Find this research useful? Add aibubblequestion.com to your Google Preferred Sources, one click, affects only your own results.

The AI Bubble Question
All pages

The Mechanism

The AI Money Loop

Chip makers investing in their own customers. A cloud part-owning its biggest tenant. Debt secured by chips, sold on promises from companies funded by the chip maker. Here is the map, figure by figure.

Updated

This page applies the historical template to the 2023-2026 AI ecosystem. One honesty rule first: much of this lives in private contracts, so every figure below is tagged with its provenance. [disclosed]] means a company filing, earnings call, or official announcement. reported means journalism, including leak-based reporting. estimate means analyst or third-party modeling. The difference matters: in [1999, the loop's true size only became clear after the write-downs.

Loop 1: Nvidia's stakes in its own customers

In January 2026, Nvidia invested $2 billion in CoreWeave, its own GPU-cloud customer, at $87.20 per share, filed in CoreWeave's own 10-Q, not any Nvidia filing [disclosed]] ([CoreWeave, Form 10-Q, Q2 2026). Follow-up reporting has put its total stake near $3.65 billion, about 9% of the company, and Nvidia at more than $6 billion of committed CoreWeave services through 2032 [reported]] ([Motley Fool; Bloomberg); neither figure is confirmed in any filing found on either company's side, Nvidia's three most recent filings and CoreWeave's own Q2 2026 10-Q were checked and neither discloses an ownership percentage or a services-purchase figure. So: Nvidia funds the customer, supplies its chips, and buys services back from it, three sides of one loop.

More broadly, Nvidia's non-marketable equity securities totaled $47.898 billion, plus $3.3 billion of equity-method investments, about $51.2 billion combined, as of 26 Jul 2026 [disclosed]] ([Nvidia, Form 10-Q, Q2 FY2027), not broken out by investee. A widely reported OpenAI stake restructured from a floated $100 billion letter of intent down to roughly $30 billion has never appeared in any Nvidia filing, including this newest one [contested]] (see [shadow-debt for the fuller treatment of the same $100B figure). No public figure exists for what share of Nvidia's order book comes from companies it part-owns. Closest historical analog: telecom vendor financing, with equity stakes playing the role Lucent's customer loans played, a difference defenders consider decisive and skeptics consider cosmetic (both arguments here).

Loop 2: The Microsoft-OpenAI circle

Microsoft has made $13.0 billion of total funding commitments to OpenAI, of which $11.9 billion was funded as of 30 June 2026 [disclosed]], and holds an approximate **25% interest** on an as-converted basis disclosed, Microsoft's own filed language, not the 27% this page previously stated. No filed valuation figure exists for the stake, so the earlier $135 billion figure is dropped, and no filed revenue-share percentage exists, so the earlier 20% figure is dropped; both were reported, not filed. FY2026 revenue from commercial arrangements with OpenAI, including revenue-sharing payments, was **$24.1 billion** disclosed, against a **$6.0 billion receivable** disclosed. Flowing back: OpenAI has committed to **$250 billion of Azure services**. Microsoft's commercial remaining performance obligations, its contracted backlog, grew **84% year over year to $678 billion** as of 30 June 2026 disclosed ([Microsoft, Form 10-K, FY2026), a filed figure this page did not previously cite; Microsoft does not disclose a customer-level breakdown of that backlog, so the earlier claim that 45% of a prior quarter's increase came from OpenAI is dropped as unfiled.

Read that back slowly: Microsoft's own 10-K puts $24.1 billion of revenue and a $6.0 billion receivable against a single money-losing company it part-owns and helps fund, inside a backlog that grew to $678 billion the same year without disclosing how much of it is OpenAI's. The exposure is real and filed; its exact share of the whole backlog is not. Closest analogs: vendor financing (funding your customer's purchases from you) crossed with keiretsu-style mutual entanglement, where the health of each party's reported numbers depends on the other's.

