The $1.4 Trillion Circle: What the IMF, the BIS and the Filings Actually Say About AI's Circular Finance
By Alpesh Patel OBE, hedge fund manager and Financial Times author
A research-led guide to AI circular financing: what it is, how the deals work, what the data says about risk, and the five gauges that would tell you it is time to worry. Ultimately this helps you hold your AI exposure on evidence, not headlines.
Data as of 29 October 2026. This is education, not personal financial advice.
The body on the floor
It is 2001 and I am watching the telecom bubble unwind from the Bloomberg desk. Lucent Technologies, once the most widely held stock in America, has just refused a $90 million loan to one of its favourite customers, a fixed-wireless start-up called Winstar. Winstar goes bankrupt. Lucent, which had promised it up to $2 billion of financing to buy Lucent's own switches, writes off around $700 million in bad debts. The salesman had been the banker. The banker had been the customer's only real source of money. When the money stopped, the sales stopped, and the loans went bad together.
Now fast-forward to this month. Nvidia has put $30 billion into OpenAI's funding round and agreed to guarantee up to $105 billion for an Ohio data centre that OpenAI will lease for 20 years and fill with Nvidia chips. Broadcom is lending Anthropic up to $42 billion to lease Broadcom chips. The IMF's October Global Financial Stability Report counts 14 deals since 2023 with about $300 billion of equity and $650 billion of debt tied to $1.4 trillion of future revenue commitments.
It looks like the same crime scene. Supplier funds customer, customer buys from supplier, supplier books revenue. If you hold a Nasdaq tracker, a global fund or almost any pension, you are standing in that room.
I have spent the past fortnight going through the filings, the central-bank research and the S-1s. My conclusion is not "relax, nothing can go wrong". It is narrower and more useful than that. The loop is real, but it is not closed. Money from outside keeps coming into it, from paying customers, from profitable businesses and from shareholders who can afford the loss. In 2000 that outside money was missing. That is the clue that solves the case. First, though, you need to see how the loop works.
"In 2001 Lucent refused its star customer a $90 million loan and lost $700 million. In the second quarter of 2026, Anthropic collected $2.08 in revenue for every $1 it spent on compute."
What circular finance actually is
Circular financing is any arrangement where a supplier provides money (equity, loans, guarantees or credits) to a customer, and that customer uses the money to buy the supplier's products. Harvard's Program on Negotiation quotes Morningstar's Brian Colello describing the Nvidia–OpenAI framework as circular because "OpenAI will likely buy gear from Nvidia, which will reinvest those profits in OpenAI". At its worst it shades into round-tripping: two parties swap money to inflate revenue without creating any real profit.
In the AI build-out it comes in six main forms:
Structure | How the money moves | 2025-26 example |
|---|---|---|
Equity for purchases | Supplier buys shares in a customer; customer spends the cash on the supplier's products | Nvidia's $30 billion into OpenAI's $122 billion round, alongside Amazon (up to $50 billion) and SoftBank ($30 billion) |
Equity for cloud commitments | Cloud provider invests; the AI lab commits to spend on that cloud | Microsoft up to $5 billion and Nvidia up to $10 billion into Anthropic, with Anthropic pledging $30 billion of Azure spend |
Warrants for orders | Supplier hands the customer an option on its own shares in return for big orders | AMD giving OpenAI the option to buy up to roughly 10% of AMD as part of a multi-year chip deal |
Vendor loans | Supplier lends the customer money to buy or lease its kit | Broadcom lending Anthropic up to $42 billion, part of a $63 billion bank and Blackstone package |
Guarantees and capital pools | Supplier guarantees a customer's lease or debt, or organises financiers to lend to its customers | Nvidia's $105 billion Ohio guarantee and its $500 billion-plus partnership with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to fund Nvidia customers |
Off-balance-sheet vehicles | Chips go into a special purpose vehicle, which leases them back and sells stakes to investors | Amazon reportedly in talks to place about $8 billion of Nvidia processors in an SPV and lease them back |
A worked example
Take an illustrative supplier, call it ChipCo, with a 75% gross margin. It invests $10 billion in a model lab. The lab spends the $10 billion on ChipCo's chips.
ChipCo books $10 billion of revenue and $7.5 billion of gross profit.
ChipCo now owns a $10 billion stake in the lab.
The lab now owns $10 billion of computing capacity.
So far, so circular. What happens next depends on one thing: does anyone outside the loop pay the lab for what it builds with those chips?
If yes, the lab grows, ChipCo's revenue repeats, and ChipCo's stake rises in value. The loop worked like venture capital with a supply contract attached.
