From Ritu Raj | Product & Market Analysis

The IMF's 14 AI Deals Rebuilt: $950 Billion Raised, $1.4 Trillion Promised

On this page

The IMF says 14 AI deals since 2023 moved $950 billion of financing in exchange for $1.4 trillion of future revenue commitments. That is $1.47 promised for every dollar raised. The IMF Global Financial Stability Report does not name the 14 deals. We rebuilt the list from company disclosures and could trace $440 billion of the promises to a document with a dollar figure in it.

Key takeaways

  • Every dollar of financing in the IMF's 14 AI deals carries $1.47 of promised revenue. The October 2026 Global Financial Stability Report counts $300 billion of equity and $650 billion of debt against $1.4 trillion of future revenue commitments.
  • Almost half the debt sits with the companies that received the money. The IMF puts $320 billion of the $650 billion with investees, including debt backed by investor guarantees and off-balance-sheet structures. That is 49%.
  • Company disclosures account for about a third of the $1.4 trillion. Four equity-linked contracts carry a dollar value totalling $440 billion. The rest is priced in gigawatts and chip counts, or not priced in public at all.
  • The IMF's correction scenario costs the US 1.8% of output by 2028. It assumes AI productivity gains fail, US equities fall 20% and US investment drops 5%. It is a stress case, not a forecast.

The short answer. The IMF did not call AI a bubble. It counted 14 deals in which investors, led by hyperscalers, financed weaker AI firms that promised to spend the money back with them. The risk it named is wrong-way risk: the stakes lose value exactly when the promised revenue fails.

$1.47Revenue promised per dollar of financing in the IMF's 14 AI deals. Source: Zan Digital calculation from the IMF GFSR, Chapter 1, October 2026.
$320BDebt carried by the AI firms that received the investment, 49% of the total. Source: IMF GFSR, October 2026.
1.8%US output lost by 2028 in the IMF's AI correction scenario. Source: IMF WEO, Box 1.2, October 2026.

Reader for this post: founders and directors, because the question is whether the vendors and platforms you depend on are funded by revenue or by promises, and that is a unit-economics question.

What the IMF global financial stability report counted on AI

The IMF published its Global Financial Stability Report on 13 October 2026. Chapter 1 names an AI-driven market correction as one of three salient risks, next to rising sovereign borrowing costs and a major cyberattack. The AI passage is short, and every word in it carries weight.

The core sentence describes "fourteen publicly disclosed transactions" since 2023. Those transactions account for about $300 billion of aggregate equity investment "led by hyperscalers in financially weaker AI firms." They also involve $650 billion of debt. In return, the investors received "$1.4 trillion in future revenue commitments."

The IMF then adds the number most coverage skipped. About $320 billion of the $650 billion of debt "is taken on by investees in circular financing deals." That includes debt backed by investor guarantees, contingent financing and other off-balance-sheet debt.

What "circular" means in the IMF's usage

AI circular financing, in the IMF's sense, is a loop with two legs. An investor puts equity or credit into an AI firm. The AI firm commits to buy compute, chips or cloud services from that same investor. The investor's revenue forecast and the investee's funding now depend on each other.

The IMF's word for the danger is wrong-way risk. Its report says equity and debt investments in investees depreciate in value precisely when the future revenue streams tied to them are impaired. A bank calls that a loan whose collateral fails when the borrower does. The mechanics of the loop itself are covered in our explainer on how AI circular deals work.

The report does not list the 14 transactions. It does not name OpenAI, Anthropic, xAI, Nvidia, Microsoft or any other party in the text we read. The only selection rule given is "publicly disclosed." That is the gap this post tries to close, and it is also why every reconstruction, including ours, carries uncertainty.

The AI circular financing coverage ratio: $1.47 per dollar

Divide the promises by the money. $1.4 trillion of revenue commitments against $950 billion of financing gives a ratio of 1.47. Each dollar that went in came with a promise of $1.47 coming back as purchases from the investor.

