independent buildout is the one that reads, correctly, as the smallest.

The cards keep turning, and the pace is the point. Apple named a number this morning and stood behind a player; Broadcom named Google on Monday; the counterparties file through the summer, and each report adds a line. Hands are being dealt across more tables than one. Standing behind a player is a confirmation; it grows more interesting still when the backers take seats of their own, as Google already has. The thing to watch is which sits down next. The filings leave one question plainly open. Who is playing at a single table, and who are the houses?

Apple did just say a lot about Google.

NOTES

(1) Apple Inc., “Apple to increase spend with Broadcom to produce billions more U.S. chips,” newsroom press release, July 8, 2026. The figure is stated as “expected to exceed $30 billion,” a multiyear total rather than an annual one; the release also states more than 15 billion U.S.-made chips, hundreds of jobs, and a $1.5 billion Broadcom investment to expand its plant in Fort Collins, Colorado. A press release is disclosure that no rule compels, and Apple filed no current report of its own on the agreement.

(2) Broadcom Inc., Form 8-K filed July 6, 2026, accession 0001193125-26-295589, Item 8.01, Other Events, signed by Chief Financial Officer Amie Thuener. It states the parties, the 2031 term, and the custom ASIC scope, states no figure, and attaches no agreement. Item 8.01 is elective; Item 1.01, Entry into a Material Definitive Agreement, would have required the agreement as an exhibit, and agreements made in the ordinary course of business are exempt from Item 1.01.

(3) Broadcom Inc., Form 10-Q for the quarter ended May 3, 2026, accession 0001730168-26-000054, Note 11, Subsequent Events, states the guarantee at a maximum exposure of $29 billion, describing the beneficiary as “a customer” and the counterparty as “an investor partner.” The roughly $35 billion vehicle it backstops, and the full reading of that note, are set out in “Anthropic, the financing before the filing” (Cape Fear Advisors, July 6, 2026).

(4) Broadcom Inc., Form 8-K, April 2026, accession 0001193125-26-144028, names Broadcom, Google, and Anthropic PBC in the collaboration the guarantee finances. The June note names neither the customer nor the investor partner; the identification runs through the April filing.

(5) Alphabet Inc., Form 10-Q for the quarter ended March 31, 2026, accession 0001652044-26-000048, accounts for the company’s data-center lease backstops as credit derivatives, governed by ASC 815 and carried at fair value. The gross notional, the maximum potential amount of future payments, is stated at $28.4 billion at March 31, up from $16.9 billion at December 31, 2025, and $6.5 billion at September 30, 2025; roughly $15.3 billion of further such agreements were entered in April 2026. The fair value is stated as a liability of $339 million and described as not material.

(6) The same Alphabet 10-Q accounts for the company’s energy-infrastructure backstops as financial guarantees under ASC 460 and states a maximum potential amount of future payments of $9.0 billion. A guarantee under ASC 460 is disclosed at that maximum; a credit derivative under ASC 815 is carried at fair value. Both treatments appear in the same filing, applied to backstops of the same broad kind.

(7) The same Alphabet 10-Q describes an approximately $40 billion investment in a private company, a $10 billion commitment and $30 billion of future funding contingent on milestones through 2030, accounted for as an equity derivative. The private company is not named in the filing. Reporting has identified the recipient as Anthropic and the figures match; the identification is reported, not filed.

(8) In the same filing’s discussion of results, Google states that its cloud business supplies gigawatts of chip capacity to customers and, in connection with certain of those agreements, provides credit backstops to support third-party data centers and power infrastructure. The counterparties are described as third parties and are not named.

The deal is public and was carefully reported. Analysts and reporters had the vehicle, the roughly $35 billion, Google’s chips, Anthropic as lessee, and Broadcom’s backstop, which they sized at $30 to $31 billion from people familiar with the terms while noting that the exact cap stayed undisclosed. (1) That was a careful approximation, and it was close. We went to the filings to confirm it and came back with something firmer: not an estimate but the figure itself, filed. Broadcom’s June quarterly report states a maximum exposure of $29 billion, in the subsequent-events note above. The confirmation was in the document, for anyone who read it there.

The same holds for the customer’s name, which the coverage gave as Anthropic, and reasonably. The June note itself says only “a customer.” Broadcom’s April 8-K names Anthropic in the collaboration the note finances, the platform’s announcement names Anthropic as its lessee, and the full identification runs through those filings and the public reporting, laid out below, under “The parties who file.”

The vehicle owns the hardware and carries the borrowing, so the debt sits with the vehicle and the $29 billion sits on Broadcom’s guarantee, neither on Anthropic’s books. That placement follows from the structure the parties chose to fund the build-out, not from a decision to keep anything off a page. The structure has a reason, and the reason is worth granting before the piece reads how the arrangement discloses itself.

The need is real, and its size explains the structure. Anthropic is meeting demand for its services that requires expensive hardware to deliver, and the capital that hardware takes runs past what an emerging company reaches through equity and ordinary credit. That is why a vehicle like this exists: the size of the need shapes the solution.

Anthropic has said as much in its own voice. Announcing its Series H round in late May, the company recorded that it had signed agreements with Google and Broadcom for five gigawatts of next-generation TPU capacity, and its chief financial officer said the funding would “help us serve the historic demand we are experiencing.” The compute the vehicle finances, Anthropic named; the demand behind it, its finance chief stated. What Anthropic has not done is file. The financing that carries the compute onto its books, or keeps it off, is what a filing would show.

The need also carries a condition, and Broadcom filed it. In its April agreement, Broadcom records that consumption of the capacity is “dependent on Anthropic’s continued commercial success.” Broadcom, with $29 billion at stake, named in its own words what the need rests on. The dependency belongs to the record, and to the hand most exposed to it.

This piece takes both as given: the need, and the size that shapes the structure. It reads only how the arrangement discloses itself across the filings: where the exposure lands, who must name it, and who may stay silent.

The parties who file

The vehicle is a special-purpose entity, the kind US GAAP treats as a variable interest entity: single-purpose, thinly capitalized, its control and its risk parceled among the parties by contract. Entities like it are among the most closely governed in the accounting standards. One set of rules decides whose balance sheet the vehicle sits on. (2) A separate set decides who discloses which interest, by the nature of the interest held. Both are detailed, both are settled, and here both are followed. Each party discloses what its rules require.

The rules produce disclosure spread across companies, because the arrangement is spread across companies. Broadcom guarantees the debt, Google supplies the chips and holds a stake in Anthropic, Apollo lends and holds, and Anthropic leases. Each files its own interest, in its own report, correctly. The vehicle becomes legible by assembling those filings, because the vehicle is assembled from those parties. Reading it takes several filings at several companies for the plain reason that the arrangement is several companies. The spread of the disclosure is the shape of the deal, faithfully reported.

The question this section asks is narrow, and it rests on top of a standard correctly applied: as the relationship crosses each company’s books, where does the name go, and where does the number?

The cast holds for any vehicle of this kind. An asset-financing vehicle like this has a fixed set of seats, because it solves a fixed problem: raise capital, buy an asset, lease it to the party that needs it, enhance the credit so the debt prices. Each seat carries its own standard, because the standard attaches to the nature of the interest, and the seat is the interest. The role names the standard.

Two of the exhibit’s columns hold for any vehicle of this type, and two vary: the role and its standard are fixed, while who sits in the seat and what they have filed change from deal to deal.

Several parties hold more than one seat. Google sells the chips and holds a stake in Anthropic. Broadcom designs the chips and guarantees the debt. Apollo sponsors the vehicle, holds its first loss, and buys its senior notes. The disclosure spreads across filings, and within a filing across notes, because a single company sits in several seats, each seat under its own standard.

The deal is public. Anthropic announced the five gigawatts of Google and Broadcom TPU capacity in its own words, and Apollo, Blackstone, and Broadcom announced the financing platform with Anthropic as its first lessee. The names are on the record.

The filings use other words for them. Broadcom’s June note calls Anthropic “a customer” and Apollo “an investor partner.” Item 5 of the same report names Apollo, the partner April’s 8-K had left “in discussions,” and still calls Anthropic “a customer.” Broadcom’s April 8-K names Anthropic plainly, “Broadcom, Google, and Anthropic PBC,” with 3.5 gigawatts of TPU compute through Broadcom, and turns to “certain operational and financial partners” at the financing. Broadcom names Anthropic at the compute and writes “a customer” at the financing, in both filings.

The identification is that simple: the customer Broadcom’s note describes, by its Broadcom-designed racks and five-year leases, is the Anthropic the announcements named. The finding is the gap. Where the announcement says Anthropic, the filing says “a customer,” and it says so at the financing, where the rules would let it write the name. Perhaps an agreement seals it; perhaps it is the ordinary practice of a filing that writes “a customer” even when the market knows who, the way CoreWeave’s filings write “Customer A” for its largest customer. The gap is on the record; the reason for it is not.

The other seats set what reaches the page in different ways. Some rules compel a number; others let a position fold into a total. Google folds: a rule lets it aggregate the sale and the stake, so the backlog carries Anthropic’s commitment under ASC 606 (3) and the mark carries the stake under ASC 321, (4) each inside a total, the positions small against Google’s results. Apollo’s insurer and asset manager report their holdings by class under ASC 320 (5) once the quarter that captures them is filed, Anthropic inside a security type. Anthropic holds the lease and, for now, files nothing public, its registration confidential and no trigger yet reached.

So Anthropic’s name reaches the page where a rule compels it, and folds into a total where a rule permits. At the financing it becomes “a customer,” where the rule would allow the name and something outside the rule holds it back.

One seat shows the materiality force by itself. The owner of the lessee holds non-marketable equity in Anthropic, marked under ASC 321. Google fills it at roughly fourteen percent and aggregates, the mark folded into other income, because at Google’s scale the position is immaterial and the rule lets the total stand unattributed. A registered fund, Destiny Tech100, fills the same seat at $100 million and names it in full, Anthropic PBC, in a prospectus supplement, because at the fund’s scale the same position is material and the rule requires the name. (6) Same seat, same standard, opposite ends of the scale, opposite disclosures. Here only materiality is at work, which makes it the clean case: the threshold moves with the filer, and the smaller holder states the name the larger folds away. What the record discloses tracks materiality to the filer, and runs opposite to size in the world.

The cash, in two parts

The money moves in two directions on two clocks.

One part is certain and moves now. The debt and the equity fund the vehicle, and the vehicle buys the chips. Reporting puts the debt near $35 billion across three tranches and the first-loss equity near $800 million, at coupons reporting also supplies. Those figures are reported rather than filed; they come from the parties describing the deal, not from a document any party has signed and submitted. What they describe is a single, upfront use of capital: the vehicle takes the borrowed money and the equity and pays Google for the hardware. That leg closes at the start and stands on its own.

The other part is contingent and arrives over time. Anthropic leases the chips and pays across five years, and those payments service the debt. The stream is future, and it carries the condition Broadcom filed: it depends on Anthropic’s continued commercial success. Behind the stream stands Broadcom’s backstop, which the same filing sizes at $29 billion and describes as growing while the racks deploy and shrinking as Anthropic pays. The contingent leg is where the structure lives, because it is the leg that can vary.

With the reported figures set aside, two numbers about this arrangement are filed rather than reported, and both describe the future. Broadcom’s $29 billion is the contingent ceiling on the backstop, gross, as its standard requires. Google’s $462.3 billion is the aggregate of Google Cloud revenue under contract and not yet earned, Anthropic’s share among it, as its standard requires. Google’s finance chief attributed the backlog’s near-doubling in part to TPU hardware sales, and told analysts that the TPU hardware agreements “are reflected in our cloud backlog,” the same $462.3 billion; the seller places this class of arrangement inside the number, an aggregate that names no one. The money that has already changed hands reaches the record only as reported. The two filed numbers both point forward: one a ceiling on a guarantee, one a backlog of revenue, each a measure of what has yet to happen.

The exposures

The $29 billion is a ceiling, and it is gross, because ASC 460 requires a guarantor to state the maximum it could pay before any recovery. (7) Broadcom filed the trigger and the remedies alongside it: the backstop is called if Anthropic defaults, and Broadcom’s remedies then, assuming the lease or selling the racks, would reduce the maximum exposure. Three parts of the exposure are filed, then, the ceiling, the trigger, and the remedies Broadcom says would reduce it, and one part is not: what those remedies would actually recover.

That recovery rests on a resale value no filing in the structure supplies, and it is the same value the CoreWeave readingturned on. What a specialized chip fetches after a lease has run and the market has moved is the quality of the asset behind the cash. Here the asset stands behind a guarantee rather than a balance sheet, and the question holds its shape: the exposure is only as covered as the collateral is worth, and the collateral is a used accelerator whose secondhand price stands outside every filing.

This is terrain Rod Dubitsky has been reading, (8) and further than we can reach from EDGAR. His forensic work on Athene’s statutory filings takes Athene’s book apart below the level the GAAP reports show: where we stop at ASC 320’s disclosure by security type, he reads the holdings themselves. On the residual-value question, and on how the rating agencies weigh a guarantee against the guarantor that gives it, his work goes where ours does not, and we point readers to it.

The net, the amount Broadcom would truly carry, is the filed ceiling less a recovery the record describes and does not size. And the condition beneath the whole arrangement, Broadcom filed as well: consumption of the compute, its April agreement records, is “dependent on Anthropic’s continued commercial success.”

