What looks like an equity stake is the push that starts a wheel, and the wheel returns purchases many times the size of the push. On the lowest published figures a placement returns nearly ten times its own size at NVIDIA’s filed margin, and that is the low end of the range rather than the middle of it. The figure that governs is the placement at which the return falls to one, and on the deal in front of everyone that means the counterparty would have to fail before buying a tenth of what it committed to.
The subject ran everywhere last week. Two examples, chosen from many. The Economist published a briefing on September 3 headed “Nvidia is the central bank of AI,” which set the arrangement against the telecom equipment lending of the late 1990s and asked whether the loans will prove sound. (1) The Atlas newsletter published an item on September 4 about NVIDIA investing in the companies that buy its chips, closing on a sentence: the concern is circularity. (2)
Both are accurate, and both name the circle. The naming is where each of them stops, and it is where nearly all of the coverage stops.
Sizing it takes one operation and four numbers. Three of them circulate with every announcement, and the fourth is filed every quarter.
Circular financing, in motion
Capital goes into motion around a wheel. Money leaves NVIDIA as an investment, arrives at a counterparty, and comes back as a purchase order for the thing NVIDIA sells.
The stake is the smallest number in the arrangement, and the purchases it sets in motion are the largest.
The wheel is the fact everyone has. The measure is what happens to a dollar as it travels: how much of it returns, and in what form.
The deal in front of everyone
NVIDIA is reported to be in talks to place about $2.5bn into Thinking Machines Lab, at a valuation near $40bn. The company is reported to have committed to at least one gigawatt of NVIDIA’s Vera Rubin infrastructure. (2)
The placement is described as under negotiation and may close on other terms. What the reported pair supplies is the shape: capital out, and a quantity of committed purchases in.
The operation
Exhibit 1
Revenue per GW × Gross margin ÷ Capital placed = Coverage
$32,500m × 74.9% ÷ $2,500m = 9.7×
Revenue per gigawatt, times gross margin, over the capital placed. That is coverage, and it asks what a supplier earns on the purchases an arrangement produces, set against what the supplier put in to produce them. (3)
One of the four inputs is filed. NVIDIA’s second-quarter Form 10-Q reports revenue of $96,221m against cost of revenue of $24,079m, a gross margin of 74.9 percent, and operating income of $63,734m, an operating margin of 66.2 percent. (4)
The revenue a gigawatt produces for NVIDIA is published in commentary and estimates, and two figures bracket it. Jensen Huang has put NVIDIA’s own take at $40bn to $50bn a gigawatt. Barclays, working bottom up, put compute spending at $32.5bn to $42bn, with a midpoint near $39bn representing about 65 percent of a $60bn total. The two are not the same quantity. Huang’s figure is what NVIDIA books. Barclays measures what the buyer spends on compute, which sits above NVIDIA’s revenue by whatever the channel keeps. (5)
The arithmetic
Exhibit 2
| Per-GW figure | Gross profit | Coverage |
|---|---|---|
| $32,500m (Barclays low) | $24,342m | 9.7× |
| $39,000m (Barclays mid) | $29,211m | 11.6× |
| $40,000m (Huang low) | $29,960m | 11.9× |
| $50,000m (Huang high) | $37,450m | 14.9× |
On the lowest published figure and the filed margin, one gigawatt of purchases returns $24,342m of gross profit against $2,500m placed. That is 9.7 times. At the Barclays midpoint it is 11.6 times, and at the top of Huang’s range 14.9 times.
Substituting operating margin for gross margin, which charges the placement with research and selling costs it did not incur alone, the low case still returns 8.6 times.
The channel takes the low case down and leaves it in the same place. If distributors and system builders keep a tenth of the compute bill, coverage is 8.7 times. At a fifth it is 7.7 times, and at three tenths it is 6.8 times.
The placement carries no expense
Ten times is not the figure that decides anything. The one that does is the placement at which coverage falls to one.
Coverage of one has a specific meaning on this rung, and a lender’s meaning is the wrong one. A placement is an investing outflow, and it stops short of operating income. The stake sits on the balance sheet, the purchases it produces run through revenue and cost of revenue, and no part of the placement offsets the gross profit it generates. What appears below the operating line later is an impairment if the stake falls or a gain if it rises. (6)
So NVIDIA holds both legs at once: the gross profit, earned with no expense set against the capital that produced it, and the stake, still carried as an asset.
The size of the stake and the class of the shares are chosen. Above the influence threshold the equity method applies and an investor books its share of the counterparty’s results, which for a company still spending to build lands as a charge against NVIDIA’s own income. A small position, or a non-voting one, keeps the demand on the income statement and the counterparty’s losses off it. (6)
Where coverage falls to one
That fixes what coverage of one means. At the lowest published figure and the filed margin, one gigawatt of resulting purchases produces $24,342m of gross profit. At that placement a total loss on the stake would leave the arrangement whole, the gross profit having returned the capital while the stake was written off beneath it.