Run the same sum against equity alone and the ratio is 4.67. Each equity dollar was paired with $4.67 of promised spending. That is the figure we would watch, because equity is the first loss when a promise fails.

The IMF's 14 AI deals: what went in, what was promised back $ billions, 14 publicly disclosed transactions since 2023 Equity $300B Investee debt $320B Other debt $330B $950B financing $1,400B Revenue commitments 1.47x Money in Promised back Source: IMF Global Financial Stability Report, October 2026. Ratio and $330B residual: Zan Digital (zandigital.in).
The red block is the part to notice. Nearly half the debt sits with the firms that made the revenue promises, so it fails in the same scenario as the promises do.

Why the ratio above 1 is not automatically bad

A ratio above 1 is normal for a supply contract. A cloud customer commits to spend over five or eight years, and the financing covers only the early build. Your own software contracts work the same way. You commit to a multi-year total and pay as you go.

The question is what funds the gap between $950 billion and $1.4 trillion. In a healthy loop it is the investee's own customer revenue. In a fragile one it is the next round of financing. The IMF number cannot tell you which, and we do not think anyone outside the companies can yet.

Rebuilding the IMF's 14 AI deals from company disclosures

We assembled every publicly disclosed deal since 2023 that pairs an investment with a commitment back to the investor. We used company press releases and filings first, and named press reports where no primary document exists. We found 13 credible candidates. We cannot name a 14th with confidence.

Candidate deals for the IMF's count of 14, with the funding leg and the commitment leg, as of 12 November 2026
#DealFunding legCommitment backFit with IMF definition
1Microsoft to OpenAI$13B funding committed$250B of incremental Azure services (Oct 2025)Clean
2Nvidia to OpenAI$30B equity in the 2026 round, replacing an "up to $100B" letter of intentAt least 10 GW of Nvidia systems (Sept 2025 letter), no dollar figureClean, unpriced
3Amazon to OpenAI$50B, of which $35B is conditional$38B AWS deal (Nov 2025), expanded by $100B over 8 years (Feb 2026)Clean
4AMD and OpenAIWarrant for up to 160M AMD shares, flowing to the buyer6 GW of AMD GPUs, first 1 GW in late 2026Partial
5Oracle and OpenAINo equityAbout $300B over about 5 years (WSJ, Sept 2025)Doubtful, debt-side
6Amazon to Anthropic$8B equity (2023 to 2024)Primary cloud and training partner, no dollar figureClean, unpriced
7Google to AnthropicAbout $3B reported, not confirmed by GoogleUp to 1M TPUs, "tens of billions of dollars" (Oct 2025)Clean, loosely priced
8Microsoft to AnthropicUp to $5B$30B of Azure compute (Nov 2025)Clean
9Nvidia to AnthropicUp to $10BUp to 1 GW on Nvidia systemsClean, unpriced
10Nvidia to xAIAbout $2B reported, plus Series E stakeGPUs bought by a vehicle and leased to xAI for 5 yearsClean, debt-heavy
11Nvidia to CoreWeave$2B of stock (Jan 2026)GPU purchases, no fixed figure. Nvidia also backstops $6.3B of capacityPartial
12CoreWeave and OpenAI$350M of CoreWeave stock to OpenAIContracts worth up to $22.4BPartial
13Meta and Blue Owl, HyperionBlue Owl funds 80% of the vehicle, about $27B of notesMeta leases the site, with a 16-year residual value guaranteeDoubtful, off-balance-sheet

"Up to" figures are caps, not amounts paid. Rows 5, 7 and 10 rest on press reports of company plans rather than primary documents. Row 13 debt size is reported by IFR and Fortune, not stated in Meta's release. Full source list at the end of the post.

The clean fits: hyperscaler equity into a lab that buys back

Seven rows match the IMF wording closely. An investor led by a hyperscaler or chipmaker puts equity into a loss-making lab, and the lab commits to its cloud or chips. Microsoft's October 2025 restructuring attached an incremental $250 billion Azure commitment to its stake in OpenAI. The November 2025 Anthropic deal paired up to $15 billion from Microsoft and Nvidia with $30 billion of Azure compute.

