NVIDIA, What Comes Back

Two disclosures, shorter than usual because we published longer versions on Wednesday and Thursday.

Nothing here identifies an error, an inconsistency, or a bad actor. Every filing is taken on its face as accurate and complete under the rules that govern it. Advance rates move, spreads move, and a lender revising terms between one facility and the next is the ordinary operation of a credit market.

The required share, the decomposition and the three tests are ours. The tag marks who answers for a figure, not who holds title to it: FILED means the filer said it, REPORTED means a publication said it, and OURS means we said it and the error is ours to carry. Anyone with the same documents reaches the same arithmetic, and a reader who prefers different inputs substitutes them and runs it again. Every figure carries its source.

The machinery has to be in place before any of this applies

The size of the artificial intelligence build is estimated everywhere and documented almost nowhere. One corner of it is documented. Where a supplier funds a customer, the customer’s lenders write down what they will advance, and that term decides how much of the customer’s spending has to come back to the supplier before the supplier is whole again. The term is filed. It has been filed four times between July 2025 and August 2026, and the requirement it sets is half again what it was.

Circular financing is a phrase doing service for several arrangements, most of which are ordinary. The one this piece measures has a specific shape, and the shape is checkable.

It also comes in two currencies. A supplier can lend to a customer and a supplier can take stock in one, and both put capital into a company that spends part of it back with the supplier. What differs is the instrument left behind and who prices it. A loan leaves a term that a credit market marks continuously and in public: an advance rate, a spread, a rating, a maturity. A placement leaves a position marked by a tape, or by nothing at all until a funding round prices. This piece reads the credit form, because the credit form writes its terms down.

A company sells something. It puts money into a customer. The money stands where a lender requires equity, and the facility that opens buys the thing the company sells. The question is whether a company is funding its own demand, and that turns on whether the commercial machinery exists.

Which sorts the positions in NVIDIA’s reported equity book into three groups, each decidable from a filing. (1)

The machinery is in place. CoreWeave qualifies: a $2,000m placement in January for about 23 million Class A shares at $87.20, four filed credit agreements defining what lenders advance against the equipment, and equipment that is NVIDIA’s to sell. Nebius holds a commercial relationship and a warrant, on a lower grade of evidence.

The machinery is incomplete, and the position stands as an investment. The equipment is there and the leverage is not. Space Exploration Technologies sits here, at $20,976m and 33.1 percent of the book, the largest position in it after Intel. Its final prospectus names NVIDIA four times, and every one of them describes equipment rather than a counterparty. Twice in the list of market and industry sources, as the titles of two SemiAnalysis reports. Twice in the business section, describing its own fleet: approximately 325,000 NVIDIA graphics processing units across two sites, sold as capacity to Anthropic under Cloud Services Agreements at $1.25 billion a month through May 2029, ramping in May and June 2026 at a reduced fee, and terminable by either party on ninety days’ notice after an initial three-month period.

So the equipment is filed and the purchases behind it can be inferred. What no filing read for this piece supplies is the third element: a credit agreement stating what a lender will advance against that equipment. Without it there is no advance rate, and without an advance rate the requirement has no first input. NVIDIA appears nowhere among principal stockholders, in related-party transactions, or in any disclosed purchase commitment. The position may be an excellent investment. It is not this one.

The relationship runs the other way. Intel at $29,989m is a manufacturing relationship, the largest single position in the book, and money enters this system from outside it. Coherent and Synopsys supply NVIDIA on what is publicly understood of their businesses, which is a lower grade of evidence than a filing, so they are named and left uncounted.

The book shows one instrument form, and that matters more than the sorting.

At June 30 it carries Nebius at 1,190,476 shares. A Schedule 13G filed three weeks later reports 9.3 percent beneficial ownership, including a pre-funded warrant over 21,065,936 Class A shares struck at $0.0001. A warrant at a hundredth of a cent is the stock in every economic sense, and the position is eighteen times what the report shows.

Warrants, purchase commitments, capacity backstops and residual value guaranties all carry equity-like exposure. None of them reaches the position report. (2)

What a placement has to capture to get its money back

In the one relationship where the machinery is filed end to end, the narrow question comes first: what has to happen for the money to come back.

A contribution of E stands beside a program of E divided by one minus the advance rate a. Some share s of that program is spent with NVIDIA, at gross margin m, so the gross profit is E over one minus a, times s, times m. Set that equal to E and solve for the share:

The share of the program that has to be spent with NVIDIA, for a placement to return its own cash, is one less the advance rate, divided by the gross margin.

