The Circle and the Bubble

## How AI's interlocking deals became the central question in global finance

Researched and written August 2026. Figures reflect reporting available at that date; many are contested or move quickly, and the article flags where that is the case. This is analysis, not investment advice.

---

1. The shape of the problem

In September 2025, Nvidia signed a letter of intent to invest as much as $100 billion in OpenAI, tied to the deployment of at least 10 gigawatts of Nvidia systems. [1] The structure was easy to describe and hard to categorise: the world's dominant seller of AI chips would put capital into the world's most prominent buyer of AI chips, and that capital would substantially return to the seller as revenue.

Ten months later, in July 2026, the same two companies were reported to be negotiating something considerably larger, a financial backstop of roughly $250 billion that would let OpenAI raise debt against Nvidia's credit rather than its own, in order to lease a 10-gigawatt data centre campus in Pike County, Ohio, being developed by SoftBank's energy subsidiary. A parallel negotiation covered chip-purchase financing that could reach a further $350 billion. The total project cost, including silicon, was reported above $500 billion, which would make it the largest data centre development ever announced. The reason the guarantee is needed is unambiguous and was reported plainly: OpenAI has not turned a profit and cannot obtain an investment-grade credit rating on its own. [2][3][4]

That single deal contains, in compressed form, nearly every element of what critics call circular financing and what participants call ecosystem investment: a supplier funding a customer, a customer's obligations underwritten by a supplier's balance sheet, an asset whose value is asserted rather than tested by an arm's-length market, and a project whose scale exceeds the annual revenue of the company that will occupy it by more than an order of magnitude.

Whether that arrangement is a flywheel or a house of cards is now, genuinely, one of the most consequential open questions in the global economy. The Bank for International Settlements, the central bankers' central bank, devoted a section of its 2026 Annual Economic Report to it, naming circular financing explicitly and warning that deal terms are typically poorly disclosed, with risks that the same asset is pledged multiple times. [5]

---

2. A taxonomy of the circle

"Circular financing" is used loosely in press coverage to mean several structurally distinct things. Separating them matters, because they carry different risks.

a) Equity for purchase commitments. The supplier takes a stake in the buyer; the buyer commits to multi-year purchases. Nvidia's OpenAI investment is the archetype. Microsoft and Nvidia's combined commitment of up to $15 billion to Anthropic in November 2025, paired with Anthropic's commitment to buy $30 billion of Azure compute and contract up to a gigawatt of Nvidia capacity, is the same shape with different names in the boxes. [6][7] (Disclosure: Anthropic makes Claude, the assistant that researched and drafted this article.)

b) Warrants for volume. AMD's October 2025 agreement with OpenAI covers 6 gigawatts of Instinct GPUs and grants OpenAI a warrant for up to 160 million AMD shares at one cent each, roughly 10% of the company, vesting in tranches tied both to deployment milestones and to AMD's own share price reaching targets as high as $600. [8][9] The customer is paid in supplier equity for buying the supplier's product, and the payment grows as the market rewards the announcement. AMD subsequently offered Meta a near-identical structure in February 2026, which suggests this is a repeatable financing template rather than a one-off concession. [10]

c) Vendor guarantees and backstops. Rather than investing, the supplier stands behind the customer's obligations. The proposed Nvidia and OpenAI Ohio backstop is the largest example. Smaller ones are already routine: Nvidia agreed in September 2025 to lease unsold capacity from CoreWeave in an arrangement reported at $6.3 billion, and Google has guaranteed roughly $44 billion of other companies' data centre rent. [11][12]

d) Off-balance-sheet vehicles. Meta's Hyperion joint venture, reported at around $27 billion, places 80% ownership with funds managed by Blue Owl and 20% with Meta, keeping much of the project debt off Meta's balance sheet. [13] Similar structures are spreading through securitisations and project finance. Independent estimates of off-balance-sheet data centre financing run to roughly $800 billion. [13]

e) Reflexive mark-to-market. This one receives the least attention and may be the strangest. Under US accounting rules, an investor must mark a private holding to fair value when a new financing round establishes a reference price. Amazon disclosed $16.8 billion of pre-tax gains from its Anthropic stake in Q1 2026, more than 40% of its pre-tax income for the period, triggered by Anthropic's Series G and the conversion of some convertible notes. In Q2 2026 the figure was $53.4 billion, roughly two-thirds of pre-tax income. Alphabet's Q1 2026 profit of $62.6 billion, up 81% year on year, included a large markup on the same holding. Fortune reported that roughly half of both companies' headline AI profits that quarter came from the stake rather than from operations. [14][15][16]