Amazon's February 2026 announcement is the largest single funding leg we can document. It committed $50 billion to OpenAI, $15 billion upfront and $35 billion when certain conditions are met. OpenAI expanded its AWS commitment by $100 billion over 8 years in the same announcement.

The partial fits: equity that flows to the buyer

Three rows run backwards. AMD gave OpenAI a warrant for up to 160 million AMD shares, vesting as OpenAI deploys AMD chips. CoreWeave issued $350 million of stock to OpenAI alongside contracts worth up to $22.4 billion. The supplier hands equity to the customer to win the order.

We think the IMF would count these. They create the same dependency, only with the arrows reversed. A drop in AMD's share price cuts the value of OpenAI's incentive exactly when OpenAI's ability to pay is also in question.

The doubtful fits: debt without an equity leg

Oracle's reported $300 billion contract with OpenAI involves no Oracle equity. Meta's Hyperion vehicle involves no equity in an AI lab. Both matter enormously to the debt total, which is why we list them. Whether they sit inside the IMF's 14 is a judgement we cannot settle from the text. The structure of the Hyperion notes is examined in our breakdown of the Hyperion bond and who holds the risk.

What the disclosures cover, and the $960 billion they do not

Add up the funding legs we can document and you reach about $123 billion. That covers the clean and partial rows, counting every "up to" figure at its cap. It is 41% of the IMF's $300 billion. The other 59% likely sits with co-investors in the same rounds, later tranches we could not verify, or deals priced only in private.

The commitment side is thinner. Only four equity-linked contracts carry a public dollar figure: Microsoft's $250 billion, Amazon's $138 billion across two agreements, Anthropic's $30 billion of Azure and OpenAI's $22.4 billion with CoreWeave. Together they total $440 billion, or 31% of the IMF's $1.4 trillion. Add Oracle's reported $300 billion and the traced share reaches $740 billion, or 53%.

How much of the IMF's total can be traced to a company document $ billions. Dark bar: IMF total. Light bar: what Zan Digital found with a dollar figure attached. Equity, IMF total $300B Equity, documented $123B (41%) Commitments, IMF $1,400B With Oracle added $740B (53%) Equity-linked only $440B (31%) Commitments priced only in gigawatts or chip counts (Nvidia 10 GW, AMD 6 GW, Google TPUs) are excluded. "Up to" amounts counted at their caps. Oracle figure is a WSJ report, not a company disclosure. Source: IMF GFSR October 2026; company releases. Compiled by Zan Digital (zandigital.in).
The gap is the story. The IMF can see the price of deals that companies only describe in gigawatts, so its total is far larger than anything a reader can add up.

Where the missing $960 billion probably sits

The largest unpriced commitments are measured in power, not dollars. OpenAI's letter with Nvidia covers at least 10 gigawatts. The AMD deal covers 6 gigawatts. Anthropic's Google deal covers up to 1 million TPUs. Nvidia's Anthropic deal covers up to 1 gigawatt. None of these has a public dollar value.

The IMF had to price them to reach $1.4 trillion, and it does not show the conversion. That is our main criticism of the report. A headline figure that a reader cannot rebuild is harder to act on, and it hands every side of the bubble argument a number it can interpret as it likes.

Where the $650 billion of AI debt risk sits

The equity in these deals is the visible part. The debt is where the IMF's concern lies, because debt has fixed payments and equity does not. $650 billion of debt against $300 billion of equity means the deals are financed more than twice over by borrowing.

The $320 billion with the investees

The IMF places $320 billion of the debt with the investees. Set that beside the $300 billion of equity they received and the investees have borrowed $1.07 for every equity dollar invested in them. These are the firms the IMF calls financially weaker. They carry the debt and they made the revenue promises.

The IMF's description of that debt matters as much as its size. It includes debt backed by investor guarantees, contingent financing and off-balance-sheet structures. A guarantee means the investor's exposure is larger than its equity stake, though it may not appear on its balance sheet until something goes wrong. The same pattern in the lease market is covered in our analysis of data centre leases that have not yet reached the balance sheet.