If no, ChipCo's revenue was a one-off funded by its own balance sheet, the stake gets written down, and both companies' numbers fall at the same time.
That second path is exactly what the Bank for International Settlements means when it says circular deals make reported AI demand "partly endogenous", meaning partly created by the industry's own financing decisions. It is a fair warning. It is also a question you can test with data.
Why companies do it at all
The BIS bulletin itself does not call these deals inherently unsound. It sets out four rational reasons: suppliers often understand a customer's prospects better than outside investors; funding a customer smooths demand for very capital-heavy products; customers invest in suppliers to secure scarce inputs; and when you are building custom chips and custom data centres, a shared stake protects both sides from being held to ransom by the other.
Anyone who has run a business knows the first reason. When I lend a regular client a little slack, it is because I know their order book better than their bank does. The question is not whether the loop exists. It is how big it is compared with the money arriving from outside.
The evidence
How big is the loop?
The scariest numbers come from the BIS. Mapping 1,246 AI firms, it found that 55.2% of investment into AI firms between 2021 and 2025 came from other AI firms, and 46.4% of AI-to-AI deal value involved an investor who was also a supplier or customer. Compute and cloud firms are financially tied to their customers in 15.2% of relationships, against a 3.3% baseline across more than 10,000 US customer-supplier pairs. The Bank of Korea estimates that circular deals make up 47% of Oracle's cloud backlog and 41% of Microsoft's.
Those numbers are real, and I will not wave them away. But notice what they measure: deals between AI firms, and backlogs, which are promises of future revenue. They do not measure where today's cash comes from. To answer that, zoom out from the deals to the cash-flow statements.
Clue one: outside money dwarfs the loop
Start with the biggest buyers. The four largest hyperscalers plus Oracle are on course for about $800 billion of capex in 2026, against roughly $707 billion of operating cash flow, according to Brookings. That operating cash flow is not circular money. It comes from Google searches, Instagram adverts, Office 365 subscriptions and Amazon parcels. The bulk of the spending on AI chips comes from businesses that already earn hundreds of billions of dollars a year from customers who have nothing to do with AI deal-making.
Now zoom in to one company, Microsoft, which publishes the clearest split.
Microsoft's revenue from OpenAI in fiscal 2026 was $24.1 billion, about 7% of its $331.8 billion total. Big, but 93% of revenue came from everyone else.
Its commercial backlog was $678 billion, up 84%; excluding OpenAI it still grew 25%, and all of the sequential growth in the June quarter came from customers outside the frontier model companies.
In the quarter just reported, Microsoft Cloud revenue rose 30% to $63.9 billion and Azure grew 48%.
In its June quarter it generated $55.4 billion of operating cash flow and still had $19.6 billion of free cash flow after $41 billion of capex.
Then Nvidia, the company at the centre of almost every loop diagram. In its July quarter it reported revenue of $96.2 billion, up 106%. Its fiscal 2026 free cash flow was $96.7 billion. Its $30 billion OpenAI cheque is real money, but it is about a third of one year's free cash flow, and Nvidia's whole book of private stakes was $51.2 billion against $229 billion of shareholders' equity at the end of July. Nvidia is not funding its own sales. Its customers are, overwhelmingly from their own operating cash.
Compare that with 2000. Lucent had about $8.4 billion of vendor-financing commitments in February 2000, on annual revenue of roughly $29 billion. The industry lent $25 billion to $30 billion at the 2000 peak, often to the fourth or fifth largest carrier in a market, after real banks had turned them down. It was loans, to weak companies, on revenue that did not exist.
There is a useful detail hidden in the Nvidia–OpenAI story too. The famous "$100 billion" was only ever a letter of intent, never a definitive agreement. Nvidia excluded it from its bookings guidance, wrote in its own filing that there was "no assurance" of a deal, and ended up writing a $30 billion equity cheque with no milestones attached. That is a supplier showing discipline, not a supplier desperate to stuff the channel.
What it means: the loop exists, but it sits on top of a vast river of outside cash. In the telecom bust the loop was the river.
Clue two: the customers can pay
This is the clue that cracks the case. In circular finance, the danger is never the loop itself. It is whether the last company in the chain has paying customers. Winstar did not. The customers today do.
Zoom out to the two labs at the centre of the web:
Anthropic's IPO prospectus shows first-half 2026 revenue of $16.3 billion, up roughly 14 times year on year. Third-quarter revenue was reported at $16.8 billion, up about 45% on the second quarter, with positive cash flow and a second straight quarter of adjusted profitability.