Loop 3: OpenAI's commitments versus OpenAI's money

OpenAI's revenue: $13.07 billion of GAAP revenue in 2025, against roughly $34.0 billion in costs, a $20.92 billion operating loss, and a $38.53 billion loss attributable to OpenAI [reported]], from leaked audited financials the Financial Times has corroborated, with an [annualized run rateTake the most recent month's revenue and multiply by 12. It shows how fast a company is growing right now, but it is a projection, not money already earned.Full definition in the glossary that has since surpassed $40 billion as of mid-August 2026, up from about $25 billion in February 2026, reported via an internal staff communication, not an audited figure [reported]] (Bloomberg, 13 Aug 2026). OpenAI's commitments: the **$250 billion Azure agreement** (the Azure figure rests on Microsoft's own filings; the rest do not) plus large multi-year compute contracts with Oracle and CoreWeave reported, plus anchor-customer arrangements with AMD and Broadcom whose values are undisclosed ([Bloomberg). OpenAI is privately held and files nothing with the SEC, so none of its own revenue or commitment figures are audited; directionally, the commitments run well beyond the company's reported annual revenue, but that gap cannot be certified as a precise multiple without OpenAI's own audited statements. The capital behind those commitments is raised largely from Microsoft, Nvidia, and other ecosystem participants. Closest analog: the railway share call: capital promised, not yet possessed, dependent on the future staying friendly.

Loop 4: SPVs and private credit for data centers

A growing share of data-center construction is financed through special purpose vehiclesA separate legal company created to hold one project and its debt, keeping both off the parent company's books.Full definition in the glossary borrowing from private credit fundsLoans made by investment funds instead of banks, with little public disclosure. Tens of billions of it now finances AI data centers.Full definition in the glossary, structures like Meta's "Hyperion" financing with Blue Owl, Apollo, and peers. Bloomberg estimates roughly $50 to 80 billion in committed facilities across these vehicles [estimate]] ([Bloomberg); no regulator filing, ratings-agency methodology, or consolidated dataset was located behind that range in this pass, so it is presented here as Bloomberg's own reporting, not an independently verified aggregate. By design this debt sits off the tech companies' own balance sheets, and the vehicles' revenue rests on offtake agreementsA long-term contract to buy a project's future output, signed before it's built. Lenders rely on these promises, so the buyer's health becomes the project's foundation.Full definition in the glossary from AI tenants who are themselves funded by the same hyperscalers and chip vendors. Rating-agency commentary worries that true AI leverage is understated in consolidated accounts [reported]]. **Closest analogs:** the [1920s trust pyramid for the layering, and Enron for the off-balance-sheet form, with the crucial caveat that these SPVs are disclosed and (so far) nobody alleges fraud. The structural echo is the opacity, not the legality.

Loop 5: GPU-collateralized neocloud debt

The neocloudsA newer, smaller cloud company built specifically to rent out AI computing power (GPUs), often financed with debt secured against the chips themselves.Full definition in the glossary, CoreWeave, Lambda, Crusoe, Nebius, have raised multi-billion-dollar credit facilities secured largely by their GPU fleets and customer contracts [reported]] ([Bloomberg), with CoreWeave's own total principal debt reaching $35.551 billion at 30 June 2026, up from $25.149 billion three months earlier [disclosed]] ([CoreWeave, Form 10-Q, Q2 2026). CoreWeave's SEC-filed Q2 2026 earnings release confirms 1.5 gigawatts of active power and 3.7 gigawatts contracted; a "more than 5 gigawatts by 2030" target reported elsewhere traces only to earnings-call commentary, not a filing, and is presented here as unfiled [reported]]. The [collateralAn asset pledged to a lender so that if the borrower cannot pay, the lender can seize and sell it. Loans against fast-depreciating collateral are riskier.Full definition in the glossary is specialized hardware that loses most of its market value within 3-5 years, and the customers guaranteeing the revenue are often the same startups funded by Nvidia and the hyperscalers. Closest analog: Japan's collateral-based lending against inflating assets, with a faster-melting asset.

Loop 6: The vendor-financing question mark

As of mid-2026, no chip maker discloses a Lucent-style loan book. What exists instead is the functional equivalent: equity-funded prepayments, extended payment terms, and cloud credits granted to startups who spend them with the grantor [estimate / commentary]] ([Bloomberg). Whether that difference in form is a difference in substance is the central dispute, argued at full strength on the Case Against page.

How much of AI's demand is recycled?

No primary-source number exists. Triangulating the disclosed pieces, over $100 billion of Microsoft backlog from one part-owned counterparty, $40+ billion of Nvidia equity in its own customers, $50-80 billion of private credit resting on intra-ecosystem offtakes, analysts have begun estimating that a double-digit percentage of headline AI bookings and revenue is ultimately funded by capital recycled within the ecosystem itself [estimate]] ([Bloomberg). Bloomberg's summary: a web of interlinked investments that "raises the risk of cascading losses if AI falls short." For the historical base rates on how long such webs hold, and what has broken them, see Circular Financing, Explained and Conclusions. The harshest independent read of this same carousel is graded claim by claim on The Critics.