The other $330 billion

Subtracting $320 billion from $650 billion leaves $330 billion of debt that is not with the investees. The GFSR text we read does not say where it sits. The obvious candidates are the investors' own bond issuance and project vehicles like Hyperion. The IMF separately notes that hyperscalers' share of investment-grade bond issuance has risen sharply since 2024, which fits that reading.

For a neocloud like CoreWeave, both sides of the loop show up in one capital structure. Its debt is financed against contracts with customers that are themselves funded by the same few investors. We traced that layering in our map of CoreWeave's debt stack and its stress points.

What the IMF AI market correction scenario assumes

The stability report describes the channels. The World Economic Outlook, published the same week, puts numbers on a downside. Box 1.2 of the October 2026 WEO models a market correction in which AI investment unwinds. It is the scenario behind the widely quoted 1.8% US figure.

The scenario assumes that productivity gains from AI investment fail to materialise. US equity prices fall 20% in 2026 and equities elsewhere fall 15%. Financial conditions tighten. US investment falls by 5% from the unwinding of AI spending alone. The shocks then fade over the outlook horizon.

Note what is not assumed. The scenario does not model a default by any of the 14 investees, and it does not model a run on private credit. It is a demand and wealth shock, not a credit event.

By 2028, US output is 1.8% below baseline and world output is 1.1% below, according to the Reuters report of the WEO and the WEO text itself. Asian output falls 1.7%, because the region exports much of the hardware. The unwinding layer on its own cuts global GDP by as much as 0.4%, investment by 1.3% and consumption by 0.6% by 2028.

IMF AI market correction scenario: output below baseline by 2028 Percent of GDP. Scenario, not forecast. October 2026 World Economic Outlook, Box 1.2. United States -1.8% Asia -1.7% World -1.1% World, unwinding only -0.4% Assumes AI productivity gains fail, US equities fall 20% in 2026 and 15% elsewhere, financial conditions tighten. Source: IMF WEO and GFSR, October 2026; Reuters. Compiled by Zan Digital (zandigital.in).
Compare the bottom two bars. Most of the damage comes from the market correction and tighter credit, not from the lost investment itself.

What the scenario tells you about the 14 deals

The gap between 0.4% and 1.1% is the financial amplification. That is exactly the channel the stability report links to circular financing. If equity values fall, the stakes in the 14 deals lose value, investees find refinancing harder and revenue forecasts at the investors fall. The WEO prices the macro outcome. The GFSR explains why the loop can make it worse.

IMF deputy research director Petya Koeva Brooks said the worst case was not likely to reduce growth to levels consistent with a global recession, according to Reuters. We agree with that reading. A 1.8% output loss over two years is a serious slowdown, not a collapse. How this sits against the July warning is set out in our earlier piece on the IMF valuation debate.

Where this reconstruction is weakest

The table above is our list, not the IMF's. The IMF did not publish its 14. If the Fund priced the gigawatt deals differently, or included SoftBank's role in OpenAI or Oracle's debt-funded build, our coverage percentages change. Treat 31% and 41% as a floor on what is disclosed, not as a measure of what the IMF saw.

Our equity total also flatters the documented side. We counted "up to" caps at full value, and $35 billion of Amazon's OpenAI commitment is conditional. If you count only money confirmed as paid, the documented equity is smaller.

There is also a fair case that the coverage ratio overstates risk. Microsoft reported $10.0 billion of revenue from OpenAI in the September 2026 quarter, more than it has ever funded in total, as shown in our map of the OpenAI money flows. Some of the promised revenue is arriving as real cash. A circular loop with paying customers outside it is a supply chain, not a pyramid.

Finally, the IMF itself said core AI firms have been profitable so far. Its warning is about amplification if things go wrong, not a claim that they will. We would not short anything on the strength of this count alone.

Frequently asked questions

What did the IMF global financial stability report say about AI?