OpenAI's annualised revenue was about $50 billion at the end of September, and it expects at least $70 billion by the end of the year. ChatGPT reaches 1.2 billion people a week.
Now zoom in to the single most important number in this debate. In the second quarter of 2026, Anthropic spent $5.6 billion on compute and generated $11.6 billion in revenue, $2.08 of revenue for every dollar of compute. Two quarters earlier, in Q4 2025, it was collecting only 67 cents for every dollar of infrastructure cost. The ratio flipped from loss-making to more than two-to-one in six months.
That is the money leaving the loop. When Anthropic pays Amazon, Google or Microsoft for compute, an increasing share of that payment is funded by its own customers rather than by its investors. And those customers are not a handful of insiders. The S-1 shows more than 11,900 customers paying at least $100,000 a year, up from about 950 a year earlier, and more than 1,900 paying over $1 million. About 80% of first-half revenue was usage-based, meaning customers pay because they are actually using it.
Compare that with the telecom bust. Bad loans as a share of Lucent's vendor book went from 2.6% at the end of 2000 to 60% a year later; Nortel's went from 25.5% to 80%. The borrowers had built networks for traffic that had not yet arrived. Today's labs are short of capacity, not short of users. Nvidia describes its own outlook for next year's roughly 70% revenue growth as "supply-constrained".
I always say cash is a fact and accounting profit is an opinion. The cash facts here are paying customers, compute bills covered by revenue, and hyperscalers funding most of their spending from their own operations.
What it means: a loop with a paying customer at the end is a supply chain with financing attached. A loop without one is a Ponzi structure. The evidence says we have the first.
Clue three: the price is not 2000
Bubbles are not defined by how much is spent. They are defined by what investors pay for it. Here the comparison with 2000 breaks down completely.
Nvidia trades at about 17 times forward earnings, below the S&P 500's roughly 19 times. Cisco, the "picks and shovels" darling of 2000, traded at about 100 times.
The Nasdaq-100 trades at about 20.8 times forward earnings, in the bottom 10% of its own five-year range. In the dot-com era, index multiples topped 60.
The five biggest tech companies trade at roughly 23 times forward earnings, against about 65 times before the 2000 peak.
Multiples have fallen during the boom because earnings have grown faster than share prices. Bubbles do the opposite. Howard Marks made the behavioural point this month: it is not yet a classic bubble, because investors are openly discussing the risks rather than ignoring them. The market reaction to OpenAI's revenue restatement in October, when its "$70 billion" run rate turned out to be about $50 billion on a like-for-like basis, proves it. AI stocks sold off across the board from Nvidia to Oracle and CoreWeave. That is a market checking the numbers, not swallowing them.
Even the official worst case is milder than people assume. The IMF's adverse scenario, weaker and less durable AI investment and profits, takes 0.4% off global GDP, 1.3% off investment and 0.6% off consumption by 2028. In the same report, a global bank stress test found banks would "remain broadly resilient" to an AI-driven market correction combined with an energy shock and higher inflation. That would be painful for AI shareholders. It is not 2008.
What it means: in 2000 investors were paying for a future that might never come. In 2026 they are paying a below-market multiple for earnings that are already here.
How I might be wrong
You should not trust an article on this topic that only gives you good news. Here is the case against my own conclusion, which I take seriously.
The cash cushion is thinning. Brookings shows hyperscaler capex crossing 100% of operating cash flow in 2026 for the first time. Alphabet posted its first negative free-cash-flow quarter since its 2004 IPO, minus $5.9 billion, and Meta's third-quarter capex of $34.9 billion exceeded its operating cash flow, producing its first outflow in more than a decade. Microsoft is the only big hyperscaler that BofA does not expect to go cash-flow negative.
Debt is replacing cash. Alphabet, Microsoft, Amazon, Meta and Oracle issued about $230 billion of bonds in the first nine months of this year, more than double last year's total. Moody's flags a wall of AI-related loan maturities in 2027-2029, and about 17% of the 607 borrowers it tracks use payment-in-kind structures that defer cash interest.
The weak links are real. Neoclouds such as CoreWeave spent $6.42 billion on equipment against $0.68 billion of operating cash flow in one quarter, more than nine times. These are the Winstars of this cycle, and the BIS and IMF are right to watch them.
Nvidia's own numbers carry warning lights. Its July-quarter net income of $59.7 billion included $7.8 billion of gains on its stakes in AI companies, which is circularity showing up in the profit line. Receivables rose by $24.6 billion in the first half, days sales outstanding stretched from 45 to 60 days, and operating cash flow was only 40% of net income that quarter. Nvidia is also reportedly in talks over a $250 billion guarantee for OpenAI's Ohio campus.