The standard, and the page

Every party followed every rule. The vehicle stands disclosed, the guarantee filed at its gross maximum, the backlog aggregated as the standard directs, the stake marked where the standard places it. Each seat met its obligation, and the framework the FASB built for exactly these structures did what it was written to do. Their compliance is the finding, and it carries weight.

The sum of that compliance is a $35 billion financing that reaches the public record in pieces, held by the parties around the vehicle, on pages that belong to them. The number at the center, the $29 billion, reaches the record because Broadcom was required to state it, in Broadcom’s filing, under the label “a customer.” Reconstructing the vehicle took a reading across several filings at several companies, because the arrangement spreads across several companies and the disclosure spreads with it. That reading was the only route to the number, and it returned more than the number. It tightened a widely-held assumption into a filed fact, and it traced a pattern that outlasts this deal: a name placed where a rule compels it, folded into a total where a rule permits, and set aside for “a customer” at the financing, until the party every exposure in the structure runs back to is the one the filings name least.

The question the piece leaves begins where compliance ends. It asks whether a standard can be fully satisfied and still keep the financing behind a commitment of this size off Anthropic’s own page, legible only through the others.

The counterparties file again through the summer, and each report can add a line: Athene’s holdings by class, Apollo’s consolidation, Google’s next mark and its backlog. Anthropic’s own filing is expected around Labor Day, later than once thought and open to change, and it is the one that can answer the balance-sheet question. It can show whether Anthropic consolidates the vehicle or holds it apart, whether the lease appears as a right-of-use asset while the debt stays with the vehicle, (9) and whether the $29 billion Broadcom disclosed appears at last under Anthropic’s own name. It cannot fix what the hardware will fetch secondhand, or whether the demand the structure rests on holds, or where the net exposure behind the gross ceiling finally sits; those wait on the market, not the page. And it can decline the question altogether, aggregating the vehicle the way the counterparties aggregate the stake and the backlog, and leave the pieces where they lie.

The standard has been satisfied. The page is still to come.

The author holds no direct position in any security discussed. Whatever indirect exposure comes through index and managed funds is neither known to him nor, he suspects, escapable.

A note on tools and conflict. This analysis was drafted and pressure-tested with AI tools, including a model built by Anthropic, the company this piece examines. No tool removes that conflict; another model would only trade it for its own. The bias a tool cannot eliminate has three answers, and the piece uses all three. It avoids the judgment call where the record allows, resting its claims on filings and accounting standards a reader can check without trusting the tools or the author. It tests against the bias rather than around it. And in the end it rests on the author, who holds the judgment that remains and the responsibility for it.

NOTES

(1) The trade coverage that carried the deal sized Broadcom’s backstop from unnamed sources and flagged the cap itself as unknown. AI Weekly, “Apollo Backs $36B Record Debt Deal for Anthropic TPUs,” May 30, 2026, stated that “Broadcom’s residual-value guarantee cap is undisclosed”; OpenTools reported the same, that the guarantee cap “is also not public.”

(2) ASC 810-10 governs consolidation of a variable interest entity. The party that both directs the activities most significant to the entity’s economic performance and bears the obligation to absorb its losses or the right to its benefits (ASC 810-10-25-38A) is the primary beneficiary, and consolidates it. Every other holder of a variable interest discloses that interest under ASC 810-10-50 without consolidating. The test decides which balance sheet the vehicle sits on; when Anthropic files, its treatment of the vehicle under this test is the line to read.

(3) ASC 606-10-50-13 requires disclosure of the aggregate transaction price allocated to performance obligations not yet satisfied, with an explanation of when it will be recognized. The standard asks for the aggregate, not the customers. Anthropic’s commitment sits inside Google’s Google Cloud remaining performance obligations as one contract among many, disclosed in total and without names.

(4) For an equity holding without a readily determinable fair value, ASC 321-10-35-2 permits the measurement alternative: cost less impairment, remeasured to fair value when an observable price change occurs in an orderly transaction for the same issuer. A new Anthropic funding round is such a transaction. The holder marks its stake to the round and records the change in earnings, in other income, where the standard places it and where the issuer goes unnamed.

(5) A creditor holding the vehicle’s senior notes discloses them under ASC 320-10-50 by major security type, defined by the security’s nature and risk (ASC 320-10-50-1B), rather than by issuer. The holdings reach the record inside a class, not under the borrower’s name.

(6) Materiality is judged relative to the reporting entity and its investors, not in absolute dollars. The SEC staff’s SAB No. 99 (codified at ASC 250-10-S99) frames significance against a particular registrant’s statements and the total mix of information a reasonable investor in that entity would weigh, and the FASB locates the same judgment with those who understand the entity’s own circumstances. A $100 million position significant to a small fund, named in its schedule of investments, can sit inside an aggregate line for a company whose results run in the tens of billions.

(7) ASC 460-10-50-4 requires a guarantor to disclose the maximum potential amount of future payments it could be required to make, undiscounted, and before reduction for any recoveries under recourse or collateral provisions. Broadcom’s $29 billion is that figure, gross. The remedies listed in the same note, assuming the lease or selling the racks, are the recoveries the standard holds separate from the ceiling.

(8) Rod Dubitsky writes forensic analysis of Athene’s statutory filings and the structures around them at roddubitsky.substack.com, including “Inside Athene’s Nesting Dolls” and the “Dangers Lurking” series. The residual-value and rating-agency observations this section points to are his.

(9) For the lessee, ASC 842-20-25-1 places a right-of-use asset and a lease liability on the balance sheet at the present value of the lease payments. The vehicle’s own debt follows onto the lessee’s books only if the consolidation test in ASC 810 makes the lessee the primary beneficiary.

The phrase belongs to the auditor’s report: an opinion on the financial statements “taken as a whole.” It is a fine phrase, and it describes a reading that, as far as we can find, no one is actually required to perform. The opinion runs to conformity with the framework, and the framework governs one statement at a time. What follows is that reading, performed on CoreWeave’s Form 10-Q for the quarter ended March 31, 2026, filed May 8. (1) Our June piece ran the quality-of-cash test down that company’s filings vertically, line by line. This one runs it horizontally, statement against statement. Everything quoted below is from the one document.

The three portraits

The income statement presents a rental business. Revenue of $2,078 million for the quarter, up from $982 million a year earlier, recognized as service is delivered under the revenue standard, with cost of revenue of $716 million and technology and infrastructure expense of $1,273 million set against it. (2) Depreciation has no line of its own. The quarter’s charge, $1,147 million, appears in the cash flow statement’s reconciliation; on the income statement it is dissolved into the captions, the machines’ wear inside technology and infrastructure, the facilities’ inside cost of revenue. (3) There are no lessor captions: no lease income unwinding, no net investment in leases, no residual accounting. The statement’s theory of the company is a provider of capacity, earning as it serves, period by period. It is prepared exactly as the rules require for that kind of business. Which kind it is, is a presentation judgment the standards commit to management. Management has made it.

The balance sheet presents an equipment lessor. Property and equipment of $36.4 billion net, plus $10.2 billion of right-of-use assets, together 84 percent of total assets of $55.6 billion. (4) The face carries the machines at one number, net. The note beneath splits them: $40.9 billion gross, $4.5 billion already consumed, and $9.6 billion of construction in progress, which neither depreciates nor yet earns, nearly a quarter of the gross. (5) Against the machines: $24.9 billion of debt net of costs, $7.5 billion of deferred customer money, and a footnote tab whose payment expectations run to month eighty-four. (6) Assets that depreciate on schedules; obligations that mature on schedules; customer money that arrived years ahead of the service it purchases. That is the right-hand and left-hand side of a term-financing business, and every line of it is prepared exactly as the rules require.

The cash flow statement presents a financing operation. Operating activities provided $2,984 million, a figure that includes the customer prepayments, against $61 million a year earlier. Investing consumed $7,708 million, of which $7,695 million was purchases of property and equipment. Financing provided $3,914 million, including $3,290 million of new debt. And after all three engines, cash fell $810 million in the quarter. (7) Interest expense, net, ran $536 million, which is 26 percent of revenue, and the quarter’s debt service totaled $1.8 billion: $1.3 billion of principal and $459 million of interest. (8) The statement’s theory of the company is an organism that ingests capital, converts it to equipment, and services the conversion. It is also, of the three, the statement that would not change. Machines are investing, debt is financing, customer receipts are operating, under any name this business carries; the cash flow statement reads identically in every characterization of the company.

Three statements, three businesses, one company. Each portrait is correct. The rules that govern each are statement-local, and nothing anywhere requires the income statement’s theory of the business to agree with the balance sheet’s. The disagreement, if it is one, never appears inside any statement. It lives in the space between them.

The sentence in the footnote

One event this quarter crossed that space, and the filing discloses it in a single sentence:

“The change in deferred revenue in the three months ended March 31, 2026 was primarily driven by reclassification of $1.3 billion to customer liabilities, revenue recognized from deferred revenue at the beginning of the period of $304 million, partially offset by invoicing in advance of performance under contracts.” (9)

The two captions are not defined alike. Deferred revenue is money received for performance still owed; it is discharged by serving, and it becomes revenue. Customer liabilities, a line within other current liabilities that stood at $137 million at year-end and $1,469 million at quarter-end, the filing does not define. (10) So the sentence records a change of terminology as much as of amount: in one quarter, $1.3 billion left the caption that resolves into revenue for one the filing names but does not explain, set against $2,244 million of unrestricted cash and $11.1 billion of total liquidity. (11) The classified balance sheet adds tenor. The filing does not map the reclassification between maturity buckets; the columns record the endpoints: deferred revenue beyond twelve months smaller by $1.1 billion over the quarter, deferred revenue within twelve months larger, and the reclassified line current by placement. Wherever the promise sat, the obligation now stands inside the year. (10)

The filing states the mechanics and stops. No cause is given. Two sentences earlier, on the other side of the Contract Balances heading, the filing notes that one customer accounted for 68 percent of accounts receivable at year-end; it does not connect the two, and neither do we. (12) What can be said from the captions alone is this: the entry is correct, and it is correct in the way a term business’s books are correct. A short-cycle rental book contains no relationship that could generate an entry of this kind; a reclassification of this size arises from a long-dated committed arrangement whose accounting basis changed mid-term, and the balance sheet recorded it exactly as such books should. The nature of the change is not filed, and we do not supply one. What changed on the record is the word: money on track to become revenue now sits under a caption the last audited statements left unnamed. The terminology is the event, recorded correctly, in a voice the income statement does not use.

Where does the event register in the cash flow statement? It does not, and the statement is correct not to carry it. Cash flows are indifferent to recognition and presentation decisions; they track cash, and no cash moved. A liability-to-liability change has no line in the body, and the supplemental non-cash schedule records conversions of form, not changes of meaning within a form: its entries this quarter were equipment payables becoming debt, $1,469 million, and, in the prior year, a customer deposit becoming debt. (13) The prepayments themselves remain in operating cash under the character they had on arrival, the only character a cash statement records. But the reclassification happened, and it left its footprint where such footprints belong, one note over, in contract balances. The statement records motion; the notes record meaning; both did their jobs. The meaning lives past the statement, in the notes, because the statement, correctly, will never carry it.

Believed to be firm

The filing discloses the tab itself two sentences further on. Remaining performance obligations of $98.8 billion, of which 36 percent is expected to be recognized in the first twenty-four months, 39 percent between months twenty-five and forty-eight, and the remainder between months forty-nine and eighty-four. (14) The footnote books year seven. A booking for year seven is not a rental; the disclosure was built for promises with tenor, its rule permitting contracts of a year or less to be omitted entirely, and this figure’s own bands place nearly two-thirds of it beyond the first twenty-four months. (15) And the figure is net of the company’s own estimates: the definition subtracts “potential reductions to the transaction price in the future, such 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.” (16) The tab is stated after the company’s forecast of its own credits, delays, and resales. Estimates, inside the aggregate, disclosed as such.

Form follows the presentation judgment here too. The tab exists as an RPO because the contracts are presented as services. Presented as leases, the same commitments would appear instead in a lessor’s maturity table: undiscounted payments to be received, year by year for each of the first five years, with a total for the years beyond. (17) The lease form of the disclosure is more granular than the service form’s bands. Whichever presentation is chosen, the chosen one here discloses less about time.

And the categorization answers less than it appears to. The RPO does not evidence the services reading; it presupposes it. It inherits the judgment, so it cannot test it. Nor is the tab on the balance sheet: the funded portion, cash received ahead of performance, is the deferred revenue liability; the remainder of the $98.8 billion is a disclosed aggregate, neither asset nor liability. The polarity of the presentation runs through everything downstream. Presented as services, the future dollars stand mostly outside the statements, and the recognized sliver is performance the company owes. Presented as leases at full depth, the same dollars become an asset, a receivable, and assets carry valuation machinery that disclosures do not: allowances, measured against the credit of named counterparties, tested each period. Under the reading on file, the accounting asks no one to assess the collectibility of the $98.8 billion, because nothing about it is an asset. Disclosures are stated; assets are tested. The one movement on the record this quarter ran in the other direction: the $1.3 billion reclassified to customer liabilities.