The accounting is entirely conventional; tax and accounting specialists quoted at the time said so. [16] The reflexivity is the issue. Amazon and Google are simultaneously Anthropic's investors, its compute vendors, and the beneficiaries of the rising valuation that their own investment helps to set. Nothing here is improper. But the earnings that result are not the same kind of object as earnings from selling something to a stranger.

---

3. How big is the circle?

The honest answer is that nobody knows precisely, which is itself the finding.

The publicly announced numbers are large. OpenAI's stacked infrastructure commitments, across Oracle, Microsoft, Amazon, Nvidia, AMD, Broadcom and CoreWeave, have been reported at above $1.15 trillion through 2035. [17][18] Analysts tracking the interlocking arrangements have put the circular component north of $800 billion. [19] Bloomberg has mapped the deals under the label "circular deals" and put overlapping commitments at roughly $1 trillion. [20]

Against those commitments sit revenues that are growing extraordinarily fast and are still far smaller. OpenAI's annualised run rate reached roughly $25 billion by early 2026, on Q1 2026 revenue of about $5.7 billion and cash burn of about $3.7 billion, with a projected 2026 loss near $14 billion. [21][22] Anthropic's reported run rate climbed steeply through the first half of 2026, though the comparison is disputed. OpenAI has argued that gross-versus-net accounting overstates the gap. [23] Run-rate figures are last-month-annualised and are not the same as booked revenue; treat all of them with caution.

The capital expenditure side is firmer. The four largest US hyperscalers guided to roughly $725 billion of capex for 2026, up from about $410 billion in 2025, per earnings compiled by the Financial Times. [24] The BIS put the five largest hyperscalers above $1 trillion of AI-related capex across 2025 and 2026 combined. [5]

The financing mix has shifted. BIS research found that outstanding private credit loans to AI-related firms grew from near zero in 2016 to about $200 billion in 2025, rising from under 2% to 8% of total private credit lending. AI firms raised $243 billion through bonds in 2025, up from $79 billion in 2023. Direct lending funds have quadrupled their exposure to AI and IT over five years, to roughly 15% of their portfolios. [5][25]

---

4. The case for alarm

Four mechanisms recur in the critical literature.

Demand cannot be distinguished from its own financing. When a supplier funds a customer who spends the money on the supplier, the resulting revenue growth is real as an accounting matter but tells you very little about end-user appetite. As the BIS put it in structural terms, hyperscalers, chip makers and labs are linked through a web of private arrangements whose terms are typically poorly disclosed. [5] The signal that ordinarily disciplines capital allocation, customers spending their own money, is partly muted.

The same asset can support several claims at once. This is the BIS's sharpest warning. [5] Data centre construction is increasingly outsourced to third parties who lease facilities back on long-dated contracts with embedded exit clauses; GPUs serve as collateral for asset-backed loans; equity stakes in the tenant are carried on the landlord's balance sheet. In a benign environment these layers are independent. In a stress event they correlate.

Depreciation is doing heavy lifting. In November 2025, Michael Burry argued that hyperscalers were flattering earnings by depreciating Nvidia-based hardware over five to six years when the real economic life is closer to two or three, estimating $176 billion of understated depreciation across 2026 to 2028 and singling out Oracle and Meta as overstated by roughly 27% and 21% respectively by 2028. [26] The estimate is difficult to verify and Nvidia pushed back, saying customers consistently observe four-to-six-year lives. [27] But the underlying divergence is documented: Amazon shortened its server useful-life estimate from six years to five in early 2025, citing the pace of AI development, while Meta extended its own. [28] The assumption is not a rounding error. It determines whether a large share of reported hyperscaler profit exists.

