The Same Money, Counted Three Times: AI's Circular Financing Loop Explained

Executive Summary

• Nvidia is reportedly in talks to guarantee $250 billion of OpenAI's debt obligations — so a chip supplier would be standing behind its own customer's borrowing, to fund purchases of the supplier's own product. The deal is unsigned. The pattern it belongs to is not.
• Roughly $820 billion in AI-related lease obligations sits off the balance sheets of the world's largest technology companies, priced on those companies' credit ratings rather than the quality of the underlying assets — a structure that concentrates risk in pension funds and insurers.
• You do not need to have bought a single technology share to be exposed. Index funds and default pension allocations have already done it on your behalf.

The Signal

In late July 2026, the Wall Street Journal and Bloomberg reported that Nvidia was in talks to guarantee approximately $250 billion of OpenAI's data-centre lease obligations, and separately to finance around $350 billion of OpenAI's chip purchases. The arrangement was subsequently reworked and downsized: by mid-August 2026, Nvidia had agreed to guarantee up to $105 billion covering the first phase of the Ohio project, rather than the $250 billion first reported. On 27 July 2026, Nvidia shares fell approximately 5% (closing down 4.99%) and the cost of insuring Nvidia's debt against default recorded its largest single-day intraday rise since the contract began actively trading in November 2025.

The credit market moved before the equity market finished arguing. That sequencing matters.

The reported deal is not an isolated transaction. It belongs to a pattern visible in ordinary regulatory filings. OpenAI's audited 2025 accounts show $13.1 billion in revenue against a $20.9 billion operating loss. A company with those numbers cannot borrow at affordable rates on its own merits. So the companies that sell it compute — Nvidia, Microsoft, Google, Amazon, Oracle — lend their balance sheets instead, through equity stakes, vendor financing, and guarantees. The money returns to those same sellers as chip revenue and cloud contracts, which supports their share prices, which supports their capacity to keep extending credit.

The loop is not secret. It is disclosed. That is precisely what makes it worth understanding.

The Noise

The mainstream debate asks the wrong question. Almost every "AI bubble" argument collapses three separate questions into one.

Is AI useful? Largely settled — yes. Microsoft reports more than 30 million paid Copilot seats. Azure passed $100 billion in annual revenue for fiscal year 2026, growing 41% for the full year; in the fourth quarter alone, Azure and other cloud services revenue grew 43% year-over-year. Google Cloud grew 82% year-over-year to $24.8 billion in Q2 2026 (the quarter ended June 30, 2026). These are real customers paying for real outcomes.

Will the AI labs make money? Genuinely unresolved.

Will the people who financed the data centres get their money back? This is the question that actually matters — and it can be answered no even when the first two are answered yes.

Britain's railways were built and are still running. Most investors who funded the 1840s mania were wiped out. The fibre laid in the late 1990s carries today's internet; the firms that laid it went bankrupt, and their cable was bought out of administration for a fraction of its cost. The technology survived. The capital structure did not. The useful question is not "is AI real?" It is narrower: is the money funding this buildout being repaid by customers who are financially independent of the people lending it?

The evidence in the filings suggests, in several significant cases, it is not.

Forensic Analysis

Silicon. The concentration in reported revenue is stark and accelerating. In Nvidia's 2023 financial year, no single customer accounted for more than 10% of revenue. By the quarter ending April 2026, three customers accounted for 30%, 18%, and 16% of Nvidia's accounts receivable balance. Oracle's contracted backlog grew 363% in a year to $638 billion — but in the quarter when Oracle disclosed that OpenAI represented 54% of that backlog, its shares sold off. The following quarter, Oracle provided no customer breakdown at all. Oracle burned more than $20 billion in free cash flow over its last four quarters after capital expenditure, and S&P cut its credit rating to BBB−, one notch above junk, in July 2026. Moody's also placed a negative outlook on the company, and S&P now forecasts the free operating cash flow deficit widening to roughly negative $42 billion in FY2027 — nearly double its prior forecast of negative $24 billion.