The labs have made enormous promises. Anthropic has $498 billion of non-cancellable compute commitments, and a quarter of its 2025 revenue came from two customers. Revenue has to keep growing very fast to cover them.
Valuations are low on forward earnings, but those earnings could be at a cyclical peak. As the same valuation data warns, in cyclical sectors the multiple is lowest when earnings peak. The S&P 500 CAPE is near 41, close to its 1999 high of 44.
None of this changes my central finding, which is that the loop is anchored to real end-demand. It does change how you should hold it. This is a capital-intensive, cyclical boom with a growing debt layer. Some participants will fail, and share prices can fall 30% or more without the AI story being wrong. I remember Cisco bouncing back over $20 in the bust and then halving again. A fall in price is not, on its own, a reason to buy, and a frightening headline is not, on its own, a reason to sell.
Solving the case: five gauges to watch
The verdict is that circular finance is an amplifier, not the cause. It speeds the build-out up when end-demand is strong and would speed the fall when end-demand weakens. So stop watching the loop diagrams and start watching the demand. These are the five numbers I track, each one available in a quarterly filing:
Gauge | Why it matters | Reassuring | Worrying |
|---|---|---|---|
AI lab revenue per $1 of compute | Shows whether outside customers are funding the loop | Above $1.50 and rising (Anthropic: $2.08 in Q2) | Falling back below $1 |
Hyperscaler capex ÷ operating cash flow | Shows whether spending is self-funded | Below 100% (Microsoft: 65% in the June quarter, i.e. $35.8 billion cash capex ÷ $55.4 billion operating cash flow) | Sustained above 120% across the group |
Backlog growth excluding related parties | Separates real demand from circular commitments | Positive double digits (Microsoft ex-OpenAI: +25%) | Flat or falling ex-related parties |
Nvidia days sales outstanding and gains on its stakes | Shows whether revenue is turning into cash | DSO back below 50 days (60 days in the July quarter); stake gains a small share of profit | DSO above 75 days; stake gains doing the heavy lifting |
Neocloud and AI credit spreads, PIK share | The weak links break first | Stable spreads, falling PIK share | Widening spreads, rising PIK, refinancing failures in 2027 |
If three of those five turn red at the same time, it is time to reduce risk. Until then, the data says the circle has an exit, and paying customers are walking through it every day.
What this taught me
What 2001 taught me, and what this research confirmed, is that you never judge a financing structure by its shape. You judge it by who pays at the end. Lucent's loop ended with a start-up that had no customers. Today's loop ends with 1.2 billion weekly ChatGPT users, nearly 12,000 businesses each paying Anthropic six figures a year, and companies such as Microsoft paying for most of it out of their own operating cash.
Circular finance is not something to be complacent about. It is something to be numerate about.
Here is the question I will leave you with: of the AI companies you own, which ones could pay their bills if their biggest partner stopped writing cheques tomorrow? That single question separates the Microsofts from the Winstars.
Want to run these gauges on your own portfolio?
I run these five gauges on the AI names my members hold, quarter by quarter, using the same cash-flow-first approach I use for every stock. Join me on the next weekly webinar and I will show you how to read them on your own portfolio.
Book a call with my team. We'll look at what you hold, how you pick investments now, and whether the Great Investments Programme makes sense for you.
Sources
Every figure in this article is linked where it appears. Here is the full list in one place. The 2001 Bloomberg-desk and Cisco recollections are my own, and ChipCo is an illustrative example, not real data. Ratios such as "a third of free cash flow" and "more than nine times" are my calculations from the figures cited.
The 2001 telecom bust
Lazonick and March (MPRA): The Rise and Demise of Lucent Technologies (Winstar $90m loan, $2bn financing, $700m write-off)
Forbes, Feb 2000: Wired and restless (Lucent, America's most widely held stock)
Computerworld: Report: SEC widens accounting investigation ($8.4bn vendor-financing commitments)
Rotman School, University of Toronto: Gale of "Creative Destruction" Engulfs Nortel (case study) (Lucent revenue)
Newsweek: The Stupid Loan Bubble ($25-30bn industry vendor loans, Lucent and Nortel bad-loan ratios)
The deals, the IMF and the BIS
Tech Times: Nvidia OpenAI investment shrinks from $100B to $30B
CNBC: Nvidia wows Wall Street with a strong quarter ($105bn Ohio guarantee)
Yahoo Finance: Broadcom to lend Anthropic up to $42 billion
Reuters: Wall Street banks and Blackstone line up $63 billion for Broadcom chips for Anthropic
AI Cassindra, on the IMF Global Financial Stability Report (Oct 2026): $300bn equity and $650bn debt riding on $1.4 trillion of AI revenue commitments
Tom's Hardware, on the IMF report: AI sector wipeout risks global financial contagion (adverse scenario, bank stress test)
Santage, on the BIS bulletin: BIS: 55% of AI funding comes from AI firms (55.2%, 46.4%, 15.2% vs 3.3%, four rational reasons)
Seoul Economic Daily: Bank of Korea warns AI funding risks (Oracle 47%, Microsoft 41% of backlog; $230bn of bonds)
Harvard Program on Negotiation: What are circular deals?