The load the tab carries points the same way. A seven-year revenue commitment, load-bearing for financing already in place and for build-out commitments already made, sits on its face at odds with a presentation of a services or rental business.

For more than thirty years, the disclosure rules had a name for this figure: backlog, and the required words were “dollar amount of backlog orders believed to be firm.” The qualifier was carried on the face of the requirement. That requirement was removed in 2020, in favor of principles, two years after the revenue standard installed the successor. (18) The successor’s name is remaining performance obligations, which asserts more and confesses less, and the figure under it is now larger than the balance sheet beneath it. The accounting board reviewed the standard that created it and closed the review, concluding the standard works as intended, and opened no project to revisit it. (19) The last time obligations of this character accumulated in footnotes at national scale, the repair took fourteen years from the study that found them to the standard that surfaced them. (20)

How it goes missing

Every statement is correct under its own rules. Every analyst reads the pieces their work relies on, and reads them correctly. And the profession’s own consistency check runs across periods, not across presentations: the auditing standard titled Evaluating Consistency of Financial Statements asks whether “the comparability of the financial statements between periods has been materially affected by changes in accounting principles or by material adjustments to previously issued financial statements.” (21) This period against last period. Nothing is skipped, and nothing here says otherwise.

What remains is a distinction, not a fault. The statements articulate: net income rolls to equity, cash rolls to cash, every figure that must equal another figure equals it, and an entire industry polices the tying. Whether the statements agree, whether the three share one theory of what the business is, is a different question. The cash flow statement, identical under every characterization of the business, is the fixed point that makes the other two statements’ choices visible as choices; and the duties broad enough to reach the question of agreement, the certification that the filings fairly present and the rule requiring whatever further information is necessary to make the statements not misleading, are tested after the fact. (22) In the meantime, the question sits with whoever holds the statements.

Which returns everything to the critical audit matter and the judgment call, because the record already points there, correctly filed. The presentation of the contracts as services is a judgment the standards commit to management. Management exercised it, and the exercise is not gradable as wrong, because it is a judgment. The auditor of the annual statements then did its own job: it identified that judgment, along with the consolidation accounting for a data-center joint venture, as the audit’s critical matters. (23) The disclosure is addressed to whoever holds the statements; it says: this choice was critical, weigh it. The two matters flagged are not details inside the portraits. They are the portraits’ boundaries, the first deciding which income statement this company has, the second whether an entity stands inside these statements or outside them. Management chose. The auditor flagged the choosing. And the weighing was handed, explicitly, to the one seat at the table with no rulebook: the holder of the whole.

The rulebook behind that flag is specific, and it was followed to the letter. A critical audit matter is any matter, drawn from those communicated or required to be communicated to the audit committee, that relates to accounts or disclosures material to the financial statements and involved especially challenging, subjective, or complex auditor judgment. (24) For each one, the auditor must do four things: identify it, describe the principal considerations that led to the determination, describe how it was addressed in the audit, and refer to the relevant accounts or disclosures. The auditor may not use the communication to publish company information that is not already public, so the channel cannot reveal the contracts; it can only point at their disclosed shadow. (25) And the standard prescribes the sentence that must precede every such communication: that communicating these matters “does not alter in any way our opinion on the financial statements, taken as a whole,” and that the auditor is not, by communicating them, providing separate opinions on the matters or on the accounts they relate to. This piece’s title is quoted from that required sentence. The rulebook separates, in one breath, the opinion on the whole from the flag on the matter; the pointing is expressly not an opinion. Identify, explain, describe, refer. Four duties, four performances, to the letter.

So the decision before whoever holds the statements is the same one the auditor flagged, and it is simple to state: is this a services business? The RPO does not answer it; the RPO takes the answer as given. The statements do not answer it; each is correct under either answer, within its own rules. The contracts answer it, privately. What the statements carry is the set of judgments built on that answer, and relying on those judgments, and testing them, is what financial statements are there for.

One more fact about the pointer: it is annual. Critical audit matters exist only in the auditor’s report on the audited annual statements. The quarterly filings between them, including the one this piece reads, are unaudited; they receive a review, inquiries and analytical procedures, which expresses no opinion and carries no critical audit matters. (26) The footprints land quarterly. The pointer updates once a year. For the three quarters between annual reports, whoever holds the statements holds them alone. Stated plainly: the critical audit matters belong to an audited annual presentation; the reclassification belongs to an unaudited quarterly update. They are different classes of document, and the record points forward on its own terms: the next audited annual statements are the venue where matters of this kind may be more fully discussed and disclosed.

Two consistent companies, and a third

Suppose the document were made to agree with itself. There are two ways, and each has a price.

Make it a rental business everywhere, and the balance sheet must follow the income statement: short obligations, minimal prepayment float, no month-eighty-four band, and, above all, no financing of the kind the filing describes. The debt footnote lists non-recourse facilities and vendor financing arrangements tied to equipment additions, and the filing describes what the term lenders hold: the facilities are “collateralized with the assets underlying the contributed contracts and the pledged contractual cash flows.” (27) That is lending against contracts. A contract can be contributed only if it exists, and its cash flows pledged only if they run long enough to repay the draws; a short-cycle rental book has neither to offer. A company financed contract by contract cannot, consistently, be a short-cycle rental operation. The rental reading, applied to all three statements, leaves the company to borrow on some other basis than the contracts, at whatever terms a pool of depreciating machines earns on its own name. The filing’s own liabilities are the evidence that nobody lent to that company.

Make it a lease business everywhere, and the change arrives at two depths. At the shallow depth, operating-lease treatment, the changes are few: the machines stay on the balance sheet, depreciation continues, the cash flow statement does not change a line, and the income statement re-captions, lease income for service revenue, the same dollars under different words, with timing differences where the consideration is variable or usage-based. The visible change is the tab, which dissolves as an RPO and reappears as the maturity table of note (17), disclosing more about time than the bands it replaces. Counted in changed lines, this is the smallest alternative on the table: no line of the cash flow statement, no dollar of the balance sheet, the captions of one statement and the form of one note.

At the full depth, for contracts that transfer substantially all the economics of the machine, dedicated assets, terms consuming most of the equipment’s life, financing components the accounting already admits, the lease is a sale with financing attached: the equipment leaves the balance sheet and a net investment in leases arrives, a receivable, a financial asset, and the credit assessment the tab has never received arrives with it. (28) The customers, at either depth, book right-of-use liabilities on their own balance sheets. At the full depth the income statement becomes a financing spread over residual risk, and the multiple, wherever the market sets it, is thereafter a multiple of the machine and not of the word.

The third company is the one on file: rental on the income statement, lessor on the balance sheet, financier in the cash flows, each per the rules. It is the version in which no one is required to book the debt, assess the credit by name, or re-price anything, because no single statement, read under its own rules, calls for it. It is internally consistent, and it holds. What it leaves unanswered is the question the three raise only together, and that question stays open until something on the record closes it.

The question

The contracts know which business this is. Whether the provider can substitute the machines, whether the customer directs their use: these are facts, held by two parties who both know the answer, in documents the public does not read. Assume the contracts are well made and do exactly what they were written to do; whatever tensions exist between the statements’ postures are resolved there, privately. Nothing in this piece requires otherwise. The filing is not required to say, and it does not.

Nothing on the record says the income statement’s presentation is the way to hold the business; the record says only that management chooses to hold it this way, and that the choice was critical. Both of those statements are correct. Neither is an answer. And the presentation is correct in a further sense: the standards commit the judgment to management, and the exercise of a committed judgment is its correctness. What stays open is not whether the presentation is correct but how it is to be read. This piece does not supply the reading either, because it cannot: the presentation may be exactly right in substance as well as in form, the contracts may be genuinely and substantively services, and each statement’s portrait may be the correct rendering of a business that is simply hard to draw. We note what the documents themselves put on the table: that the lenders’ facilities and the income statement’s caption rest on different answers to the same question, and that the three statements, held together, put that question to anyone willing to hold them. That is all, and it is enough.

The register solved this problem before the industry existed, which is why it keeps returning to these pieces. At the drawer, the word and the cash were one event: the sale, the record, and the bell, simultaneous, in public. Modern statements distribute that event across classifications, correctly, and so the classifications are where the listening happens now. Words are not the decoration of these documents; they are the definitions a change must pass through to become visible. The reclassification is noted here, and it merits notice, because it is exactly that: a word changing on the record.

This piece reaches no conclusion, because the reading it performs does not produce one. It produces a requirement: a comprehensive reading of this company needs everything, three statements, the notes beneath them, the annual report and its critical matters, the credit documents behind the debt table, held at once, because the meaning is distributed across them and resides in no single place. And it produces a baseline. The next quarterly update re-measures the balances quoted here: the customer-liabilities line and how it resolves, the deferred revenue that fell $0.7 billion in the quarter while debt rose $3.5 billion, the bands of the tab. (29) The next annual report renews the audited presentation and its critical matters, and that comparison, audit against audit, this reading against the next, is where more may become visible. Nothing here concludes. The nearer question, what the reclassified line is, resolves inside the coming year, as the balance discharges, or is explained, or moves again; any of those is an answer. The larger question, whether this is a services business, keeps the audit’s timetable, not the quarter’s. The tab still stands at $98.8 billion; the record, held whole, has now been read once, so the next reading has something to be held against, and we will be reading for both.

NOTES

(1) CoreWeave, Inc., Form 10-Q for the quarterly period ended March 31, 2026, filed May 8, 2026 (unaudited condensed consolidated financial statements). All quotations and figures in this piece are from this filing unless otherwise noted.

(2) Condensed consolidated statements of operations: revenue of $2,078 million (three months ended March 31, 2026) versus $982 million (2025); cost of revenue $716 million; technology and infrastructure $1,273 million; operating loss $(144) million; net loss $(740) million.

(3) Condensed consolidated statements of cash flows, reconciliation of net loss to operating cash: depreciation and amortization of $1,147 million (three months ended March 31, 2026) versus $443 million (2025). MD&A: technology and infrastructure expense “consists of costs associated with our infrastructure, such as depreciation and amortization related to our servers, switches, networking equipment and internally developed software”; cost of revenue includes “depreciation of power installation and distribution systems.” The statements of operations carry no separate depreciation caption; each expense caption absorbs its allocable share, and the allocation between captions is not quantified in the filing.

(4) Condensed consolidated balance sheets: property and equipment, net, $36,424 million; operating lease right-of-use assets $10,182 million; total assets $55,573 million. The two asset captions sum to 84 percent of total assets.

(5) Property and equipment note: technology equipment $26,627 million; software $827 million; data center equipment and leasehold improvements $3,878 million; furniture, fixtures, and other assets $20 million; construction in progress $9,581 million; total property and equipment $40,933 million; less accumulated depreciation and amortization $(4,509) million; net $36,424 million. Construction in progress is 23 percent of gross property and equipment. The note states depreciation and amortization on property and equipment of $1.1 billion for the quarter, and $97 million of interest capitalized.

(6) Total debt, net of unamortized discount and issuance costs, $24,859 million (principal $25,149 million); deferred revenue, current and non-current, $7.5 billion at March 31, 2026.

(7) Condensed consolidated statements of cash flows: net cash provided by operating activities $2,984 million (versus $61 million); net cash used in investing activities $(7,708) million; net cash provided by financing activities $3,914 million, including proceeds from issuance of debt, net, $3,290 million; net decrease in cash, cash equivalents, and restricted cash $(810) million.

(8) Interest expense, net, $(536) million (versus $(264) million). MD&A, liquidity: “our cash flows dedicated for debt service requirements totaled $1.8 billion, which includes principal payments of $1.3 billion and interest payments of $459 million, inclusive of $95 million related to capitalized interest.” The property and equipment note separately reports $97 million of interest capitalized during the quarter; the two figures measure payment and capitalization respectively, and are the filing’s own.

(9) Notes to condensed consolidated financial statements, Contract Balances, quoted in full as it appears.

(10) Other current liabilities note: customer liabilities of $1,469 million at March 31, 2026 and $137 million at December 31, 2025. The filing does not further define the caption or state the cause of the reclassification. The caption does not appear in the Company’s Form 10-K for fiscal year 2025 (filed March 2, 2026); at December 31, 2025 the $137 million the quarterly filing labels customer liabilities was carried within that annual report’s undifferentiated other current liabilities of $162 million. The Form 10-Q introduces the line, and neither filing defines it. Balance sheet: deferred revenue, current, of $2,130 million at March 31, 2026 and $1,709 million at December 31, 2025; deferred revenue, non-current, of $5,393 million and $6,476 million at the same dates. The bucket movements, $1,083 million down beyond twelve months and $421 million up within, net to the $0.7 billion decline in total deferred revenue.

(11) Cash and cash equivalents of $2,244 million at March 31, 2026, excluding restricted cash. MD&A reports total liquidity of $11.1 billion, of which $8.8 billion is availability under the revolving credit facility and delayed draw term loan agreements; the filing describes borrowings under the term facilities as used primarily to finance the acquisition and installation of computing infrastructure, and remaining revolver capacity was $686 million.

(12) Concentration disclosure: Customer A at 68 percent and Customer D at 11 percent of accounts receivable, net, as of December 31, 2025; as of March 31, 2026, Customers A, B, and C at 39, 17, and 22 percent. The concentration sentence closes the passage directly above the Contract Balances section, two sentences before the reclassification sentence; the adjacency is a fact of the document’s layout, no connection is stated in the filing, and none is asserted here.