The reported placement here is $2,500m, which is 10.3 percent of that line. So the placement is covered once a little over a tenth of the committed gigawatt has shipped, and for the arrangement to lose money the counterparty has to fail before buying a tenth of what it committed to.
Coverage of one is a condition on timing as well as on size, and stating it that way is what makes it checkable. At 9.7 times on the full gigawatt, a total loss on the stake still leaves the purchases 8.7 times the placement ahead.
The test runs on recognition. A shipment becomes revenue and gross profit when it ships, and the cash arrives on the terms NVIDIA extends, which ran to an average near 60 days in the second quarter against a receivable of $63,059m. A tenth of the gigawatt shipped is a tenth recognized, and the collection follows behind it. (4)
Scaled up, NVIDIA is reported to have pledged over $70bn into startups across three years. (1) At the same low case, that entire book is covered by 2.9 gigawatts of resulting purchases. The campus at Piketon carries residual value guaranties on 4.25 gigawatts, with credit support available at NVIDIA’s discretion on about 3.8 more, and it stands here as a measure of scale rather than as coverage, because a guaranty places no capital. (7)
What settles it, and when
Three inputs here come from publication instead of from filings, so the reading runs bracketed. Each bracket narrows on a document with a date on it.
The placement. A closed investment appears in NVIDIA’s own equity and commitments disclosure, and the counterparty’s purchase obligation appears in its filings once it has any. At that point the numerator and the denominator both come from documents.
The channel, which thins the revenue and the margin together. NVIDIA’s filings put cloud service providers in an indirect class reached through board makers, distributors, original design manufacturers, original equipment manufacturers and system integrators. (8) Hardware that travels through a channel earns the margin on the sale into the channel, and the price the buyer pays sits above what NVIDIA books. A segment split, or a filer disclosing what it paid, converts a buyer-side estimate into a supplier-side figure and the company-wide 74.9 percent into the rate that applies here.
The financing behind the purchase. Each of these arrangements is funded, and the funding carries an advance rate. That rate says what a lender will lend against the equipment, which is the market’s own estimate of the thing NVIDIA stands behind. It becomes visible with each facility, and it is the one input the required share needs. (3)
The one term that stays an assumption is the counterfactual: some of these purchases would follow without the placement. Where that share is unknown, coverage is an upper bound on what the placement bought, and the bound is stated rather than modeled.
The same three components
A frame this shelf published in August sorts nine forms of the same help into a ladder: cash on the barrel, trade credit, supplier finance, customer prepayment, component intermediation, vendor lending, equity in the counterparty, contingent support, and wrapped structures. (3)
An equity placement sits on the seventh rung and is the simplest of them, which is why it is the one worked here. Every arrangement on the ladder decomposes into the same three components: the capital placed, the leverage arranged against the asset, and the gross profit realized on the purchases that follow.
A residual value guaranty places no capital and arranges a great deal of leverage. A prepayment sends the capital in the other direction. A supplier finance program puts a bank between the two. The forms differ and the three components hold, which is what makes one operation run across all of them.
Where that leaves it
Circular financing exists, it is disclosed, and the coverage last week described it accurately.
The question is how much capital an arrangement can carry before the purchases stop repaying it, and that number is computable from figures already public. On the deal reported last week it sits around $24bn against a placement of $2.5bn.
What moves it is a smaller figure per gigawatt, a larger channel cut, or a placement that grows faster than the purchases it produces. The first two settle as filings arrive. The third is visible on the face of every announcement.
Standing Disclosure
Anthropic is the developer of Claude, which is used in preparing this research. NVIDIA, whose placement economics this piece computes, holds a position in Anthropic taken in late 2025, and NVIDIA supplies the infrastructure that serves Anthropic. Among the other companies named, Thinking Machines Lab and OpenAI are competitors of Anthropic, and SB Energy is developing capacity that hosts compute for one of those competitors. The finding here runs in NVIDIA’s favor, and that is the direction in which the nearness needs saying. That nearness cannot be fully checked away, which is why no claim here rests on trust in the tool: every figure carries a public source, every construction is tagged as ours, and the record grades the rest. Companies not named here may hold positions or supply relationships that bear on the filers discussed, and that possibility is part of why every piece is re-checked for bias, ground facts, and filings instead of being read against a fixed list.
Notes
(1) The Economist, “Nvidia is the central bank of AI,” interactive briefing, September 3, 2026. The briefing’s own sources line reads press reports, the Financial Times, and The Economist. The $70bn pledged into startups over three years and the roughly $300bn of potential customer liabilities are the briefing’s figures and are REPORTED.
(2) GenAI Works, Atlas newsletter, “Nvidia builds an AI empire,” September 4, 2026, for the Thinking Machines Lab placement, the $40bn valuation and the one gigawatt commitment. The newsletter attributes the placement to The Information. The Information’s report was not reached, so the newsletter is the route rather than the source, and the whole worked example rests on that chain. REPORTED, and the placement is described as under negotiation.