In plain terms. NVIDIA placed $2,000m. The August facility advances seventy cents against each dollar of equipment cost, so that placement stands beside a program of about $6,667m. If forty cents of every program dollar comes back as an NVIDIA sale, the gross profit on those sales is about what was placed. Everything above forty cents is return, and the stock sits on top of it.

The unknown is gone. The share was the one term no filer discloses, and asking what the share has to be removes it. Both remaining inputs are filed: the advance rate, in each credit agreement, and the gross margin, in the income statement. No other input enters it.

And there is no time in it. The gross profit arrives as the equipment ships, against cash that left once, so the question is one of size and no discount rate, horizon or cost of capital enters it.

What the line decides is whether an arrangement of this kind pays for itself on its own commercial terms. Above the requirement the sale has repaid the cash and the shares are a residual, so the arrangement clears whatever they do. Below it the sale has repaid part, and the shares carry the balance. That is the whole accretive-and-dilutive question for a circular arrangement, and the threshold that decides it is set by two filed inputs. (3)

Four rates, and the requirement rises by half

Each agreement defines what its lenders advance at the funding date. Set beside what each release says about the customers behind it, and beside what each rate then requires: (4)

Under the July 2025 facility, a placement returned its cash if about twenty-seven cents of every program dollar came to NVIDIA. Under the August 2026 facility it takes about forty. The requirement rises by half, and the margin cancels out of that ratio, so the rise is a function of the filed advance rates alone.

The rise is one step and then a plateau. Between July 2025 and May 2026 the requirement rises 42.9 percent. Between May and August it rises a further 5.0 percent. About 86 percent of the rise from July 2025 to August 2026 was in place by May, so the August facility continues where May sat instead of stepping up from it.

March runs the other way, and its own release explains it. DDTL 4.0 “received ratings of A3 by Moody’s and A (low) by DBRS, representing the first investment-grade rated financing secured by HPC infrastructure and an associated customer contract.” (5)

It reached investment grade at the highest advance rate of the four, and a rating ordinarily falls as loan to value rises. A3 at 90.00 percent inverts that, which holds where a rating rests on the counterparty contract in place of collateral cushion. So March is where a customer’s credit substituted for the cushion, and it makes a poor base for measuring a trend in the cushion. The comparison that survives is the pair at each end: DDTL 3.0 and DDTL 5.5, both $2.6bn, neither stating a customer rating, July 2025 against August 2026.

One limit on all of it. These are four facilities and at least three underwriting models, so the rate series and the model series move together. DDTL 3.0 lends against an asset. DDTL 4.0 lends against execution, at more than three times the size of the facilities at either end, and it is the only one described as non-recourse. DDTL 5.0 and DDTL 5.5 state neither model. The advance rate cannot be read apart from what each lender was underwriting, and the agreements do not say what that was for the two most recent. (6)

Three things moved, and the filings name all three

The counterparty, on one step. DDTL 5.0’s release states that proceeds “will support the purchase and deployment of infrastructure dedicated to customer contracts with two large, non-investment grade customers,” against March’s A3 and A (low). Between those two the rate falls about eighteen points and the stated credit falls with it.

The collateral mechanics, across the run. DDTL 3.0 capped outstanding loans each month at 85 percent of the equipment’s depreciated value, so the borrowing shrank as the fleet aged. That continuing cap is absent from the three later agreements, “Test Date GPU Amount” carries zero uses in all three, and “Stabilized” appears thirty-five times in DDTL 4.0 and not once after. (7)

And renewal risk on the equipment, which is the filer’s own account of the newest facility:

“Unlike prior delayed draw term loan facilities backed by customer contracts extending through the maturity of the debt, the DDTL 5.5 Facility carries an approximate five-year maturity while its underlying customer contracts average approximately three years in length. By financing these shorter-date commitments, lenders are signalling confidence in the long-term value of NVIDIA GPUs running on CoreWeave’s cloud platform and a willingness to underwrite renewal risk.”

The facility outlives its contracts by about two years, and on the last step the company puts its lenders on the machines. Three explanations, each sourced to a filing, and the requirement runs the same under all three, because it takes the rate as given.

Forty cents is where the requirement starts

The requirement converts program dollars into returned cash at gross margin, which is the widest conversion available and therefore produces the smallest requirement. Every narrower conversion raises it.

Gross profit stands above operating cost, so converting at an operating margin raises the requirement, and gross profit is taxed as it is earned, so converting after tax raises it again.