(13) Supplemental disclosures of cash flow information, non-cash investing and financing activities: “Reclassification of liabilities related to property and equipment additions to debt upon execution of OEM financing arrangements, $1,469 million” (three months ended March 31, 2026); “Reclassification of customer deposit to debt, $230 million” (three months ended March 31, 2025). The $1,469 million figure coincides numerically with the customer-liabilities balance in note (10); they are different items. Three separate $1.3 billion figures likewise appear in this filing (the reclassification to customer liabilities; the outstanding balance of the DDTL 4.0 facility; the quarter’s principal payments) and are unrelated.

(14) Contract balances note: $98.8 billion of unsatisfied remaining performance obligations, with the timing bands as quoted.

(15) ASC 606-10-50-14 permits omission of the disclosure for contracts with an original expected duration of one year or less; the expedient is elective, and the filing does not state whether it is applied. The two-thirds figure is the complement of the 36 percent band at note (14).

(16) Contract balances note, definition of RPO, quoted verbatim.

(17) ASC 842-30-50 requires a lessor to disclose a maturity analysis of lease payments to be received, showing undiscounted cash flows on an annual basis for a minimum of each of the first five years and a total for the years thereafter. The RPO bands are as quoted at note (14).

(18) Regulation S-K Item 101(c), as in effect for more than thirty years, required disclosure of “dollar amount of backlog orders believed to be firm”; the SEC’s modernization of Items 101, 103, and 105 (adopted August 26, 2020, effective November 9, 2020) replaced the prescribed list with a principles-based approach under which backlog is disclosed if material.

(19) FASB, Post-Implementation Review Report: Topic 606—Revenue from Contracts with Customers, presented to the Financial Accounting Foundation Board of Trustees in November 2024: the report concludes the standard meets its intended purpose, observing that Topic 606 “provides understandable and operational guidance that can be applied as intended”; the IASB’s parallel review of IFRS 15, published September 30, 2024, concluded the requirements are “working as intended,” with remaining application questions deferred to a future agenda consultation.

(20) SEC staff, Report and Recommendations Pursuant to Section 401(c) of the Sarbanes-Oxley Act of 2002 On Arrangements with Off-Balance Sheet Implications, Special Purpose Entities, and Transparency of Filings by Issuers (June 15, 2005), estimating by extrapolation that there “may be approximately $1.25 trillion in non-cancelable future cash obligations committed under operating leases that are not recognized on issuer balance sheets, but are instead disclosed in the notes to the financial statements,” and recommending “that the accounting guidance for leases be reconsidered”; ASU 2016-02 (Topic 842) was issued February 2016 and effective for public companies in 2019, fourteen years after the study. The report is by the Commission staff, which notes the Commission expressed no view on its findings.

(21) PCAOB AS 2820, Evaluating Consistency of Financial Statements, paragraph .02, quoted verbatim.

(22) Exchange Act Rules 13a-14 and 15d-14 (certifications that the report does not contain untrue statements or omissions and that the financial statements “fairly present in all material respects the financial condition, results of operations and cash flows”); Regulation S-X Rule 4-01(a) (financial statements must include “such further material information as may be necessary to make the required statements, in the light of the circumstances under which they were made, not misleading”). Certification language conformed to Exhibits 31.1 and 31.2 as filed with the Form 10-Q.

(23) Deloitte & Touche LLP, Report of Independent Registered Public Accounting Firm, CoreWeave, Inc. Form 10-K for the fiscal year ended December 31, 2025 (filed March 2, 2026), discussed in our June 13 piece. Two critical audit matters. The first, Revenue Recognition, names among the significant judgments for certain customer contracts “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, and the identification and treatment of contract terms that may impact the timing and amount of revenue recognized.” The second, Consolidation Accounting for a Variable Interest Entity, concerns whether the company is the primary beneficiary of, and so must consolidate, a joint venture with a third-party infrastructure developer formed to acquire and develop a multi-phase data center campus. A critical audit matter identifies decisions whose making involved especially challenging, subjective, or complex judgment and commends them to the attention of whoever holds the statements; it is not a ranking of business risk, and it is not a criticism.

(24) PCAOB AS 3101, The Auditor’s Report on an Audit of Financial Statements When the Auditor Expresses an Unqualified Opinion, Appendix A (definition of critical audit matter) and paragraphs .11 through .14 (determination, communication, and documentation requirements); the four communication elements are as enumerated in the standard.

(25) AS 3101.15 (prescribed language preceding critical audit matters, quoted in part); Note 2 to paragraph .14 (the auditor is not expected to provide information about the company that has not been made publicly available by the company, unless necessary to describe the principal considerations or how the matter was addressed).

(26) The condensed consolidated financial statements in the Form 10-Q are unaudited. Interim financial information of SEC registrants is reviewed by the independent auditor (PCAOB AS 4105, Reviews of Interim Financial Information; Regulation S-X Rule 10-01(d)); a review consists principally of inquiries and analytical procedures, and expresses no opinion. Critical audit matters are communicated only in the auditor’s report on the audited annual financial statements.

(27) Debt note: the DDTL 4.0 facility is described in the filing as “a non-recourse delayed draw term loan facility,” with an outstanding balance of $1.3 billion at March 31, 2026; OEM financing arrangements per note (13). MD&A, liquidity: the delayed draw term loan facilities “are collateralized with the assets underlying the contributed contracts and the pledged contractual cash flows, generally from investment grade counterparties.” The underlying credit agreements and their collateral and covenant terms are filed as exhibits and are the subject of continuing review.

(28) Under the lease reading, lessor classification (sales-type, direct financing, or operating) would be determined contract by contract under ASC 842; a net investment in leases is a financial asset within the scope of the credit-loss standard (ASC 326). The invariance of the cash flow statement asserted earlier rests on ASC 842-30-45-5, under which a lessor classifies cash receipts from leases within operating activities “regardless of whether the lease is classified as a sales-type, direct financing, or operating lease” (rationale at BC335 of ASU 2016-02: leasing is the lessor’s revenue-generating activity); the lone exception, depository and lending lessors within the scope of ASC 942, does not apply to the company. Machine purchases remain investing and debt remains financing under every reading, and the commencement recognition of a net investment is non-cash. No restatement is asserted or implied; the paragraph describes the disclosure architecture each consistent reading would produce.

(29) Deferred revenue of $8.2 billion at December 31, 2025 and $7.5 billion at March 31, 2026; total debt principal of $21,615 million and $25,149 million at the same dates; scheduled principal payments of $6,066 million for the remainder of 2026 and $5,652 million for 2027.

The author holds no position in the securities discussed.

It is the start of a long holiday weekend in the United States, the 250th Fourth of July, which seems the right occasion for a lighter key, and for a piece about words. One word in particular: cloud, which arrives today in variants, neoclouds, hyperscalers, sovereign clouds, plain cloud computing, and which carries the majority of Oracle’s revenue, a $56 billion buildout, and forty-year money. The country has spent two and a half centuries arguing over what fixed words govern as the world moves under them. The computer industry ran the same experiment twice, faster, with the same company standing in the middle both times: once with SQL, the ledger’s language, which held still for forty years, and once with cloud, which held still for none of them. The difference was the test.

The industry that coins these words began, in Dayton, Ohio, in 1879, as an honesty machine. James Ritty, a saloonkeeper whose bartenders were stealing, built a device to make the cash tell the truth and named it for its purpose: the Incorruptible Cashier. A drawer that opened only when the sale rang. A bell, so the whole room heard the transaction commit. In time, a paper tape: the journal, written by the machine. John Patterson, a coal dealer whose own books came straight within months of installing two of them, bought the company, renamed it the National Cash Register Company, and invented the modern sales profession to sell one product: trust in the drawer. (1)

The line from that machine to the computer is an org chart. Patterson’s star salesman, Thomas Watson, fired in 1914, carried NCR’s sales gospel to a holding company of tabulators, time clocks, and scales, three more ledgers, for people, hours, and weight, and renamed it International Business Machines. (2) NCR itself crossed from brass to electronics in its turn, registers becoming calculators becoming computers. And when IBM’s researchers, two generations on, formalized what a computer must promise when it touches money, all or nothing, a correct state to a correct state, durable once done, they were restating the register’s warranty as mathematics. The transaction was defined at the same IBM project that produced SQL, in the same body of papers. (3) The language and the ledger arrived in one envelope.

The papers in that envelope began in 1970, when a mathematician at IBM Research named Edgar Codd published a paper proposing that data be stored in relations, tables, and queried by logic rather than by location. (4) IBM built a research prototype, System R, and its team invented a query language for it, eventually called SQL. (5) IBM published the papers and then hesitated to commercialize the work, protecting an older product line.

Three men at a small California consultancy read those papers and did not hesitate. Larry Ellison, Bob Miner, and Ed Oates built a commercial implementation of IBM’s published language and shipped it in 1979, as Oracle Version 2, the first commercial SQL database on the market, arriving before IBM’s own. (6) The company’s commitment to the standard word was total: Ellison telephoned IBM’s Don Chamberlin, an author of the language, to ask for System R’s error numbers, so that Oracle would match the original down to its failure codes. IBM declined to share them. (7) The founders divided their own labor by comparative advantage, and, the others being the stronger engineers, Ellison took sales. (8) The company began as an act of reading someone else’s disclosure, and its founder began as its salesman.

Vendors found the new word quickly. As “relational” became fashionable in the early 1980s, older technology was relabeled to wear it, and Codd published his twelve rules to define what the word required, fighting what one account calls a sometimes bitter campaign against the veneer. (9) What settled the matter was a referee.

In 1988 the federal government adopted SQL as a Federal Information Processing Standard, FIPS 127, revised in 1990 and again in 1993. (10) The standard did not suggest.

“Relational model database management systems acquired for Federal use shall implement FIPS SQL.” (11)

Whether developed internally, bought, leased, or contracted. And the government did not stop at defining: NIST built and administered a conformance test suite, programs and schemas and validated-products lists, and the largest buyer of information technology on earth put its procurement money behind passing. (12) A participant in that era later wrote that the arrangement worked precisely because the standard was developed, maintained, interpreted, and tested by an independent member of the user community rather than by the vendors, which made vendors’ conformance claims worth believing. (13)

Under the tested word, SQL stayed SQL. The language Oracle implemented from IBM’s papers in 1979 is, extended but recognizable, the language running today. And Oracle’s own documentation still carries the certificate, in an appendix that reads, in full seriousness:

“Oracle complied fully with last Federal Information Processing Standard (FIPS), which was FIPS PUB 127-2. That standard is no longer published.” (14)

The company still attests conformance to a test nobody administers. It sits in the current manuals like a fossil in limestone, and it is the last artifact of the era when a word in this industry had a referee.

The testing ended in the mid-1990s, when NIST terminated its database standardization work, including the SQL validation program. (13) The standard itself was formally withdrawn in February 2005, in a batch of seventeen obsolete FIPS, for a stated reason that has governed every word coined since: federal agencies were now directed “to use technical standards that are developed in voluntary consensus standards bodies,” so standards that duplicated industry’s own “are no longer needed.” (15) The government handed the dictionary to the vendors. It was done deliberately, lawfully, and in the Federal Register. Every word this industry has coined since has lived in the world that notice made.

Cloud arrived into that world, and the most clear-eyed description of what would happen to it was delivered by Ellison himself, at Oracle’s analyst meeting in September 2008, in a riff that has been quoted for laughs ever since and deserves to be quoted for accuracy instead:

“We’ve redefined cloud computing to include everything that we already do.”

He compared the industry’s vocabulary to fashion, orange as the new pink, cloud as the new SaaS, and pronounced the term gibberish. And then he told the room, in plain terms, exactly what Oracle would do about it:

“We’ll make cloud computing announcements, because if orange is the new pink, we’ll make orange blouses. I’m not gonna fight this thing.” (16)

That is a salesman describing, in advance and in public, the mechanics of a word without a referee: the term will be worn, not met. He then executed, brilliantly. Within two years Oracle’s clusters were, in his phrase on an earnings call, “now called private clouds... the more-fashionable term for clusters.” (17) The decade of selling that followed was run, from September 2010, by a co-president whose credential Oracle’s own announcement stated in one line, that he had “pioneered the integration of hardware with software when Teradata was a part of NCR”: Mark Hurd, twenty-five years at National Cash Register, a junior register salesman out of San Antonio risen through Dayton’s database division, who rebuilt Oracle’s salesforce and trained cohorts of graduates to sell the cloud. (18) The Incorruptible Cashier’s company had trained the man who sold the word.

The filings followed, and calendar 2011 produced two definitions of the same word. The government tried to be a referee again that year: in September 2011, in its sixteenth and final version, NIST published its definition of cloud computing. (19)

A model of ubiquitous, on-demand network access to a shared pool of configurable computing resources: five essential characteristics, on-demand self-service, broad network access, resource pooling, rapid elasticity, measured service; three service models; four deployment models.

