Nvidia's customer financing is disclosed. Its effect on revenue is not.

Nvidia closed fiscal 2026, the year ended 25 January 2026, with $215.9 billion of revenue, up 65%, and data-center revenue of $193.7 billion. In the same period it became one of the most active strategic investors in technology. The FY2026 10-K, filed 25 February 2026, discloses $17.5 billion invested in private companies and infrastructure funds "primarily to support early-stage startups." The Q1 FY2027 10-Q, filed 20 May 2026 for the quarter ended 26 April 2026, discloses a further $18.6 billion of such investment in a single quarter.¹

Many recipients of that capital also buy Nvidia silicon. That is the circularity worth testing: how much of the revenue Nvidia books comes from counterparties Nvidia itself finances? This note builds the ledger from filings only, then states plainly what the ledger can and cannot prove.

What the filings actually disclose

The revenue concentration is real but anonymous

The FY2026 10-K states that for fiscal 2026, sales to one direct customer represented 22% of total revenue and sales to another direct customer represented 14%, both primarily attributable to the Compute & Networking segment. For fiscal 2025, one direct customer had represented 12% and two others 11% each. The Q1 FY2027 10-Q, in Note 13 and again in MD&A, states that for the first quarter of fiscal 2027, three direct customers represented 21%, 17% and 16% of total revenue. The Q3 FY2026 10-Q, filed 19 November 2025, had disclosed four direct customers at 22%, 15%, 13% and 11%.

None of these customers is named. Nvidia labels them only as "direct customers," a category it defines to include add-in-board makers, distributors, ODMs, OEMs, cloud service providers, AI model makers and system integrators. The accounts-receivable note in the Q1 FY2027 10-Q (Note 7) adds that three direct customers accounted for 30%, 18% and 16% of the receivable balance as of 26 April 2026, against 25%, 18% and 13% at 25 January 2026.

The single most important sentence for this question is qualitative. In the Q1 FY2027 10-Q MD&A, under "Concentration of Revenue," Nvidia writes that it estimates "one AI research and deployment company contributed to a meaningful amount of our revenue by purchasing cloud services from our customers in the first quarter of fiscal year 2027." That is an indirect-customer acknowledgement, unnamed and unquantified. It is as close as any Nvidia filing comes to connecting a financed AI counterparty to its revenue.

The investment book

The balance-sheet line "Non-marketable securities" rose from $3.39 billion a year earlier to $22.25 billion at 25 January 2026, and then to $43.36 billion at 26 April 2026. Marketable equity securities rose from $12.89 billion at 25 January 2026 to $30.24 billion at 26 April 2026, driven by the Intel position converting to a marketable holding and by large unrealised gains: the Q1 FY2027 10-Q recorded $13.4 billion of net unrealised gains on publicly held equity securities and $2.6 billion on non-marketable securities in the quarter. Note 6 of that filing discloses total investment commitments of $27 billion as of 26 April 2026 "subject to certain contingencies," which Nvidia expects to fund through the remainder of fiscal 2027, plus $1.0 billion of equity-method infrastructure-fund investments with maximum loss exposure of $2.3 billion.

The FY2026 10-K also discloses how the balance got there, and the composition matters more than the total. Against an opening balance of $3.39 billion, Nvidia recorded $17.44 billion of net additions, $2.37 billion of unrealised gains, $0.85 billion reclassified out to marketable securities, and $0.10 billion of impairments and unrealised losses, closing at $22.25 billion. Capital actually deployed, rather than mark-to-market appreciation, accounts for roughly 92% of the increase. That distinction matters for what follows: a book that grew on revaluation would say little about what Nvidia has committed, while one that grew on additions says a great deal.

The ledger

The CoreWeave case is the cleanest two-way loop

CoreWeave buys Nvidia GPUs, which makes it Nvidia's customer; Nvidia owns equity in CoreWeave; and under the order form dated 9 September 2025, Nvidia "is obligated to purchase the residual unsold capacity through April 13, 2032," an arrangement with an initial value of $6.3 billion (CoreWeave 8-K, Item 1.01, filed 15 September 2025). The equity side is documented on both sides of the transaction. Nvidia's own Schedule 13G reported 24.2 million CoreWeave shares, roughly 7%, as of 31 March 2025. Following the January 2026 investment, CoreWeave's Schedule 13G/A reported Nvidia holding 47,213,353 Class A shares, or 11.5%, as of 23 January 2026. The MSA note in the 8-K confirms the dual role directly: "In addition to the MSA, NVIDIA supplies the Company with NVIDIA GPUs and is a stockholder of the Company."