The two ways money comes back are taxed differently, and the difference carries weight. Gross profit on a sale is operating income, taxed in the period it is earned, and the cash leaves for the tax authority in that period. An increase in the carrying value of the stock is a mark: it prints in earnings, it carries deferred tax rather than a payment, and it can reverse to nothing without a dollar ever having moved. So a pre-tax dollar of sale and a pre-tax dollar of mark are not the same dollar. On a cash-return test the sale dollar is worth more. On a reported-earnings test in the quarter it prints, the mark counts the same as the sale.

The requirement is a cash test, which makes forty cents its floor. The upper end of the range lands against the fleet-wide equipment share. (8)

What no filing supplies is the answer

Forty cents of every program dollar is a requirement. Whether it is met is a different question, and no filer discloses NVIDIA’s share of anything CoreWeave buys.

The filings supply one figure in the neighborhood, and it works as a scale reference. Technology equipment was $33,823m of CoreWeave’s $52,622m of gross property and equipment at June 30, or 64.3 percent, against 61.6 percent at December 31, and additions in the June half ran 69.2 percent. That describes the fleet, and the facilities fund GPU servers and related infrastructure specifically, so a program is more equipment-heavy than the fleet it joins. NVIDIA also sells networking. So the fleet share gives a sense of scale, and nothing in the agreements read here caps what one vendor may capture. (9)

So the requirement stands on its own, next to a blank. Forty cents in the dollar is what has to be true. The one number that would settle whether it is true is disclosed by nobody, and it would take one line in a segment note to settle it.

What it looks like if the requirement is met

What the arrangement returns depends on the share captured, which nobody files, so the example is run at three shares.

At the requirement it is a dollar for a dollar and the stock is the whole of the return. Below it the sale returns less than it consumed, and the shares have to close the gap or the arrangement does not clear. (10)

On the after-tax conversion only the third clears, which is the range from the section above arriving in the example.

One assumption belongs to the example alone. The identity prices a marginal dollar: money placed where a lender requires equity stands beside a program of its own size over one less the advance rate, whatever else is in the capital structure. The example goes further and stands a single placement beside a whole program, as though it were the only equity behind it. CoreWeave raised equity and equity-linked capital from several sources across the same period, so a placement supplying part of the equity behind a facility stands beside a proportionally smaller program, and the recovery per dollar falls with it. The example takes the most favorable reading available, which is a second sense in which forty cents is a floor. The cross-facility ratio is untouched by any of it, because the same assumption stands on both sides and divides out. (11)

One qualification travels with the whole table, and it is separate from the conversion. The dollar placed carries whatever it costs the participant, so a company funding a placement from cash needs less margin above the requirement than one funding it from borrowings. On identical terms the same arrangement can clear for one participant and fail for another. (12)

One percent of the earnings, and a much larger share of the story

At the middle of those three shares, the placement opens roughly $2,288m of gross profit before tax. Against what NVIDIA already earns, gross profit of $153,463m in fiscal 2026, and $61,157m in the April quarter which annualizes to $244,628m, that is 0.94 to 1.49 percent of the earnings base. (13)

Which is the wrong denominator for the question the market is asking.

NVIDIA introduced a sub-segment in the April quarter called AI Clouds, Industrial and Enterprise, splitting data center demand between the four named hyperscalers and everyone else. The stated purpose is to show demand broadening beyond a handful of buyers. Data-center compute ran $60.4bn in that quarter, up 77 percent. (14)

Self-funded demand is a rounding error in the earnings base, and it sits inside the cohort whose growth carries the claim that demand is broadening beyond four customers. The first half of that is computed. The second half waits on a denominator: the $60.4bn is data-center compute overall rather than the sub-segment, and the sub-segment’s own size is not disclosed.

So whether self-funded demand matters to that cohort’s growth rate is the pre-registered question rather than a finding here. Wednesday either supplies the denominator or does not, and either outcome is informative. What can be said now is that the level and the margin are different questions, and that the line deciding the second one is written in a credit agreement, where it moved three times this year.

One borrower files the terms, and the rest of the market does not

The apparatus producing an advance rate against this equipment is a set of defined terms inside the credit agreements, and a full-text search of EDGAR run on August 22 returns them from one filer. “Funding Date GPU Amount,” ten filings, all CoreWeave. “GPU Depreciated Amount,” ten filings, all CoreWeave. “Test Date GPU Amount,” one filing, CoreWeave. (15)

One looser phrase is unresolved. “Loan to cost” alongside GPU returned eleven filings across four filers on August 21 and one filing on the August 22 rerun, and the two results have not been reconciled, so neither is used here. The phrase also appears in CoreWeave’s March presentation, which quoted 90 percent loan to cost rising to about 102 percent after stabilization, and which is a company document rather than a filing. (16)

The four rates above rest on those defined terms, and the terms come from one borrower’s filed schedules.