Concentration converts a sector problem into a systemic one. The BIS reported that AI firms' share of total US market capitalisation rose from 24% to 40% between 2022 and 2025, and noted that household equity exposure is higher than in past cycles, so a correction would transmit to consumption through wealth effects more forcefully than before. [5][25]

---

5. The precedent everyone reaches for, and how to use it properly

The comparison is to late-1990s telecom vendor financing, and it is apt in mechanism. Lucent committed roughly $8.1 billion in vendor financing, Nortel extended about $3.1 billion, and Cisco promised around $2.4 billion in customer loans, lending to cash-strapped carriers so they could buy equipment. [29][30] When the carriers failed, the losses returned to the vendors: Lucent booked bad-debt provisions of roughly $2.2 billion in 2001 and $1.3 billion in 2002. [30] Global telecom stocks lost more than $2 trillion of market value between 2000 and 2002, and WorldCom and Global Crossing went bankrupt. [31] Jim Cramer, invoking exactly this history in July 2026, offered the compressed lesson: don't lend to the people buying your goods. [32]

But the precedent cuts two ways, and the second cut is usually omitted. The fibre laid in that build-out, derided at the time as "dark fibre" and a monument to waste, became the cheap capacity that carried broadband, cloud computing and video streaming through the following two decades. [31] The technology thesis was correct. The capital structure was not. Investors were destroyed and the infrastructure was useful.

That is the more precise framing of the current question. It is not "is AI real?", a question with an increasingly obvious answer. It is whether the specific financial structures built around AI can survive the gap between when capital is spent and when returns arrive, and who absorbs the loss if they cannot.

---

6. The case against alarm

The bull case is substantive and deserves engaging on its merits rather than as a foil.

The spenders are profitable. In 2000, the archetypal bubble company was pre-revenue. The companies driving 2026 capex are among the most profitable enterprises in history, funding much of the build-out from operating cash flow. J.P. Morgan Asset Management made this the crux of its response to the circularity concern: the better question is not whether today's deals resemble the dot-com era, but whether the fundamentals do, and 1990s build-outs were financed by firms with limited profitability and heavy reliance on external capital. [33]

Demand is contracted, not speculative. Neoclouds have hedged with take-or-pay structures: CoreWeave draws the overwhelming majority of revenue from multi-year fixed commitments, against a contracted backlog reported at $99.4 billion as of March 2026. [34] Capacity is being built against signed contracts rather than into a vacuum.

Revenue growth is genuinely steep. Google Cloud grew 63% year on year in Q1 2026 and 82% by Q2; AWS grew 37% in Q2 2026, its fastest in eighteen quarters. [35][36] Enterprise generative-AI spending roughly tripled between 2024 and 2025. [37] These are not eyeball metrics.

The market is not pricing catastrophe. Prediction markets have consistently put the probability of a defined AI bubble burst during 2026 at roughly 17% to 21%. [38]

Being short on compute may be more dangerous than overspending. If a small number of firms will dominate, the rational strategy for each is to over-invest. The BIS models this explicitly as a contest and, notably, finds that the contest logic itself is the risk: as competitive pressure drives capex higher, the net economic surplus for the sector as a whole declines and could turn negative in adverse scenarios. [5] Individually rational, collectively expensive.

---

7. What the instruments are actually saying, right now

The most useful development of mid-2026 is that equity and credit markets have begun to disagree, and credit is usually the earlier signal.

The BIS observed in January 2026 that equity prices had run far ahead of debt-market pricing, and by June it noted that credit spreads of some AI firms had already begun widening even as equity markets continued pricing significant upside. [5][25]

That divergence has since widened. Reporting in July 2026 described credit default swap spreads on the large hyperscalers more than doubling since the start of 2025 and standing well above their 2022 bear-market peaks; Oracle's five-year CDS reached a three-year high; Nvidia's reached record levels. Meta's Texas data-centre financing priced at yields comparable to B- rated debt. Oracle's stock was down roughly 35% year to date and Meta's roughly 17%, while the Nasdaq was up around 8% and trailing the S&P 500 for the first time since 2022. [39][40][41]

Demand for the paper is thinning. Coverage ratios on hyperscaler bond issues fell from roughly 5x in February 2026 to below 2x by July, while the equivalent measure for investment-grade bonds overall barely moved. [42] Record investment-grade issuance from the AI complex, around $182 billion in 2026 by late July, is meeting a shallower bid. [39]

Meanwhile the structures themselves are showing where the marginal buyer's tolerance ends. CoreWeave's $8.5 billion delayed-draw term loan in March 2026, anchored on a Meta contract, was rated A3, the first GPU-backed loan to reach investment grade. Its follow-on $3.1 billion facility two months later, collateralised by two non-investment-grade customer contracts, was rated Ba2/BB+ and priced higher. [11] The mechanism is legible: as the highest-quality counterparty relationships are consumed as collateral, each successive tranche costs more.