The calculation and its limits

The brief asks for a circularity ratio. The honest answer is that the filings do not support a clean one, and here is precisely why.

To compute "share of revenue from financed counterparties" you need the revenue each financed counterparty books with Nvidia. Nvidia never discloses this. Its revenue-concentration note uses anonymous "direct customer" percentages, and the financed neoclouds and model-makers, CoreWeave, Nebius, OpenAI, Anthropic and Lambda, appear, if at all, as unnamed indirect customers. The one place Nvidia connects the two, the "one AI research and deployment company" sentence, is deliberately unquantified. Nothing in the filings lets a reader attach a revenue number to CoreWeave, to Nebius, or to any company Nvidia has financed.

The timing mismatch

The capital is committed up front and any revenue follows later. The CoreWeave capacity backstop runs to April 2032, the facility lease guarantees run 5 to 7 years, and the $27 billion of investment commitments is expected to be funded through the remainder of fiscal 2027. Revenue from the GPUs those counterparties buy is recognised on delivery, quarter by quarter. A commitment made now therefore backs sales that may arrive across the rest of the decade, or not at all. The exposure is disclosed and current; what it is meant to unlock is neither.

What is reported but not filed

These items dominate the popular "circular financing" narrative yet are not confirmed as line items in any Nvidia SEC filing reviewed for this note. They are segregated here deliberately. Press coverage is used only to locate them, never as proof of a filed fact.

• OpenAI, up to $100bn letter of intent, announced 22 September 2025. A non-binding LOI. Nvidia CFO Colette Kress said at the UBS Global Technology and AI Conference in Scottsdale, Arizona, on 2 December 2025: "We still haven't completed a definitive agreement, but we're working with them." She added that Nvidia's roughly $500bn of Blackwell and Vera Rubin bookings guidance for 2025-26 excludes the OpenAI work.

• OpenAI, $30bn equity investment, reported as part of a $110bn round announced 27 February 2026, with Jensen Huang referencing it publicly on 4 March 2026. It is not named in the Q1 FY2027 10-Q. Critically, the $18.6bn total private-company investment Nvidia discloses for that quarter is smaller than the reported $30bn, so the reported figure cannot be reconciled to the filing and may sit in a later period or a different structure.

• OpenAI, roughly $250bn Ohio data-center financing backstop, reported by the Wall Street Journal on 26 July 2026 and confirmed by Reuters and CNBC. The guarantees would backstop lease and construction debt, not the chips, for a 10-gigawatt campus in Pike County, Ohio, on a former uranium-enrichment site being developed by SoftBank's SB Energy; the first phase of about 800MW is targeted for 2028 and total project cost is put above $500bn. This is NOT confirmed in any SEC filing as of this writing. For scale, Nvidia's largest disclosed guarantee book is the $3.5bn aggregate facility lease guarantee in the Q1 FY2027 10-Q, roughly one seventy-first of the reported Ohio figure.

• OpenAI, up to $350bn of chip-purchase financing, reported alongside the backstop. Not filed.

• Anthropic, up to $10bn, announced 18 November 2025 alongside Microsoft's up to $5bn, with Anthropic committing to purchase $30bn of Azure compute and to an initial compute commitment of up to 1 gigawatt on Nvidia Grace Blackwell and Vera Rubin systems; the deal valued Anthropic at about $350bn, up from $183bn. The OpenAI opportunity and the Anthropic and Intel investments were named in the Q3 FY2026 10-Q risk factors under "no assurance that any investment will be completed on expected terms, if at all." That language does not carry into the Q1 FY2027 10-Q, whose text names none of these companies.

• Lambda, $1.5bn GPU leaseback, reported September 2025. Not confirmed as a named filing line.

• Corning up to $3.2bn and IREN up to $2.1bn, reported May 2026 as rights to invest. Not confirmed as named filing lines.

What would settle the open questions

• A related-party or disaggregated revenue note naming CoreWeave, Nebius or any financed neocloud would convert the qualitative "one AI research and deployment company" sentence into a number.

• An 8-K filing the OpenAI $30bn subscription agreement, or the Ohio backstop guarantee, would move those items from the reported column to the filed ledger.

• CoreWeave's own 10-K or 10-Q disclosure of revenue attributable to Nvidia, or purchases from Nvidia, would let an analyst triangulate one side of the loop from the counterparty's filings.