In the second quarter of 2026, Alphabet reported approximately $98 billion in other income, driven by roughly $99 billion of net unrealized gains on equity securities — primarily from revaluing its stake in SpaceX, which was valued at $94.1 billion at quarter-end, with a smaller contribution from its Anthropic stake. Amazon booked $53.4 billion from revaluing Anthropic. These are non-cash, non-recurring gains drawn from the rising private valuations of companies that are themselves buying compute from the firms booking the gains. Circularity has stopped being a structural concern. It is now a line item in reported profit.

Stone. The risk does not stay inside the technology sector. Moody's estimates that hyperscaler lease obligations total roughly $1.2 trillion, of which approximately $820 billion relates to leases on data centres not yet built and therefore not yet commenced — meaning they are largely not yet recognised on balance sheets under current accounting rules. The BIS, in its March 2026 Quarterly Review, described hyperscaler lease arrangements as off-balance-sheet financing that deepens the links between hyperscalers and non-bank investors — the phrase 'shadow borrowing' came from Bloomberg's coverage of that report, not from the BIS itself. Separately, in its June 2026 Annual Economic Report, the BIS identified AI capital-expenditure sustainability as one of several key global financial stability risks, alongside financial vulnerabilities, fiscal pressures and the return of inflation. Its warning was that a reversal of AI optimism could carry major financial consequences, precipitating a sharp pullback in asset prices and disruptive macro-financial feedback loops. The BIS did not itself draw a comparison with the last financial crisis.

The structural echo is 2006, not 1999. An off-balance-sheet vehicle, holding a long-dated asset, funded by outside investors, backed by an implicit promise from a highly-rated sponsor, and sold to insurers and pension funds on the strength of that sponsor's name — that is, in its essentials, the structured investment vehicle of the pre-crisis era. The BIS said so directly. The underlying assets are real buildings, not tranched subprime claims, and the sponsors are among the most cash-generative companies in history. This is not 2008 waiting to happen. But the machinery is recognisable.

Strategic Implication

Three specific impacts on anyone holding a pension, running a business, or planning capital expenditure.

Your pension is already exposed, whether you chose it or not. The top ten companies represent roughly 40% of the S&P 500 — the most concentrated in decades. The US is approximately 72% of the MSCI World index. Default pension funds track these indices. If the AI revenue thesis disappoints, an earnings-justified premium concentrated in one technological bet unwinds faster than a diversified one. The exposure is not hypothetical; it is structural.

The depreciation assumption is a tell worth watching. Chips are depreciated over an assumed useful life — a management estimate, not a fact. Microsoft, Alphabet, and Oracle assume six years. Meta extended its depreciation schedule in 2025, worth $2.9 billion to that year's profit, and Amazon shortened its assumed server life from six years to five as of January 2025 (disclosed on its February 2025 earnings call). Independent analysis suggests that if every major operator moved to a four-year life, reported operating income across the group would fall by roughly $46 billion in 2026 alone. If another hyperscaler follows Amazon and shortens its schedule, that is management signalling — in the one language that is hard to spin — that the assets are wearing out faster than the financing assumes.

Credit markets are the honest signal; watch them, not the equity commentary. In late July, Meta and BlackRock brought a $12.5 billion data-centre financing that drew what International Financing Review called "uncomfortably low coverage for a deal of such significance" — well under half the normal order-book coverage for investment-grade paper. Barclays' Scott Schulte put it plainly: "Not only has the debt volume been heavy but incrementally none of them have performed, so as a debt buyer you feel as though you are perpetually losing money." A pulled or failed deal would be a genuine inflection point. A return to normal coverage would be a genuine reassurance.

The Long View

The question is not whether artificial intelligence works. It does, and it will keep working. The question is whether a structure in which the supplier funds the customer, the customer's promises justify the debt, the debt is priced off the supplier's credit rating, and the paper ends up in retirement portfolios — whether that structure has worked out well before.

It has not.

The railways got built. The fibre got laid. The technology arrived exactly as promised. The people who financed it did not get their money back. That is not a reason to disbelieve in AI. It is a reason to ask a different question about it — and to be clear, before the listings arrive and the index rules trigger, about who is expected to hold the paper when the answer comes.

Stone Truth: The loop is disclosed. Disclosed is not the same as safe.