None of this is a crisis. All of it is the market beginning to price the thing that the equity indices have not.

---

8. What would actually resolve the question

Five indicators, in rough order of informational value:

1. Free cash flow relative to capex at the hyperscalers. Reported profit can be managed through depreciation assumptions; cash cannot. Amazon's free cash flow collapsed roughly 95% in Q1 2026. [15] Watch whether that reverses.
2. Whether reported earnings survive an honest useful-life debate. Any further shortening of server lives at a major hyperscaler would validate the depreciation critique in the most concrete way available.
3. Credit spreads and bond coverage ratios, not equity prices. Credit has been the leading indicator so far in this cycle.
4. Enterprise ROI attribution. The single most damaging data point in the bear case has been the repeated finding that a large majority of enterprise generative-AI deployments show no measurable return. [37][43] The corresponding bull signal is specific and testable: a non-technology S&P 500 company attributing concrete margin expansion to AI on an earnings call.
5. Whether the announced structures actually close. The Nvidia and OpenAI Ohio backstop was, as of late July 2026, still under negotiation and subject to change. [3][4] Announced commitments and executed ones are different objects, and the gap between them is where narratives go to be tested.

---

9. Conclusion

Circular financing is not fraud, and treating it as such misses what makes it interesting. Every element is legal, disclosed to the degree required, and defensible on its own terms. A chip maker taking equity in a lab it supplies is a coherent strategy. Marking an appreciating stake to a fresh funding round is what the accounting standards require. Guaranteeing a customer's lease is a way of accelerating a build-out that both parties want.

The problem is what the aggregate does to information. Prices, revenues and credit ratings are useful because they compress the independent judgements of many parties. When the same small group occupies the position of investor, supplier, customer, landlord and lender simultaneously, those judgements stop being independent, and the numbers they produce stop carrying the information they appear to carry. The BIS's phrase for the consequence, the same asset supporting multiple claims, is a precise description of what happens when a system loses the ability to check itself. [5]

The likeliest outcome is neither vindication nor collapse but the historically ordinary one: the technology proves transformative, the infrastructure proves useful, and a significant fraction of the capital structure built around it does not survive the wait. That was the telecom outcome. It is not a comforting resolution for anyone holding the paper, and it is not an argument against the technology. It is an argument for reading the credit market more carefully than the equity market, and for treating any number that has travelled around the circle more than once with the scepticism it has earned.