Oracle’s annual report, filed that June, gave the word its first definition in the company’s own disclosure. Cloud computing, it explained, had:

“evolved from technologies and services that Oracle has provided for many years.” (20)

It is the 2008 joke, filed: the redefinition of the word to include everything the company already did, converted from mockery into accounting in three years, without changing a clause. The government’s is a careful document. FIPS 127 had arrived with a test suite, a validated-products list, a shall, and procurement money behind it; the cloud definition arrived alone. NIST later published guidance for evaluating whether a service meets it, which is advice, and advice is not a test. (21) So one year, one word, two definitions: the government’s, with five characteristics and no referee, and Oracle’s, with one characteristic, that Oracle already did it. Only one of them governed a set of financial statements.

From there the word moved through Oracle’s filings the way an untested word moves. In fiscal 2013 it reached the money: a segment was retitled “new software licenses and cloud software subscriptions,” and the word had a revenue line. (22) In fiscal 2019 it did something more consequential, and quieter. Through fiscal 2018, Oracle’s releases reported a standalone cloud figure, disaggregated: in the third quarter of that year, SaaS revenue of $1.2 billion, platform and infrastructure of $415 million, total cloud revenue of $1.6 billion, up 32 percent. (23) The next fiscal year the standalone figure was gone. In its place stood a fused line, “cloud services and license support,” $6.6 billion in its first quarter: the new word and the old installed-base annuity, reported together, the split undisclosed. (24) The word’s territory quadrupled in a quarter, by absorbing the business next door. No bell rang. The fused line ran seven annual reports, fiscal 2019 through fiscal 2025, and closed its final year at $44.0 billion, two-thirds of the company.

Then the word moved again, in the filing this shop finished reading last week. The fiscal 2026 income statement splits the line back apart: Cloud, $33,989 million; Software, $24,541 million, and the recast prior years show the software line sitting essentially where it sat two years ago, the annuity filed flat, on the statement’s face, for the first time. (25) The prior periods were restated to the new labels, and the change itself travels in one sentence of the accounting-policies note, reclassifications made “to conform to the current period’s presentation,” which “did not affect total revenue.” No bell rang this time either. And the boundary breaks any reader who steps across it: the line named cloud carried $44 billion in fiscal 2025 and $34 billion in fiscal 2026, a ten-billion-dollar shrinkage in the year the business named cloud grew 39 percent. The word changed territory, and every comparison across the change breaks on contact.

Notice the shape of the move. The split arrived in the year the buildout became the story, and it de-fused exactly one side of the statement. Revenue, the growth story’s side, now comes apart into cloud and software. Cost stays fused: “cloud and software” expenses, one line, $17.6 billion. (26) The disaggregation the growth story wants is filed. The disaggregation the margin question needs, the cost of delivering a cloud dollar, remains inside the combined segment and its single expense line, which is the one column this shop spent three thousand words reading from the outside. The fusion has a birthday, fiscal 2019. The question it created is still open in fiscal 2026, one lawful presentation at a time.

By fiscal 2026 the word carries the majority of Oracle’s revenue, a $56 billion capital program, and the forty-year notes priced in that piece’s table. The word Ellison called gibberish in 2008 is now the word in whose name his company borrowed money due in 2066, and it has a line of its own again, at last, with the costs still standing behind the curtain.

Everything in this walk is good business. A man who mocks a word in September 2008, announces in the same breath that he will wear it, and then wears it better than anyone alive is executing a strategy he stated in advance. He assessed, correctly, that the word had no referee and that fighting fashion is a losing trade, and he has moved this industry, partly to his benefit and partly to everyone’s, more than once, on exactly that kind of read. It is a record of competence, and it is read here as one.

Financial reporting runs on the second model. The segment rules ask how management views the business; the filer defines its own labels; the auditor audits the numbers under the label and leaves the label to the filer. There is no validated-products list for “cloud services and license support.” That is the world the 2005 Federal Register notice made, applied to disclosure, and it is lawful, and it is the environment every reader of a modern filing works in, whether they know it or not.

So: when the word had a test, it held still for forty years, and Oracle still carries the certificate. When the word had only definitions, it traveled, from an insult, to an environment, to a segment, to a fused line that swallowed the business next door, to the collateral description under forty-year money. Every comparison across those boundaries quietly broke, because the word under the numbers was not the same word. Which is the lesson, and it is short enough for a holiday weekend: a definition tells you what a word should mean. Only a test tells you what it does mean. The first machine this industry ever sold understood that; it rang a bell, loud enough to be heard through the first cloud this industry ever knew, the smoke of a Dayton saloon. Filings come with definitions. The test is the reading.

My thanks to Pops, who encourages us all to listen for the bell.

NOTES

(1) James Ritty’s cash register, patented as “Ritty’s Incorruptible Cashier” (Dayton, Ohio, 1879), built to deter theft at his saloon; the drawer, bell, and printed receipt roll were added as the machine was commercialized. John H. Patterson, a Ritty customer for his coal business before he was an owner, acquired the company in 1884 and renamed it the National Cash Register Company; his “American Selling Force” and the N.C.R. Primer are generally credited as the first organized sales force and first sales training manual.

(2) Thomas J. Watson Sr., NCR’s sales chief, convicted with Patterson and other NCR executives under the Sherman Act in 1913 (overturned on appeal, 1915), dismissed by Patterson, and hired in May 1914 to run the Computing-Tabulating-Recording Company, Charles Flint’s 1911 amalgamation of the Bundy and International Time Recording companies, the Computing Scale Company, and Hollerith’s Tabulating Machine Company; renamed International Business Machines in 1924.

(3) The transaction abstraction was defined by members of the System R team: K.P. Eswaran, J. Gray, R.A. Lorie, and I.L. Traiger, “The Notions of Consistency and Predicate Locks in a Database System,” Communications of the ACM (1976); the ACID formulation is T. Härder and A. Reuter, “Principles of Transaction-Oriented Database Recovery,” ACM Computing Surveys (1983). J. Gray received the 1998 Turing Award for contributions to database and transaction processing research.

(4) E.F. Codd, “A Relational Model of Data for Large Shared Data Banks,” Communications of the ACM, June 1970.

(5) IBM, System R project history; the language was initially SEQUEL, renamed SQL for trademark reasons.

(6) Oracle Version 2, released 1979 by Relational Software, Inc. (formerly Software Development Laboratories), the first commercially available SQL relational database. Version 1 was internal; there was no commercial Version 1.

(7) Account in the foreword to “Oracle SQL: The Essential Reference” (O’Reilly): RSI developed “a commercial product that was as compatible as possible with the prototype being developed at IBM’s research facilities,” to the point that “Larry Ellison himself called Don Chamberlin at IBM to request the error numbers that the system used.” IBM’s refusal to share System R error codes is separately recorded in standard histories.

(8) Standard accounts of the 1977 founding of Software Development Laboratories: Ellison, Miner, and Oates allocated roles with Ellison on sales, the others being the stronger engineers.

(9) E.F. Codd’s twelve rules (1985), published amid what one account describes as a campaign “to prevent the term from being misused by database vendors who had merely added a relational veneer to older technology.”

(10) FIPS PUB 127 (1988); FIPS PUB 127-1 (February 1990); FIPS PUB 127-2 (1993), keyed to ANSI X3.135-1992 (SQL-92), with Entry, Transitional, and Intermediate conformance levels.

(11) FIPS PUB 127-2, Section 11.2, Acquisition of SQL Implementations: “Relational model database management systems acquired for Federal use shall implement FIPS SQL,” with conformance required across internal development, procurement, leasing, and contracted programming services.

(12) NIST SQL Test Suite (final version 6.0), covering Interactive SQL, Embedded SQL in multiple host languages, and Module Language bindings; NIST published validated-products information during the program’s life.

(13) “The Role of NIST in SQL Standardization” (TDAN, 1997), written at the termination of NIST’s database standardization participation, including SQL validation test development.

(14) Oracle Database documentation, appendix “Oracle Compliance with FIPS 127-2”: “Oracle complied fully with last Federal Information Processing Standard (FIPS), which was FIPS PUB 127-2. That standard is no longer published.”

(15) Federal Register, February 8, 2005, “Announcing Approval of Withdrawal of Seventeen (17) Federal Information Processing Standards (FIPS) Publications,” including FIPS 127-2, Database Language SQL; withdrawal rationale per the National Technology Transfer and Advancement Act of 1995 (Pub. L. 104-113).

(16) Oracle financial analyst meeting, Oracle OpenWorld, September 2008. Transcribed variants of the remarks appear in CNET (”Oracle’s Ellison nails cloud computing,” September 26, 2008), CircleID’s transcript, InformationWeek (from the meeting audio), and InfoWorld’s 2008 quotes roundup; the sentences quoted are common to the transcripts, and the sequence of the announcements, blouses, and fight remarks follows InformationWeek’s transcription of the audio.

(17) Oracle earnings call remarks, quoted in InformationWeek, “Oracle’s Larry Ellison Embraces Cloud Computing’s ‘Idiocy’” (2010).

(18) Oracle Corporation press release, September 2010 (Form 8-K exhibit), announcing Mark Hurd as president, quoting Lawrence J. Ellison: “Mark pioneered the integration of hardware with software when Teradata was a part of NCR.” Hurd joined NCR Corporation as a salesman in San Antonio in 1980, spent twenty-five years there including as head of its Teradata division and as president, chief operating officer, and chief executive officer, and served as Oracle’s president from September 6, 2010, and co-chief executive officer from 2014 until his death in October 2019. His rebuilding of Oracle’s salesforce and the “Class Of” program that hired and trained college graduates to sell Oracle’s cloud services are matters of public account.

(19) NIST Special Publication 800-145, “The NIST Definition of Cloud Computing” (September 2011), the sixteenth and final version of a working definition begun in November 2009; public draft January 2011.

(20) Oracle Corporation Form 10-K, fiscal year ended May 31, 2011: cloud computing described as having “evolved from technologies and services that Oracle has provided for many years.”

(21) NIST SP 500-322, “Evaluation of Cloud Computing Services Based on NIST SP 800-145.”

(22) Oracle Corporation Form 10-K, fiscal year ended May 31, 2013: segment retitled “new software licenses and cloud software subscriptions.”

(23) Oracle fiscal Q3 2018 earnings release (Form 8-K, Exhibit 99.1, March 19, 2018): “Cloud Software as a Service (SaaS) revenues were up 33% to $1.2 billion. Cloud Platform as a Service (PaaS) plus Infrastructure as a Service (IaaS) revenues were up 28% to $415 million. Total Cloud Revenues were up 32% to $1.6 billion.”

(24) Oracle fiscal Q1 2019 earnings release (Form 8-K, Exhibit 99.1, September 2018): “Cloud Services and License Support revenues were $6.6 billion.” The fused presentation coincided with Oracle’s adoption of the new revenue recognition standard at the start of fiscal 2019 and ran through the Form 10-K for fiscal 2025; Oracle’s fiscal 2025 fourth-quarter release (June 11, 2025) reported cloud services and license support revenues of $44.0 billion, up 12 percent, the fused line’s final annual figure.

(25) Oracle Corporation Form 10-K, fiscal year ended May 31, 2026, consolidated statements of operations: revenues of cloud, $33,989 million (with fiscal 2025 and 2024 recast at $24,506 million and $19,774 million); software, $24,541 million ($24,724 million and $24,690 million); hardware, $3,084 million; services, $5,743 million; total revenues, $67,357 million. The accounting-policies note states that prior year amounts were reclassified “to conform to the current period’s presentation for all periods presented” and that such reclassifications “did not affect total revenue, income from operations or net income.” The standalone cloud revenue line debuted in Oracle’s fiscal Q1 2026 earnings release, September 2025. Fiscal 2026 cloud growth of 39% is computed from the recast prior-year figure.

(26) Form 10-K, fiscal 2026, consolidated statements of operations: operating expense line “cloud and software,” $17,597 million, one line against the two revenue lines. The segment presentation likewise reports cloud and software as a single business; the MD&A sentence quoted in the prior piece, naming “cloud and software support agreements” together, appears verbatim in the fiscal 2026 filing.

Where the round stands

OpenAI announced the round on February 27 at $110 billion and closed it on March 31 at $122 billion committed, an $852 billion post-money valuation. The company named SoftBank as co-lead, alongside a16z, D. E. Shaw Ventures, MGX, TPG, and accounts advised by T. Rowe Price, and reported more than $3 billion raised from individual investors through bank channels. OpenAI named the three anchor amounts; each investor’s own filing set out the terms behind its piece.

Amazon committed $50 billion: $15 billion at the close, and $35 billion that arrives, in the companies’ words, “in the coming months when certain conditions are met.” The filing behind the round states the rest. The $35 billion is callable on five business days’ notice, and it is tied to the cloud agreement that accompanies it. If that agreement ends, the $35 billion obligation ends with it. Alongside the equity, OpenAI agreed to consume about two gigawatts of Amazon’s Trainium capacity.

NVIDIA committed $30 billion as equity at the close, a straight stake in place of the September 2025 letter of intent, which would have paid in up to $100 billion as OpenAI deployed its systems. Alongside the equity, OpenAI committed to about five gigawatts of NVIDIA’s Vera Rubin systems.

SoftBank committed $30 billion through Vision Fund 2, subject to customary closing conditions, financed initially through bridge loans and replaced over time through asset sales. Upon completion, its cumulative stake is expected to reach about $64.6 billion, near 13 percent. By SoftBank’s own terms, the closings accelerate if OpenAI lists.