What the book can carry, and the direction it is traveling

The CoreWeave position was $4,700m at June 30. Total assets were $259,474m at April 26, so the position is 1.81 percent of them, and the whole reported equity book is 24.45 percent. The two dates are just over nine weeks apart and the ratios are stated on that unmatched basis. (17)

The book carried without an observable price grew 3.7 times as fast as the balance sheet holding it, and it is nine times the size of the CoreWeave position. Two observations are a line. The third arrives Wednesday. (18)

And the positions are the smaller half of what is committed. On August 17 NVIDIA entered residual value guaranties over leases for approximately 4.25 gigawatts of IT load at a site in Pike County, Ohio. The aggregate payment obligation is “cumulatively capped at $105 billion.” It is filed under Item 2.03, as a direct financial obligation or an obligation under an off-balance-sheet arrangement. That is $24,706m per gigawatt of guaranteed minimum value, on a twenty-year tail from each lease commencement, conditioned on ready-for-service dates expected from 2028, with the tenant reimbursing whatever is paid. Credit support for a further 3.8 gigawatts sits at the company’s sole discretion. Alongside it, a program announced August 10 states residual value support of up to 25 percent of an opportunity across platforms designed to mobilize over $500bn, and a reported $6,300m capacity backstop stands behind the CoreWeave relationship. (19)

Whether those overlap is stated nowhere, so they are listed one by one. Taken singly, the largest is $105,000m against $259,474m of total assets, which is 40.5 percent. Against that, an equity position at 1.81 percent is the part of the arrangement that can be absorbed without anyone noticing.

One relationship, and a large one in absolute terms. The placement was $2,000m at cost and the position stood at $4,700m at June 30. It sits beside a program of roughly $6,667m at the August advance rate, with a reported $6,300m capacity backstop behind the relationship. Against $259,474m of total assets it is under two percent, and it is less than a twentieth of the $105,000m written in Ohio. It is also a $4,700m position standing beside a $6,667m program, between two named parties, on filed terms. The system is measured in trillions and it moves in capillaries, and this is one capillary with both ends visible. (20)

The Ohio guaranties also terminate on the earliest of several events, one of which is “OpenAI achieving a satisfactory credit rating.” The support is written to end when the rating arrives. That is the March substitution, seen from the other side, in a second filer’s operative text.

The last two times, and what moved first

This has run before, and the analogues are here to be used. Both episodes were readable in filings while they were still running, on measures that were in the documents at the time, and the disclosure each one produced is part of why the measures used here exist at all.

Lucent committed $7bn to $8.1bn of customer financing and drew $1.6bn, then provisioned $3.5bn of loan losses across 2001 and 2002. Nortel committed $3.1bn and drew $1.4bn. Cisco committed $2.4bn and drew about $600m, and its investment book moved from $190m of net gains in fiscal 2001 to $858m of net losses the year after. The measure analysts used at the time was commitments against drawn, and it is filed every quarter. (21)

Seven years later the same variable ran the second episode. A repo haircut is one less an advance rate, and across nine collateral classes it stood at 0.0 percent in the first half of 2007, reaching 17.1 on AA-AAA mortgage paper and 53.5 on AA-AAA collateralized debt obligations during 2008. (22)

In both episodes the advance rate moved late. What moved first in 2007 was the seller’s own retained equity. Countrywide carried $3,040.6m of retained interests in its own securitizations, roughly a fifth of shareholders’ equity. It impaired $697.7m of them in the first half of 2007, against $69.2m a year earlier, before any haircut moved. Second was the gap between committed and drawn, widening as third-party capital left. Third was the advance rate. (23)

A retained interest is a seller’s continuing exposure to a sale it has already made, carried at a mark rather than at a price anyone paid. That is the class of instrument that moved first, and an equity position taken alongside a sale sits in the same class. (24)

So the rate is a price and an amplifier. It sizes how violent an unwind becomes once something else moves, and the something else has an address: the equity carried at a mark, the commitments not yet drawn, and the performance of what has already been sold.