---

References

1. Reuters / CNBC reporting on Nvidia's letter of intent to invest up to $100 billion in OpenAI, September 2025. https://www.cnbc.com/2025/10/06/openai-amd-chip-deal-ai.html
2. Wall Street Journal, first report of the proposed ~$250bn Nvidia backstop, July 26 to 27, 2026 (paywalled; summarised in items 3 and 4).
3. CNBC, "Nvidia and OpenAI in talks for up to $250 billion backstop to fund AI infrastructure plans," July 27, 2026. https://www.cnbc.com/2026/07/27/nvidia-and-openai-in-talks-for-up-to-250-billion-dollar-ai-backstop.html
4. Data Center Dynamics, "Nvidia considers $250bn backstop for OpenAI's planned 10GW Ohio data center," July 2026. https://www.datacenterdynamics.com/en/news/nvidia-considers-250bn-backstop-for-openais-planned-10gw-ohio-data-center-report/
5. Bank for International Settlements, Annual Economic Report 2026, Chapter I: "Progress and peril," June 28, 2026. https://www.bis.org/publ/arpdf/ar2026e1.htm
6. Bloomberg, "Microsoft, Nvidia to Invest Up to $15 Billion in Anthropic," November 18, 2025. https://www.bloomberg.com/news/articles/2025-11-18/microsoft-nvidia-to-invest-up-to-15-billion-in-anthropic
7. Axios, "Anthropic lands $15 billion investment from Microsoft, Nvidia," November 18, 2025. https://www.axios.com/2025/11/18/anthropic-microsoft-nvidia-15-billion
8. AMD investor relations, "AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs," October 6, 2025. https://ir.amd.com/news-events/press-releases/detail/1260/amd-and-openai-announce-strategic-partnership-to-deploy-6-gigawatts-of-amd-gpus
9. CNBC, "OpenAI looks to take 10% stake in AMD through AI chip deal," October 6, 2025. https://www.cnbc.com/2025/10/06/openai-amd-chip-deal-ai.html
10. IntuitionLabs, analysis of the OpenAI and AMD structure and the February 2026 AMD and Meta replication. https://intuitionlabs.ai/articles/openai-amd-ai-hardware-partnership
11. TechTimes, analysis of Nvidia neocloud equity and backstop arrangements and CoreWeave's DDTL facilities, July 2026. https://www.techtimes.com/articles/320239/20260712/nvidia-circular-financing-249b-coreweave-debt-puts-pension-funds-risk.htm
12. The Next Web, "Nvidia in talks to guarantee $250bn of OpenAI data centre debt," July 2026. https://thenextweb.com/news/nvidia-openai-250-billion-backstop-ohio-data-centre
13. Forbes, "Bond Investors Push Back As AI Debt Heads Toward $570 Billion," July 17, 2026. https://www.forbes.com/sites/robertszczerba/2026/07/17/bond-investors-push-back-as-ai-debt-heads-toward-570-billion/
14. The Next Web, "Amazon Q1 revenue hits $181.5B but $16.8B Anthropic gain inflates net income," May 2026. https://thenextweb.com/news/amazon-q1-2026-anthropic-aws-earnings
15. Digital Applied, analysis of Amazon Q2 2026 earnings and the $53.4bn non-operating mark-to-market gain, July 30, 2026. https://www.digitalapplied.com/blog/amazon-q2-2026-earnings-ads-aws-anthropic-gain
16. Fortune, "Half of Google's and Amazon's 'blowout AI profits' came from a stake in Anthropic," April 30, 2026. https://fortune.com/2026/04/30/google-amazon-ai-profits-anthropic-stake-bubble-earnings-2026/
17. Morningstar, "Ahead of IPOs, AI Giants Keep Making Circular Deals. Here's Why That's a Risk," May 2026. https://www.morningstar.com/stocks/ahead-ipos-ai-giants-keep-making-circular-deals-heres-why-thats-risk
18. Tomasz Tunguz, "OpenAI's $1 Trillion Infrastructure Spend," breakdown of vendor-by-vendor commitments. https://tomtunguz.com/openai-hardware-spending-2025-2035/
19. Columbia University (CUIT blogs), "The Hidden Risk in AI's Circular Financing Ecosystem," November 2025. https://blogs.cuit.columbia.edu/gjb2124/circular-financing/
20. Bloomberg Graphics, "AI Circular Deals: How Microsoft, OpenAI and Nvidia Keep Paying Each Other," January 2026. https://www.bloomberg.com/graphics/2026-ai-circular-deals/
21. Luminix, OpenAI financial fact sheet, June 2026 (compilation of The Information, Reuters and Sacra reporting). https://www.useluminix.com/reports/company-overviews/openai-financial-fact-sheet-june-2026
22. Value Add VC, "OpenAI Revenue 2026: $25B ARR, a $14B Loss," June 2026. https://valueaddvc.com/blog/openai-revenue-2026-25b-arr-14b-loss-and-why-profitability-waits-until-2030