A broader pool of about $12 billion completes the round, the $122 billion less the three anchor checks.

The total is $122 billion committed. At least $35 billion of it is contingent. The dollar figures are disclosed; the split between cash and services behind them stays with the company.

The lead chair moved

The lead on OpenAI’s rounds has changed hands since 2019.

Microsoft anchored the first two, in 2019 and 2023. The lead passed to Thrive Capital in October 2024, to SoftBank in March 2025, and SoftBank held it into March 2026, co-leading Series 7. Amazon and NVIDIA wrote the two largest checks in that round, $50 billion and $30 billion, and came in as anchors, behind the lead.

Across those rounds the valuation moved from about $29 billion to $852 billion. The parties writing the largest checks in the last round, and the parties holding the lead, are different parties.

What the filings say

Each commitment was disclosed by the party that made it. Read together, they share parties and reference points.

Amazon’s $35 billion and OpenAI’s Trainium purchase sit in one agreement, linked by its terms. NVIDIA’s equity and OpenAI’s purchase of NVIDIA systems were announced in the same round. SoftBank routes its $30 billion through the fund whose stake it carries on its own books, and its chairman, Masayoshi Son, chairs Stargate, the venture behind the buildout the round helps fund.

Two of those parties also appear in OpenAI’s audited 2025 accounts. About nine percent of OpenAI’s 2025 revenue came from Microsoft and SoftBank, both equity holders, and about half of the 2025 cost stack flowed to Microsoft. The compute the round helps pay for runs, in large part, on contracts with Oracle and Amazon that we read separately as well.

Each of these is on the record, in the filings and announcements of the parties to it. The round puts them on one valuation and one timetable.

What the S-1 will settle

OpenAI filed a draft registration statement with the SEC on June 8, confidentially. A confidential draft keeps the price, the share count, and the date open. The public S-1, when it lands, carries the items the round’s terms turn on.

The first is the cash-and-services split. The round reports dollar totals; the S-1 reports cash flows, and the share of the $122 billion that arrives as cash will show against the share that arrives as compute and services.

The second is the timing of the listing. Some of the round’s terms point to it. By SoftBank’s own terms, its closings accelerate if OpenAI lists, and the bridge that funds its stake comes due in March 2027. On the $100 billion that became $30 billion, NVIDIA’s Jensen Huang told an investor conference the larger figure was “probably not in the cards” because OpenAI is going public. Amazon’s $35 billion is conditional as well, by both companies’ announcements, though the filing redacts the milestones; reporting has placed them at a listing or an AGI declaration, which the companies leave unconfirmed.

The company has set the same event in front of itself. Sam Altman told staff he expects OpenAI to go public “within the next year,” and that “many things could cause it to be sooner or later in that range.” The S-1 announcement said a listing may be a while off, with steps the company would rather take private first.

So the “when” is a calculation before it is a date. It runs on the dated terms, the deadlines and the contingencies, and on the cash requirement: what OpenAI spends, and the cash the round delivers to meet it. The cash on hand and the burn are on the record now. Ed Zitron’s reporting of OpenAI’s audited 2025 accounts, verified by the Financial Times, put the year-end cash at roughly a year of runway against the operating loss, which we added up separately. The committed total is one quantity, the cash and the burn another. What still joins them is the cash-and-services split, how much of the $122 billion is cash and when it lands. The S-1 is where that comes from, and the listing is the event the company has set its terms against.

Analysis: Cape Fear Advisors. Full figure-by-figure sources below.

Sources

Figures come from the parties’ own disclosures wherever those exist, and from named reporting where they do not. Each entry lists what it supports.

Primary disclosures

OpenAI, “OpenAI raises $122 billion to accelerate the next phase of AI,” March 31, 2026. The round closed at $122 billion in committed capital and an $852 billion post-money valuation, up from the $110 billion announced February 27 at a $730 billion pre-money valuation. OpenAI’s announcement states the round was anchored by Amazon, NVIDIA, and SoftBank, with continued participation from Microsoft, and that SoftBank co-led alongside a16z, D. E. Shaw Ventures, MGX, TPG, and accounts advised by T. Rowe Price Associates. More than $3 billion came from individual investors through bank channels. The credit facility was expanded to $4.7 billion. OpenAI stated the three anchor amounts of $50 billion, $30 billion, and $30 billion at the February announcement.

OpenAI and Amazon, “OpenAI and Amazon announce strategic partnership,” February 27, 2026 (carried on openai.com and aboutamazon.com). Both companies state the $50 billion as $15 billion now and $35 billion “in the coming months when certain conditions are met.” Neither names the conditions. OpenAI commits to consume approximately 2 gigawatts of AWS Trainium capacity, and the two expand their existing $38 billion cloud agreement by $100 billion over eight years. AWS becomes the exclusive third-party distributor for OpenAI’s Frontier platform.

OpenAI, round announcement, February 27, 2026, on the NVIDIA commitment. OpenAI states it committed to “3GW of dedicated inference capacity and 2GW of training on Vera Rubin systems,” five gigawatts in total, alongside NVIDIA’s equity.

SoftBank Group, “Follow-on Investments in OpenAI,” February 27, 2026. $30 billion follow-on through SoftBank Vision Fund 2, subject to the satisfaction of customary closing conditions. Cumulative investment expected to total $64.6 billion, an ownership interest of about 13 percent. The consideration is to be financed initially through bridge loans and other arrangements from major financial institutions, then replaced over time through existing assets. Closing dates may accelerate in the event of a public listing.

Jensen Huang, Morgan Stanley Technology, Media & Telecom Conference, March 4, 2026 (remarks reported by CNBC). On the September 2025 $100 billion figure: “probably not in the cards,” because OpenAI is going public. The $30 billion “might be the last time” NVIDIA invests before an IPO.

OpenAI and SoftBank, “Announcing The Stargate Project,” January 21, 2025 (carried on openai.com and group.softbank). SoftBank and OpenAI named as Stargate’s lead partners, SoftBank with financial responsibility and OpenAI with operational responsibility. “Masayoshi Son will be the chairman.”

From the filing

Series C Preferred Stock filing, as read by GeekWire, March 2026. Amazon’s $15 billion is due March 31. The later tranche is stated at $34,999,999,447.98 and is callable on five business days’ notice once milestones are met. The milestones are redacted. The $35 billion equity commitment is contractually linked to the cloud agreement: if the Joint Collaboration Agreement terminates, the $35 billion obligation terminates with it.

As reported

Financial Times, February 2026 (originating report). NVIDIA’s $30 billion is a dedicated equity stake, not tied to deployment milestones, removing the earlier requirement that OpenAI direct spending toward NVIDIA hardware. It replaces the September 2025 letter of intent of up to $100 billion, which was conditioned on each gigawatt of NVIDIA systems deployed and never closed. NVIDIA declined to comment. Reuters reported that at the announcement it was not confirmed whether the $30 billion replaced the September commitment, and that OpenAI and NVIDIA did not clarify. Huang confirmed the replacement at the Morgan Stanley conference on March 4, 2026, resolving that question.

The condition on Amazon’s $35 billion (The Information; The Verge). With the milestones redacted in the filing, reporting has placed the trigger at an OpenAI IPO or an AGI milestone. Neither Amazon nor OpenAI has confirmed a specific trigger. This piece leaves the condition open.

Reuters, March 2026. SoftBank’s $30 billion paid in phases, the first roughly $10 billion around April 1.

Analysis (Cape Fear Advisors)

The roughly $12 billion broader pool is the round’s $122 billion less the three anchor commitments of $50 billion, $30 billion, and $30 billion. “At least $35 billion contingent” is Amazon’s conditional tranche. The total carries no cash figure because the split between cash and services across the strategic commitments is not disclosed.

The reading that NVIDIA’s equity is not deployment-conditioned is ours, drawn from what the parties did disclose: a direct stake at the round’s valuation, paired with a separate binding commitment to NVIDIA systems, and Huang’s March 4 statement that the $100 billion was off. The parties did not use that phrase.

Earlier rounds

2019: Microsoft’s initial $1 billion investment in OpenAI. January 2023: a tender offer was reported by the Wall Street Journal to value OpenAI at about $29 billion, the starting point for the valuation arc; the closing was later reported at $27 billion to $29 billion. OpenAI declined to comment. 2023: Microsoft described its investment only as “multiyear, multibillion dollar.” The amount and the implied valuation were reported, not disclosed. October 2024: $6.6 billion at a $157 billion post-money valuation, led by Thrive Capital. March 2025: up to $40 billion at a $300 billion post-money valuation, led by SoftBank, whose effective share was $30 billion after syndicating $10 billion to co-investors (SoftBank Group, “Announcement Regarding Follow-on Investments in OpenAI,” March 31, 2025).

Oracle reported a record

On June 10, 2026, Oracle reported results for the fiscal year just ended. One figure led them: remaining performance obligations, or RPO, the revenue Oracle holds under contract and will book in later periods. It reached $638 billion, a record, up 363 percent from a year earlier and up $85 billion in the quarter, from $553 billion.

That figure drew the attention. Oracle reported a second one.

Most of the increase, Oracle wrote, came from large AI contracts “where the customer prepaid Oracle for the purchase of the GPUs, or the customer bought and supplied the GPUs to Oracle.” It gave the size: “The prepaid and customer supplied hardware portions of our large AI contracts now total $75 billion.” And it added what that did for the company: “This substantially reduces the amount of capital Oracle must raise to build out our AI datacenters.”

So $75 billion of the record is, by Oracle’s own account, hardware a customer paid for in advance or supplied outright. Oracle names no customer behind it, and its 10-K, due within weeks, may name one.

Most of the record arrives years out. On its earnings call, Oracle put roughly 12 percent of the $638 billion converting to revenue within a year, another 34 percent over the two years after, and the rest, more than half, beyond three.

The demand comes from public contracts

The contracts behind the record are public. In September 2025, OpenAI and Oracle announced one: OpenAI would buy about $300 billion of computing from Oracle over roughly five years, starting in 2027. It is the largest contract publicly tied to that demand, a promise to pay rather than revenue earned. The contract is signed; the compute comes later.

To build the Abilene, Texas site that will run it, Oracle ordered roughly $40 billion of NVIDIA’s GPUs. And NVIDIA has offered a letter of intent to invest up to $100 billion in OpenAI, paid in as OpenAI deploys NVIDIA’s systems.

NVIDIA invests in OpenAI. OpenAI commits its spending to Oracle. Oracle commits to buy NVIDIA’s chips to deliver it. The same three names appear in each.

One name recurs across all of them. SoftBank holds about 13 percent of OpenAI, roughly $64.6 billion, chairs Stargate, the venture behind the buildout, holds 40 percent of its equity, and carries what the project calls its financial responsibility.

OpenAI is the company at the center

OpenAI loses money. Its own audited numbers, which we added up separately, show costs that run well past revenue and roughly a year of cash at the end of 2025. It funds what it commits by raising.

The lead on those raises has changed hands. Microsoft anchored OpenAI’s first rounds, in 2019 and 2023, and has led none since. Thrive Capital led in 2024, SoftBank in 2025, and Amazon, NVIDIA, and SoftBank led the $122 billion round that closed in March 2026.

Its valuation stands behind the commitments

That valuation is set in private. In March 2026, OpenAI’s last round put it at about $852 billion. The shares have no public market; the figure is the price investors agreed in a private round.

Three positions rest on it. OpenAI raises against that value to fund what it has committed. SoftBank, holding about 13 percent, carries the private mark on its own books. Oracle’s contracted revenue depends on OpenAI paying for the compute it ordered. The same number sits under each.

SoftBank’s exposure carries a date. To fund the most recent $30 billion of its OpenAI stake, paid across 2026, it took a $40 billion bridge loan in March, unsecured, due March 25, 2027. It has been covering the rest by selling: in late 2025 it sold its entire NVIDIA stake, $5.83 billion, and $9.2 billion of T-Mobile shares in the same quarter. SoftBank has said it will repay the bridge through asset sales and other financing, and the first source it has named is OpenAI’s public listing.

Another route ran through the stake itself: borrowing against the OpenAI shares rather than selling them. That loan stalled. SoftBank sought $10 billion backed by its OpenAI shares, cut the target to $6 billion in May, and on June 10 the talks broke off. The obstacle, as reported, was the collateral: OpenAI trades nowhere, so the shares carry no market price, and a lender could not value or sell them quickly if the loan soured. SoftBank’s stock fell 8.3 percent that day.

S&P had marked the same spot in March. On the 3rd it cut its outlook on SoftBank to negative, citing the OpenAI investment’s drag on liquidity and on the credit quality of its assets. It called OpenAI the weakest credit among them and projected that SoftBank’s unlisted assets would pass half the portfolio.

So SoftBank owes $40 billion against a fixed date, and when it moved to raise cash against the stake, the lenders declined to price it. The source it named first is a public listing that does not yet exist.

Its IPO is the answer everyone names

The resolution repeats in account after account: OpenAI goes public.

A listing would do two things at once. It would raise capital for a company that spends past its revenue and funds what it commits by raising. And it would set a public price on the shares its investors already hold, the first the market has marked. A great deal rests on each.