Which side of the arrangement the amplifier sits on follows from where the leverage is. A supplier whose cash came back through a sale holds a mark and a completed transaction. A borrower whose equipment secures the facilities holds the rate itself, on every facility, at every test date. The requirement measures what the supplier needs. The advance rate is the borrower’s term, and it moved ten points between July 2025 and August 2026.

The same arithmetic runs on the other instruments

The requirement above is one arrangement solved, and the method under it runs on any of them.

Every one of these separates into four pieces, each filed in a different place. Cash that left, which is filed. The commercial claim the cash opened, rarely disclosed and derivable from a leverage rate and a margin. The equity residual, filed in position reports, and only for one instrument form. The contingent obligation written along the way, filed when it becomes an obligation and sized late.

An undisclosed input then takes one of three treatments. It cancels, where it stands unchanged on both sides of a comparison. It bounds, where the sign is known and the size is not, so the answer becomes an inequality in a stated direction. It decomposes, where the parts separate. An item of unknown sign reaches none of the three, and the $6,300m capacity backstop sits there, carried at zero because that is the treatment its disclosure supports. (25)

One structural difference is named here and worked separately. A placement is cash out, and it is cash out because the lender requires equity in that position for the facility to open. A residual value guaranty leaves the cash in place. It improves a facility where a placement releases one, it costs a contingent obligation where a placement costs a dollar, and it opens the same revenue gate. In Pike County the equipment sells and the gross profit arrives with no placement standing underneath it. The requirement above divides by a dollar that left. Where no dollar leaves, the price of the same revenue is written somewhere else, and it is written as a contingency. (26)

A fourth move is the one this piece uses. Where an input is undisclosed, solve for it. Ask what it would have to be, and the answer is made of filed numbers even when the input is not.

And it runs in reverse. Read backwards, the same line splits a carried position in two. It says how much of the dollar that left the gross profit has already returned, and what remains of the carrying value stands on the mark alone. The placement the cash bought is carried as an asset, and the gross profit it opened is recorded separately, so the accounts show both. (27)

The comparison that travels sets the commercial claim against the cash that left. Where the claim repays the cash, the equity is residual and the arrangement clears whatever the shares do. Where it does not, the arrangement requires the shares to appreciate, and the mark stops being a byproduct and becomes the return.

Run across the instrument forms, that decomposition is the next piece. The credit terms are where this one found its numbers. The residual is where the valuation consequence lands, and the residual is an equity question.

Three tests, and a reprint schedule

Whether the commercial machinery is in place, which decides whether any of this applies.

What a placement has to capture to return its cash, and which way that requirement is moving.

Whether the funder carries the position to zero without consequence, and whether that capacity is rising.

All three describe a position and leave the price of every security alone. A reader who prefers different inputs substitutes them and runs the arithmetic again, which is the argument for publishing the method. We searched for a published version of this construction and found none. Others may exist. (28)

A phase change in a market of this kind shows up as evidence being demanded, and then priced. Each of these is a place where that demand is written down, dated, and reprinted on a schedule somebody else sets, which is what makes it readable from filings.

NVIDIA reports the quarter ended in late July on Wednesday. Three figures here reprint that day and one of them moves the requirement directly: the gross margin, which is the only assumption-free input the requirement has besides the advance rate. The non-marketable book, the total assets it sits inside and the other income line carrying the marks reprint with it. The form of the Ohio guaranties files as an exhibit to the same report. And the segment granularity either lets an outsider separate funded demand from organic demand inside that 77 percent, or does not.

The other clock belongs to CoreWeave. DDTL 5.0’s draw window closes September 30 and DDTL 5.5’s on December 31, and the third-quarter report in November is the first observation of what this borrower does at a deadline with the contribution unfunded.

What it does to a growth rate

What artificial intelligence sells today is estimated. The aggregates in circulation are assembled from leaks, private disclosures and inference, and every forecast that supports a valuation compounds from a base of that kind.

The requirement measured here says nothing about the size of that base. It says something about its composition.

Where a supplier funds a customer and the customer spends part of the proceeds with the supplier, some part of what is counted as revenue in this system was supplied from inside it. No filer discloses that part. The size is unavailable and the sign is available: to whatever extent revenue is funded by its own vendor, the base a forecast compounds from is larger than the demand underneath it, and the growth required to reach any given valuation is higher rather than lower. An inequality in a stated direction is available where a level is unavailable. (29)

Which is why a segment note decides more here than an earnings line does. The earnings line settles what was sold. The segment note settles who paid for it, and whether an outsider can tell the two apart.

Four credit agreements produced four requirements, and the arithmetic that turns one into the other is two lines long. It runs again when the fifth one files.