23. Value Add VC, comparison of OpenAI, Anthropic and Alphabet valuations and the disputed gross-versus-net revenue methodology, July 2026. https://valueaddvc.com/blog/openai-vs-anthropic-vs-google-valuation-2026-852b-965b-and-4-2t-compared
24. Tom's Hardware, summarising Financial Times compilation of Q1 2026 hyperscaler capex guidance, April 30, 2026. https://www.tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion
25. Bank for International Settlements, BIS Bulletin No 130, "AI and the global economy: implications for central banks," July 28, 2026. https://www.bis.org/publ/bisbull130.pdf ; see also BIS Bulletin No 120, "Financing the AI boom: from cash flows to debt," January 7, 2026. https://www.bis.org/publ/bisbull120.htm
26. CNBC, "'Big Short' investor Michael Burry accuses AI hyperscalers of artificially boosting earnings," November 11, 2025. https://www.cnbc.com/2025/11/11/big-short-investor-michael-burry-accuses-ai-hyperscalers-of-artificially-boosting-earnings.html
27. National Law Review / Deep Quarry, "Useful Lives of GPUs: Key Considerations," including Nvidia's response. https://natlawreview.com/article/deep-quarry-useful-lives-gpus-key-considerations
28. Dave Friedman, "Neoclouds Hold More Than $20 Billion in GPU-Backed Debt," February 2026, documenting hyperscaler useful-life changes. https://davefriedman.substack.com/p/neoclouds-hold-more-than-20-billion
29. Tomasz Tunguz, "Circular Financing: Does Nvidia's $110B Bet Echo the Telecom Bubble?" October 2025. https://tomtunguz.com/nvidia_nortel_vendor_financing_comparison/
30. Columbia University (CUIT blogs), as above, on Lucent, Nortel and Cisco vendor financing and subsequent bad-debt provisions. https://blogs.cuit.columbia.edu/gjb2124/circular-financing/
31. The Bubble Bubble, "The Late 1990s Telecom Bubble," on the crash and the dark-fibre aftermath. https://www.thebubblebubble.com/telecom-bubble/
32. CNBC, "Jim Cramer warns AI's circular financing frenzy echoes the dot-com bubble," July 27, 2026. https://www.cnbc.com/2026/07/27/jim-cramer-warns-ai-circular-financing-echoes-dot-com-bubble.html
33. J.P. Morgan Asset Management, "Does circularity in AI deals warn of a bubble?" October 2025. https://am.jpmorgan.com/us/en/asset-management/adv/insights/market-insights/market-updates/on-the-minds-of-investors/does-circularity-in-ai-deals-warn-of-a-bubble/
34. Luminix, deep dive on the neocloud GPU rental industry, June 2026. https://www.useluminix.com/reports/industry-analysis/deep-dive-on-the-neocloud-gpu-rental-industry-coreweave-lambda-crusoe
35. Yahoo Finance / Futurum, hyperscaler Q1 2026 results and cloud growth rates. https://finance.yahoo.com/sectors/technology/articles/hyperscalers-hit-700-billion-2026-111243744.html
36. Digital Applied, Amazon Q2 2026 results (AWS +37%). https://www.digitalapplied.com/blog/amazon-q2-2026-earnings-ads-aws-anthropic-gain
37. IntuitionLabs, "AI Bubble vs. Dot-com Bubble: A Data-Driven Comparison," including enterprise GenAI spending growth and the MIT ROI finding. https://intuitionlabs.ai/articles/ai-bubble-vs-dot-com-comparison
38. Polymarket, "AI bubble burst by...?" market, accessed August 3, 2026. https://polymarket.com/event/ai-bubble-burst-by
39. Benzinga, "Big Tech's $182 Billion AI Debt Spree: How META, NVDA and AMZN Are Reshaping Credit Markets," July 2026. https://www.benzinga.com/markets/equities/26/07/60490919/big-techs-182-billion-ai-debt-spree-how-meta-nvda-and-amzn-are-reshaping-credit-markets
40. Benzinga, "NVIDIA Stock Is Still Up, But $250 Billion AI Risk Has Spooked The Debt Market," July 31, 2026, citing Financial Times reporting. https://www.benzinga.com/markets/tech/26/07/60832453/nvidia-stock-is-still-up-but-250-billion-ai-risk-has-spooked-the-debt-market
41. Motley Fool, "The S&P 500 and Dow Are on Track to Beat the Nasdaq for the First Time Since 2022," July 29, 2026. https://www.fool.com/investing/2026/07/29/the-sp-500-and-dow-are-on-track-to-beat-the-nasdaq/
42. Fortune, "The AI boom is increasingly built on debt, but investor demand is plunging," July 17, 2026. https://fortune.com/2026/07/17/ai-boom-debt-blitz-investor-demand-hyperscaler-bond-issuance/
43. StockCram, "Are We in Another AI Bubble?" June 2026, on the enterprise-ROI evidence and the symmetric bull signals. https://www.stockcram.com/blog/are-we-in-another-ai-bubble