AI Stock Wire: AI's circular deals explained (OpenAI's $122bn round)
The Economic Times: Tech firms channel billions into AI infrastructure (Microsoft/Nvidia/Anthropic, AMD warrants)
US News: Will circular financing pop the AI bubble? (Nvidia's $500bn-plus financier partnership)
Universal Asset Owners: Daily brief, 2 October 2026 (Amazon SPV talks)
Hyperscaler cash flow and Microsoft
Brookings Papers on Economic Activity: The AI build-out (slides) ($800bn capex vs $707bn operating cash flow)
Microsoft: Form 10-K, fiscal year 2026 ($24.1bn revenue from OpenAI; $331.8bn total revenue)
Microsoft: FY2026 Q4 earnings call ($678bn backlog, +25% ex-OpenAI, $55.4bn operating cash flow, $35.8bn cash capex)
Stock Titan: Microsoft earnings press release (Microsoft Cloud and Azure growth)
Barchart via Yahoo Finance: Microsoft may be the only hyperscaler to stay cash-flow positive ($41bn capex, $19.6bn free cash flow, BofA view)
XenoSpectrum: Big Tech's AI spending still relies on outside funding (capex ÷ operating cash flow by company, CoreWeave)
TMT Finance: 2026 hyperscaler capex tops US$700bn (Alphabet's negative free-cash-flow quarter)
Admiral Markets: AI capex stock sell-off (Meta capex vs operating cash flow)
Nvidia
Nvidia: Financial results for Q2 fiscal 2027 (revenue, receivables)
Nvidia: Q2 fiscal 2027 earnings call transcript (DSO, "supply-constrained" 70% growth outlook)
TickerScout: Nvidia filings summary (fiscal 2026 free cash flow)
Zero One Investment Research: Nvidia 2Q FY27 results (private stakes vs equity, DSO, cash conversion)
The Daily Star (AFP): Nvidia doubles revenue ($7.8bn gains on AI stakes)
Ground News, on the WSJ report: Nvidia in talks to guarantee $250 billion for OpenAI data centre
Anthropic and OpenAI
Fortune: Anthropic's S-1, by the numbers ($16.3bn revenue, 67 cents per $1, $498bn commitments)
Fortune: Anthropic's IPO prospectus ($2.08 revenue per $1 of compute)
Yahoo Finance: Anthropic takes another step toward its IPO (Q3 revenue)
The Indian Express: Anthropic IPO filing explained (11,900 customers above $100k, 80% usage-based)
Crunchbase News: The AI IPO wave (customer concentration)
Reuters: OpenAI's annualised revenue $20 billion less than previously signalled
Bloomberg: OpenAI expects $70 billion in annualised revenue by end of 2026
The Decoder: OpenAI revenue keeps surging (1.2 billion weekly users)
Ground News: Tech stocks drop after OpenAI revenue report
Valuations and credit
Chartloom: Nvidia forward P/E
Shattered: Nvidia at 17x forward earnings (Cisco at about 100x)
Dollar Liquidity: Forward P/E: S&P 500, Nasdaq-100, tech
Admiral Markets: Nasdaq-100 commentary (dot-com index multiples)
Luminix: AI bubble vs dot-com: valuations compared (top-five tech multiples, Howard Marks)
The Herald Business: Moody's warns of AI debt maturity surge (607 borrowers, 17% PIK)
@allday_stocks on X: Tech valuations vs dot-com peak (S&P 500 CAPE)
Risk disclaimer: This article is for educational and informational purposes only and is not financial advice, a personal recommendation or an offer to buy or sell any financial instrument. Alpesh Patel Ventures Ltd provides general investment education and is not authorised by the Financial Conduct Authority to give regulated investment advice. Investing involves risks, including the possible loss of the money you invest. The value of investments can go down as well as up. Past performance is not a reliable indicator of future results. Figures are taken from company filings, central-bank and IMF research and financial media as cited, and are correct to the best of my knowledge at 29 October 2026.
Comments