The capital pays the commitments. OpenAI funds its spending on Oracle, on the compute that runs on NVIDIA’s chips, on the buildout itself, out of what it raises, and the listing is the raise now in view. The price frees SoftBank’s $40 billion bridge: a listed stake can be sold or borrowed against, where an unlisted one, as its lenders just showed, cannot. OpenAI’s funding, the vendors’ payment, and SoftBank’s bridge converge on the one offering.

On June 8, 2026, that offering took its first formal step. OpenAI said it had filed a draft registration statement with the SEC, confidentially. “We expect it to leak,” the company wrote, “so we’re just announcing it.”

A confidential draft sets no price, offers no shares, and fixes no date. OpenAI has said the timing remains open, that a listing may be a while off, with things it would rather do as a private company first. Reporting places the earliest window in the fall. The underwriters are Goldman Sachs and Morgan Stanley.

The listing would carry the valuation already in play: OpenAI’s last private round set it at about $852 billion, the same mark the lenders declined to price weeks earlier, when SoftBank asked to borrow against it.

Masayoshi Son, whose bridge falls due in March 2027, has said OpenAI “will be listed eventually and, in my belief, will become the most valuable company in the world.” He framed it as belief and as eventuality, not as a date.

The answer every party leans on is real, filed, and unbuilt. One party’s documents show it leaning on something else.

What Microsoft did

The party is Microsoft, which owns more of OpenAI than anyone. Its stake of about 27 percent was valued at $135 billion in OpenAI’s October recapitalization.

On April 27, 2026, Microsoft and OpenAI restructured their partnership. Microsoft’s license to OpenAI’s technology became non-exclusive, running through 2032. The revenue Microsoft owed OpenAI ended; the revenue OpenAI owes Microsoft continues, now capped. OpenAI gained the freedom to run on any cloud. Its forward buildout has gone elsewhere, to Oracle and Amazon.

Microsoft made no separate filing for the change. The terms were announced, and its formal mention of OpenAI came two days later, folded into the quarterly results on April 29 as an item it backs out. Microsoft reports earnings “excluding the impact from investments in OpenAI.” Across the first nine months of its fiscal year, that impact ran to a net gain of $5.9 billion. Most of it came from the October recapitalization: as new investors entered and Microsoft’s share of OpenAI fell, the round’s higher valuation marked the smaller share as worth more. What three other balance sheets rest on, Microsoft’s accounts carried this year as a gain it sets to one side.

Microsoft kept the largest stake in OpenAI and the lightest forward commitment among its backers. The contracts, and the capacity still to be built against them, are Oracle’s now, not Microsoft’s. The same exposure crossed from one balance sheet to the other, and the filings record both ends.

So the company that owns the most of OpenAI is, by its own documents, not the one paying for what comes next. That leaves the question of who is.

A question of timing

NVIDIA’s investment in OpenAI, OpenAI’s compute commitments to Oracle and Amazon, SoftBank’s stake and the bridge that funded it: all of it rests on one company, valued in private at a number the lenders declined to price. Each of them is timed against one event: OpenAI’s listing.

The listing is confidential, undated, and by OpenAI’s own account may wait.

The obligations keep their own dates. SoftBank’s $40 billion comes due in March 2027. OpenAI spends past its revenue every quarter. The compute it has contracted is being built now, on debt and on customers’ prepayments, against revenue still to come.

So the question, beyond whether the listing will justify the valuation, is a simple one: when will it happen.

If it comes late, the cash has to come from somewhere. Microsoft’s own filings have already removed one candidate: itself. What remains is the others, the company at the center of them, and a date in March that does not move.

— — —

Analysis: Cape Fear Advisors. Sources: Oracle Corp. Form 8-K, June 10, 2026; the OpenAI partnership announcements with Oracle, NVIDIA, Amazon, and SoftBank, 2025 and 2026; Bloomberg, on SoftBank’s financings and the S&P outlook action; Microsoft Corp. third-quarter results, Form 8-K and Form 10-Q, both April 29, 2026; OpenAI’s confidential S-1 announcement, June 8, 2026. Figures verified against the primary filings.

The three weeks

May 20: Q1 FY27 earnings. Revenue $81.6 billion, free cash flow $48.6 billion, gross margin 75%. Segment reporting restructured into Data Center (Hyperscale and ACIE sub-segments) and Edge Computing. The new architecture organized the growth story around China’s absence, with sovereign AI up 80% year over year, AI clouds growing nearly three times faster than hyperscalers, and a freshly named $200 billion Vera CPU opportunity. Kress on the call: NVIDIA still seeing no revenue from China despite US authorization, with Beijing actively protecting domestic semiconductor manufacturing.

May 21: Huang joined the Tsinghua University School of Economics and Management advisory board, the 65-member channel through which US executives maintain dialogue with Chinese leadership. The appointment came eight days after the public concession that NVIDIA has “largely conceded” the China advanced AI chip market to Huawei.

May 22: Constellation groundbreaking in Taipei. Annual Taiwan commitment raised to $150 billion, roughly equal to NVIDIA’s annualized operating cash flow.

May 23: Huang publicly stated his $200 billion CPU market forecast “includes China.” The China position moved from “permanently gone” in the segment redesign back into the new growth narrative as forward opportunity.

June 1: GTC Taipei keynote. Vera in full production with 150 Taiwanese supply chain partners. Early customers explicitly named: Anthropic, OpenAI, SpaceXAI. Microsoft partnership announced with the N1X laptop SoC and the RTX Spark superchip, introducing a new consumer-device category positioned as the post-PC interaction paradigm.

NVDA closed today at approximately $224, roughly flat from the post-earnings close eleven trading days earlier.

The three pillars

The architecture moved from a concentrated dependence on China — historical addressable market roughly $50 billion, structurally lost through export restrictions and Beijing’s domestic protection policy — to a portfolio of three pillars with distinct risk profiles.

Sovereign AI and AI clouds. Growth driven through the customer-investment network NVIDIA built over the past eighteen months. The capital deployment ratio across the first five months of 2026 ran approximately ten dollars into ecosystem stakes for every dollar into operational capex. The Vera customers Huang named today — Anthropic, OpenAI, SpaceXAI — are three of NVIDIA’s equity-stake portfolio companies. Kress in prepared remarks at the Q1 call: “The vast and trusted marketplace for NVIDIA Compute is a critical foundation on which $ billions in AI infrastructure spending is being financed by the ecosystem.” Jensen, in response to a question on AI native clouds: NVIDIA architecture is “the easiest to finance.” Apparent demand sustained by continued capital deployment is the structural feature management itself names as competitive advantage.

The Microsoft consumer-device extension. N1X laptop SoC and RTX Spark superchip distributed through Microsoft’s established OEM channels. Microsoft is not in NVIDIA’s equity-stake portfolio. The PC OEMs are not in NVIDIA’s equity-stake portfolio. If the AI PC category materializes as the announcements project, the revenue is structurally independent of NVIDIA capital deployment to customers. Different risk profile than the first pillar: requires market acceptance of agentic AI as the post-mouse-and-keyboard paradigm, competes against Apple Silicon and Qualcomm Snapdragon, operates against OEM economics that don’t favor premium chips in price-sensitive consumer segments.

The China optionality. The May 23 statement returned China to the $200 billion Vera CPU TAM as forward opportunity. The Tsinghua board seat maintains the diplomatic channel. Current revenue zero; future optionality preserved without operational dependence on current performance.

Each pillar has structural problems distinct from the others. None alone carries the growth narrative. The portfolio is robust to single-pillar attack because no single pillar is load-bearing.

What integrates across them

The pillar diversification is the surface architecture. The integration across the pillars is one operator.

Each pillar requires Huang personally as the channel through which it operates. The China optionality requires the Tsinghua board seat that only he occupies — Tim Cook chairs the 65-member board; Musk, Nadella, Zuckerberg, Dell, Dimon, and Fink are among the other members. The diplomatic channel is the seat, and the seat is personal. The equity-stake portfolio company pillar requires the relationships with Anthropic’s, OpenAI’s, and xAI’s principals that Huang personally holds. The Microsoft partnership requires the executive relationship with Satya Nadella that produces partnership-scale announcements. The narrative-geographic shift to Taipei requires the credibility to make Taipei the announcement-architectural home of the AI revolution.

The pillar diversification distributes the growth narrative across categories. It does not distribute the operator. Every channel runs through one person, in parallel, at speed, without margin for faltering across any of them.

The diversification did not reduce the architecture’s risks. It transferred them. The first pillar carried market-structural risk: China was lost to geopolitical realities outside the company’s control. The new portfolio replaces external structural risk with internal operator risk. Three pillars with distinct risk profiles, each requiring continuous management, each requiring coherent coordination with the others, all running through the same person. The risks the diversification absorbs are not eliminated; they accumulate around the conductor.

Each new pillar adds operational workload and operational complexity. The portfolio coherence depends on continuous coordination across the channels — the China optionality cannot be maintained without the diplomatic channel; the equity-stake portfolio company pillar cannot operate without the relationships; the Microsoft partnership cannot scale without the executive credibility; the supply-chain commitment cannot deliver without the manufacturing relationships. The coordination is the architecture; the architecture only works with one person continuing to conduct it.

The framework reading named the dependency at the moment of the Q1 disclosure. Twelve days later, the dependency has deepened. The architecture is structurally accumulating dependency on Huang at an accelerating rate — not just for execution within each channel, but for coordination across an increasingly complex portfolio. The conventional disclosure language — standard “loss of key personnel” risk factor — is valid as far as it goes and structurally understates the dependency it describes.

What the segment-level instruments miss

The conventional analytical environment processes each disclosure separately. Q1 earnings beat. Segment restructure as a presentational change. Tsinghua appointment as a soft-power channel. Taiwan commitment as supply-chain investment. Microsoft partnership as new market category. May 23 China statement as forecast methodology clarification. Vera launch as growth-driver confirmation.

Each item satisfies its own analytical instrument. None of the instruments aggregates to the integrated picture. The flat stock price across three weeks of massive structural news is the empirical evidence of the integration gap. The market has digested each item without producing a decisive directional read because the items don’t decisively integrate at the segment-level instruments the market is using.

Three pillars, one operator, accumulating dependency at a rate the disclosure language cannot aggregate, with each new pillar adding both workload and coordination demands the trillion-dollar valuation already assumes.

Below the calm stock price, the architecture restructured. The twelve days were corners taken at speed. The configuration entering the long straight is what they produced. The question moves from cornering to flat-out horsepower and speed. Not all of the restructure has been incorporated by the conventional analytical environment. The integrated picture lives at a level the segment-level instruments do not aggregate.

This piece extends the framework reading laid out in The Twelve Elements of NVIDIA’s $4 Trillion Valuation, the Q1 FY27 subscriber note, and Huang Is the Real Trillion Dollar Man.

Last week answers it qualitatively. The Q1 FY27 earnings call disclosed more than any other in NVIDIA’s history. Customer concentration intensifying, supplier obligations of $182 billion, the China data-center exit, a non-marketable equity book doubled in a quarter to $42 billion. Real challenges remained to absorb, China most prominent. Huang managed the cycle and maintained the growth expectations the $4 trillion valuation requires. Since the call, he has made three more deliberate moves: taking a board seat in Beijing, announcing a $150 billion-per-year supply commitment in Taipei, and preparing to keynote in Taipei on Monday. Each is a busy executive doing the job at scale. Read together with the context of the call, the moves answer the question. The answer is “very.”

The week

Wednesday the Financial Times reported that Huang had joined the advisory board of Tsinghua University’s School of Economics and Management, the institution that educates China’s senior official class. The 65-member board is chaired by Tim Cook. Other members include Elon Musk, Satya Nadella, Mark Zuckerberg, Michael Dell, Jamie Dimon, and Larry Fink. The board does not lobby; its members participate in periodic dialogues with Chinese leadership. Membership is the soft-power channel for US corporations whose business requires China access despite a deteriorating bilateral relationship. The appointment came eight days after Huang told analysts on the Q1 FY27 earnings call that NVIDIA has “largely conceded” China’s advanced AI chip market to Huawei. The public concession releases political pressure in the US. The private appointment maintains the channel in Beijing. The two operate as complementary moves, in the same week, by the same person.

Earlier in the week in Taipei, Huang spoke at the groundbreaking of Constellation, NVIDIA’s new 4,000-worker campus, and announced annual Taiwan spending would rise from roughly $10–15 billion several years ago to $150 billion. He called Taiwan “the epicenter of the AI revolution.” Annual spending at that scale is roughly equal to NVIDIA’s current annualized operating cash flow. The commitment concentrates in a single jurisdiction whose security depends on US treaty and force-projection commitments. The cost of failing to maintain those commitments now includes the destruction of the supply chain NVIDIA just committed to. The geographic commitment functions as a request for US defense at scale, made structurally rather than explicitly.

On Monday June 1, in Taipei, Huang will co-keynote COMPUTEX with Marvell CEO Matt Murphy at the Taipei Music Center. NVIDIA holds GTC Taipei as a separate event from its US GTC. The venue and the date, immediately following the Constellation groundbreaking, make Taipei the announcement-architectural home of the AI revolution and not only its manufacturing home. The location of major announcements is itself a structural choice, and it has been shifted from the US to Taipei.

Three moves. Each makes a different argument about the same proposition.