(3) Coverage belongs to the same family as the required share set out in “The Required Share,” and the two are different computations. The required share asks what portion of a program’s spending has to reach the supplier for a placement to return itself, and it runs one less the advance rate over the margin, taking the advance rate from the customer’s credit agreement and the margin from the supplier’s income statement. Coverage takes an observed placement and an observed purchase commitment and returns the multiple. The advance rate is the input the required share needs and this arrangement has not disclosed, which is why the coverage form is the one that runs here. The Required Share carries no company in it. The fourteenth position runs the operator as a taxonomy rather than as a multiple, in the sentence stating that coverage names the forms it takes, being a margin, an occupancy, a capacity, an exclusivity or a guarantee avoided. The word appears in that piece in three senses, the operator, press coverage, and the debt service coverage tests the filing describes, and the first of those is the one meant here. The operator run as a multiple, on the figures here, is ours.
(4) NVIDIA Corporation, Form 10-Q for the quarter ended July 26, 2026, accession 0001045810-26-000075, filed August 26, 2026. Three months ended July 26, 2026: revenue $96,221m, cost of revenue $24,079m, gross profit $72,142m, total operating expenses $8,408m, operating income $63,734m. FILED. The gross margin of 74.9 percent and the operating margin of 66.2 percent are ours from those figures, floored rather than rounded. The CFO Commentary furnished as Exhibit 99.2 to Form 8-K, accession 0001045810-26-000073, states gross margin of 75.0 percent for the same quarter and corroborates the figures; a furnished exhibit is not a filed one, and the arithmetic here runs on the 10-Q. The same 10-Q reports accounts receivable of $63,059m at July 26, 2026 against $38,466m at January 25, 2026, FILED; the average collection period near 60 days, and the extended terms for certain investment grade customers behind it, are from the furnished CFO Commentary and are tagged accordingly. The margin is disclosed for the company and not by platform, the reportable segments being Compute and Networking and Graphics, so the company-wide rate is the one applied here, as on this shelf’s earlier NVIDIA work. Where a data center rate sits above the blend, the rate used here understates the gross profit and understates coverage with it, and that direction is ours rather than filed.
(5) Huang’s $40bn to $50bn a gigawatt and Barclays’ $32.5bn to $42bn with a $39bn midpoint at about 65 percent of a $60bn total are REPORTED through coverage. The two are measured at different points in the chain. Huang’s is NVIDIA’s own revenue; Barclays’ is the buyer’s compute spend, which includes whatever the channel keeps. The channel sensitivities of a tenth, a fifth and three tenths are ours and illustrative, and no filing supplies the share. The figure 8.7 appears twice in this piece as two different constructions that floor to the same number: coverage at the low case with a tenth of the compute bill kept by the channel, and the multiple the purchases leave after a total loss on the stake at the full gigawatt. The first is 8.763 and the second is 8.737. Huang separately put the cost of building a one gigawatt facility at $50bn to $60bn in remarks at the G20, which is a facility cost rather than a supplier’s revenue and is set aside here.
(6) A non-marketable equity security without significant influence is carried under ASC 321 at cost less impairment, adjusted for observable price changes, with remeasurement running through other income rather than through operating income. At the reported figures the stake would be about six percent, far below the twenty percent presumption in ASC 323 that would put it on the equity method and change the treatment. The pair supports 5.88 percent if the reported valuation is pre-money and 6.25 percent if it is post-money, the reports leave the basis open, and about six percent covers both. Under the equity method the investor recognizes its share of the investee’s earnings or losses in its own income, and the presumption runs on voting interest, so a non-voting class stays outside it however large. The twenty percent presumption is rebuttable in both directions, and a supplier holding board representation or protective provisions could be argued into significant influence below it; the terms of this placement are not filed, so the ASC 321 reading is the one the reported facts support rather than one a document settles. The reading that the size and the class are chosen to keep that share off the page is ours. The placement is an investing cash outflow on the statement of cash flows. The reading that the placement produces no operating expense, and that the gross profit on the resulting purchases arrives without an offset from it, is ours from that treatment.
(7) NVIDIA Corporation, Form 8-K dated August 17, 2026, accession 0001045810-26-000069, Item 1.01: residual value guaranties with SB Energy relating to leases for approximately 4.25 gigawatts of IT load at the Portsmouth Site, with credit support available at NVIDIA’s sole discretion for approximately 3.8 additional gigawatts, and an aggregate payment obligation cumulatively capped at $105 billion for the initial commitment. FILED. Read at length in “SB Energy, Quality of Cash: The Fourteenth Position.”
(8) NVIDIA’s Form 10-Q defines direct customers as add-in board manufacturers, distributors, original design manufacturers, original equipment manufacturers and system integrators, placing cloud service providers in a separate indirect class. FILED. The consequence for the realized revenue and margin is ours.
Analysis: Cape Fear Advisors.
This piece also appears on Substack. Cape Fear Advisors is an independent advisory firm based in Portsmouth, NH.
Contact Cape Fear Advisors