The case the week is making

The proposition is that NVIDIA’s $4 trillion valuation depends on this specific human continuing to do these specific things, at this level of performance, simultaneously.

Huang is driving a car custom-built for him. No one else has the keys, and the road it travels was paved to accommodate this specific vehicle. The Tsinghua appointment is the diplomatic seat only he occupies. The Taiwan commitment is the supply-and-security architecture only he can announce with credibility at that scale. The Computex venue choice is the narrative-geographic anchor only he can move. The earnings call’s disclosure-cycle management is only legible because he is the one explaining the company to the market.

Standard CEO compensation arithmetic compares one executive against peer-group benchmarks. The arithmetic the work this week implies runs differently. The relevant comparison is what NVIDIA’s market capitalization is, what fraction of it depends on his continued personal participation, and what that fraction’s dollar value is. The market is currently valuing the assumption of his continued personal participation in the multiple it carries. A trillion is the floor of what that valuation implies. Any compensation discussion that follows will run against a market already valuing the assumption at this magnitude. The negotiation occurs inside that valuation, not against it.

What stops when the work stops

The architecture of the case is also the architecture of the risk. Every move that demonstrates indispensability also demonstrates dependency. The two readings are not in tension; they are the same observation from opposite directions.

The risk is not bounded by incapacitation or retirement. It is the ongoing requirement that Huang continue operating at this level, across this many channels, at this speed, without faltering. Twelve cylinders at 300 miles per hour have no margin for a single mis-fire. Institutional functions transfer to a successor: chip engineering continues, CUDA continues, TSMC operational relationships continue. The personal channels do not transfer in any reasonable timeframe, and not to any single successor, any outside hire of any pedigree, or any combination of people. The integration across the channels is itself the value. Decompose the role across multiple people and the integration is lost; replace him with one person and that person inherits the operational positions without the relationships, the founder credibility, the China-channel access, or the narrative authority the relationships require.

Tim Cook is the canonical case of CEO-as-strategic-asset. Apple’s continuity, supply-chain mastery, services-growth strategy, and operational discipline are inseparable from his leadership, and Apple’s 10-K reflects that reality through standard “loss of key personnel” disclosure. NVIDIA uses the same language for Huang. The language is valid in both cases, and in both cases it structurally understates the dependency it describes. The understatement at NVIDIA is materially greater because the dependency is. Cook leads a company whose core business is mature and whose succession-readiness has been a stated focus for years; Huang leads a company that is still being built around him, whose narrative engine is his personal voice, and whose most consequential international relationships exist because he personally holds them. The work this week further deepened the dependency on his continued personal participation, to a degree the standard disclosure language, valid as far as it goes, was never built to capture at the magnitude it now represents.

This observation is descriptive, not prescriptive. The architecture is the architecture. The work the architecture rests on is the work Huang has been doing, and every holder of NVIDIA exposure, direct or through index, has a rational interest in his continuing to do it. The framework reading does not contest that interest. It surfaces what the standard disclosure language cannot: the magnitude of dependence on his continued personal participation, accumulated to a level the standard language was never designed to disclose.

NVIDIA is approximately 7% of the S&P 500’s total market capitalization. A Huang-specific event (incapacitation, departure, or sustained performance below this level across any meaningful subset of the channels he runs) does not stay with NVDA shareholders. It propagates through the semiconductor sector, the AI infrastructure complex, and the hyperscaler capex linkage. A combined market response in the $1.5–$2.5 trillion range, 3 to 4 percent of the entire index, is the order of magnitude the dependency now operates at. This is not an NVDA shareholder risk or a board risk. It is a global market risk that every index fund, every 401(k), every passive holder of US large-cap exposure is carrying without having chosen it.

Twelve cylinders, all firing, none redundant, in a car custom-built for one driver, on a road the index has paved for him to drive on, at speeds the machine has no margin to forgive. The trillion is what the market is currently valuing as the assumption of this car being driven, by this driver, at this speed. The trillion is also what stops working when the assumption stops holding. Both numbers are the same number. The market has been valuing the same assumption, in real time. That is the architecture this week made visible.

This piece extends the framework reading laid out in The Twelve Elements of NVIDIA’s $4 Trillion Valuation and the Q1 FY27 subscriber note. The published SpaceX series runs the same instrument on a different case.

NVIDIA operates at scale that conventional analysis struggles to absorb. $4 trillion in enterprise value. Approximately $216 billion in fiscal 2026 revenue, up 65 percent from the prior year. Customer concentration at 61 percent from four customers in the most recent quarter. The scale itself constrains how the company can be managed. Failures at this scale produce corrections in the hundreds of billions rather than in the billions. Each managed element operates at magnitudes that change the nature of the management problem.

The foundational chip business produces real, durable value through six elements. Chip design through Blackwell and Rubin generations executes at standards the industry recognizes. The CUDA software ecosystem creates developer mindshare and switching costs. These produce durable competitive advantage. The TSMC manufacturing partnership operates at scale and reliability alternatives cannot match.

Customer support across the buildout delivers against commitments at hyperscaler scale. Competitive positioning defends against AMD, against custom silicon from hyperscalers, and against Chinese domestic alternatives. Operational financial discipline produces gross margins above 70 percent, operating margins above 60 percent, and free cash flow of approximately $97 billion in fiscal 2026.

The six elements together support substantial enterprise value. They do not alone support $4 trillion. The gap between what the foundation supports and what the price reflects gets filled by additional elements that have expanded over the past year.

The additional elements

China policy and trade relationships. Huang spent the past year as a go-between with Washington and Beijing on chip-export questions. The China revenue at stake runs approximately $20 billion against $4 trillion in enterprise value. The management attention runs at levels disproportionate to the revenue share because the larger question matters more: whether Chinese domestic chip development displaces NVIDIA’s foundational position if the relationship deteriorates. The Tuesday boarding signals how much that question matters at this moment.

Political channel cultivation with the Trump administration. Direct executive relationships. Acceptance of the federal-revenue-cut structural arrangement on China sales. Sovereign-AI customer development through administration channels. Administration officials raised national-security objections to advanced-chip sales in November 2025, marking the boundary the cultivation approaches.

Customer-investment network. NVIDIA committed $40 billion in equity stakes in customers and ecosystem partners in the first five months of 2026. $30 billion to OpenAI. Stakes in CoreWeave (7 percent), Anthropic Series G, xAI Series E, Nebius, IREN, Corning. Capital expenditure for fiscal 2026 ran approximately $6 billion. Ecosystem investment exceeded operational capex by roughly seven to one. Wedbush placed these investments “squarely in the circular investment theme.”

AI demand narrative. The hyperscaler capex trajectory at $805 billion projected for 2026 (Morgan Stanley) and $1.1 trillion for 2027 assumes sustained end-customer AI demand at the scale the buildout requires. Independent analytical work has identified scaling limits in current LLM technology that question whether validation arrives at the scale and timing the projections assume. The narrative sustains customer capex commitments through the validation window.

Accounting classification. Accounts receivable grew 67 percent in fiscal 2026 to $38.5 billion against revenue growth of 62 percent. Inventory grew 112 percent to $21.4 billion. Supply commitments grew 220 percent to $95.2 billion. Each growth rate exceeds revenue growth. The positioning prepares for accelerated demand the narrative supports.

The CEO as personal brand. Huang’s keynote appearances, media presence, and public relationships have become material components of the company’s market identity. The Air Force One boarding produces the kind of personal-brand maintenance Musk performs at Tesla and SpaceX. The brand component requires Huang’s sustained personal participation in high-visibility moments. The valuation now reflects the credibility Huang brings to the AI narrative and to the strategic decisions the company stewards.

These six additional elements operate alongside the foundational six. Twelve elements together produce the current price. Conventional commentary discusses several of them individually. The combined pattern has not yet received the same engagement.

What the elements address

The chip business will produce projected revenue if AI end-customer demand realizes at the scale the buildout assumes. Whether it will is the foundational question NVIDIA’s $4 trillion valuation embeds.

The additional elements operate as the conditions under which the foundational business produces the projected revenue during the validation window.

The customer-investment network sustains GPU demand from companies whose foundational economics develop while validation runs. The political channel maintains regulatory conditions the buildout requires. The China policy management defends against competitive threats that emerge if validation runs slower than projected. The narrative maintenance sustains customer capex commitments through the validation window. The accounting classification positions inventory and supply commitments against demand the narrative supports. The personal brand sustains the credibility the integrated system requires.

The integrated system captures full projected value if AI demand realizes within the window. The same system produces correlated correction across components if validation runs longer than the system can sustain.

The configuration also represents a shift in what the company is managing. The foundational six produce the chip business. The additional six produce the conditions under which the chip business operates. Management attention has expanded toward the second group as the valuation has expanded toward what only the second group sustains.

This produces specific exposure the current price reflects. Each element carries its own pressure. Each pressure interacts with the others. Failure at any single element removes a condition the others depend on. Adverse outcomes propagate across the integrated system rather than remaining contained.

The configuration requires twelve simultaneous successes over the validation window. Two or three independent elements at high individual probability would produce combined probability without strain. Twelve compound it down. The current price assumes near-certainty across the integrated configuration. The compounding probability operates lower than that assumption requires.

The compensation signal

NVIDIA’s proxy filing on Tuesday showed Huang’s total compensation for fiscal 2026 at $36.3 million. Stock awards came to $24.8 million, down 36 percent from the prior year as share-price appreciation moderated. Insider stock sales over the past three months ran approximately $164 million under pre-scheduled plans.

The arrangement remains conventional. Stock awards tied to operational performance. No stock-price-based hurdles. No trigger-on-trigger architecture that locks in valuation regardless of subsequent performance.

The conventional arrangement reflects foundational-management compensation. It also runs substantially below the levels companies whose CEO operates as material brand component have established. Tesla shareholders ratified approximately $1 trillion in equity compensation for Musk in November 2025 at a market cap roughly one-third NVIDIA’s. The current NVIDIA arrangement sits below founder-brand-asset precedents at comparable scale.

The configuration produces an unusual feature for the eventual compensation discussion. NVIDIA’s valuation depends on Huang’s continuation in the brand-asset role the configuration now requires. The board’s negotiating leverage on compensation faces structural pressure from what the valuation requires. Huang effectively operates on both sides of the eventual restructuring discussion — as the executive whose compensation is being set and as the brand-asset whose continued participation the valuation depends on. The eventual restructuring, when it arrives, will reveal which framing the board accepts.

The current arrangement does not yet address this. Whether and when it does becomes a watchable signal in its own right.

Signals to track

Specific conditions across the twelve elements produce signals over the coming quarters.

China-relationship outcomes from the Beijing summit. Approval or restriction. Commercial scale of any approved sales.

Political-channel conditions. Administration internal positioning. Regulatory environment shifts. Trade policy changes.

Customer-investment-network performance. OpenAI IPO pricing and trajectory. CoreWeave debt service. Anthropic commercial traction against the $30 billion annual revenue baseline. Individual hyperscaler capex announcements.

AI demand realization. End-customer revenue at scale from generative-AI applications. Hyperscaler capex moderation or sustained projection.

Accounting classification developments. Receivables collection patterns. Inventory write-down requirements. Supply commitment cancellations or modifications. Revenue recognition reclassifications.

Compensation arrangement structure. Restructuring toward founder-brand-asset arrangements or maintenance of the current arrangement.

Each signal supplies information independent of the others. The combined set produces a watchable configuration.

What the decomposition produces

NVIDIA at $4 trillion rests on twelve simultaneously managed elements. Six foundational elements produce the chip-business performance. Six additional elements address the question of whether AI demand realizes at the scale the buildout assumes. The integrated system captures projected value if all twelve sustain through the validation window.

The structural feature operates independent of management capability. Each element carries its own risk. Each risk interacts with the others. The configuration produces multiplied exposure across the integrated system rather than independent risks at each element. The aggregate operates at sovereign scale because NVIDIA does.

Conventional analytical commentary engages with several elements individually. The full decomposition adds the additional elements, the foundational uncertainty they address, the personal-brand component the valuation has come to reflect, the compensation signal the eventual restructuring will provide, and the compounding probability twelve simultaneous successes carries.

The framework that produces this decomposition reads across configurations conventional analysis applied to single dimensions cannot reach. The NVIDIA configuration represents one application. The structural reading produces the same kind of multi-element decomposition wherever it operates.

This raises a question conventional disclosure practice does not yet address. Standard materiality applies to single-dimension risks at conventional scale. The configurations the framework reads operate across twelve or more interacting elements at scales where individual elements carry material exposure and the integrated system carries multiplied exposure. Whether current disclosure standards capture what now matters in these configurations becomes a separate analytical question worth pursuing. The framework supplies the structural reading. The disclosure question follows from what the reading reveals.

The decomposition supplies the analytical infrastructure. The valuation judgment belongs to the reader.

A continuation of our independent analytic work. Prior pieces examined disclosure at sovereign scale, the Apple-Google AI substrate consolidation, the SpaceX TAM and forced-inclusion mechanism, the AI infrastructure context, the trigger-on-trigger architecture, the documented checks and balances, the IPO as M&A currency creation, the Terafab capital-gap diagnostic, and the $165 billion cumulative cash gap. A review of SpaceX’s forthcoming S-1 follows.