The October 2026 Global Financial Stability Report named an AI-driven market correction as one of three salient risks. It counted 14 publicly disclosed circular financing transactions since 2023, involving about $300 billion of equity and $650 billion of debt, in exchange for $1.4 trillion of future revenue commitments. It warned that rising leverage, concentrated exposures and circular financing could make a correction worse.

What is AI circular financing?

AI circular financing is an arrangement in which an investor funds an AI company that then commits to buy the investor's chips, cloud or compute. The investor's revenue depends on the investee's funding, and the investee's funding depends on the investor. The IMF calls the resulting danger wrong-way risk, because the investment loses value exactly when the promised revenue fails to arrive.

How much debt is tied to AI circular deals according to the IMF?

The IMF counts $650 billion of debt across its 14 transactions. About $320 billion of that, or 49%, is taken on by the investees, the financially weaker AI firms that received the investment. That figure includes debt backed by investor guarantees, contingent financing and other off-balance-sheet structures. The report does not specify where the remaining $330 billion sits.

Which companies are in the IMF's 14 AI deals?

The IMF did not name them. The only selection rule it gives is that the transactions were publicly disclosed since 2023. Based on company disclosures, the strongest candidates involve Microsoft, Amazon, Google, Nvidia and AMD as investors, with OpenAI, Anthropic, xAI and CoreWeave as investees. Any list, including ours, is a reconstruction rather than the IMF's own set.

How much would an AI market correction cut US GDP?

In the IMF's October 2026 World Economic Outlook scenario, US output ends 1.8% below baseline by 2028 and world output 1.1% below. The scenario assumes AI productivity gains fail, US equities fall 20% and US investment drops 5%. It is a stress scenario, not a forecast, and the IMF said it would not produce a global recession.

Is the AI bubble real according to the IMF?

The IMF did not call it a bubble. It said core AI firms have been profitable so far and that weaker AI investment could trigger a correction in AI-linked asset prices. Its concern is amplification: leverage, concentration and circular financing could turn a valuation correction into wider financial stress. That is a statement about fragility, not a prediction that prices will fall.

How to use the IMF count this quarter

Pick the two or three AI vendors your business depends on most. For each, find out whether it appears on either side of the table above, as an investor or an investee. A vendor that is an investee carries the debt and the promises. Ask its sales team one question: what share of its compute commitments is funded by customer revenue rather than investor capital?

Then put one date in the calendar. The next GFSR is due in April 2027. If the IMF's count of transactions or the $320 billion of investee debt rises again, treat that as a reason to shorten the length of the contracts you sign with the vendors in the loop.

Related in this series

For the tests that would show the AI cycle turning, read our scorecard of falsifiable AI bubble tests. For how auditors treat revenue from investors, see the auditor tests for circular revenue.

References

  1. International Monetary Fund, Global Financial Stability Report, October 2026, Chapter 1, 13 October 2026. Used for the 14 transactions, $300B, $650B, $1.4T, $320B and wrong-way risk.
  2. International Monetary Fund, Global financial markets face rising risks from AI, bond yields and cyberattacks, IMF Blog, 13 October 2026. Used for the summary of the circular financing finding.
  3. International Monetary Fund, World Economic Outlook, October 2026, Chapter 1, Box 1.2. Used for the market correction scenario assumptions and results.
  4. Reuters via BusinessDay, IMF warns of risks from AI market correction and war, 13 October 2026. Used for the 1.1% and 1.8% figures and the Koeva Brooks comment.
  5. Microsoft, The next chapter of the Microsoft and OpenAI partnership, 28 October 2025, and Microsoft, Nvidia and Anthropic announce strategic partnerships, 18 November 2025.
  6. Amazon, Amazon and OpenAI strategic partnership and investment, 27 February 2026. Used for the $50B and $100B figures.
  7. AMD, AMD and OpenAI announce strategic partnership to deploy 6 gigawatts of AMD GPUs, 6 October 2025.
  8. CoreWeave, CoreWeave expands agreement with OpenAI by up to $6.5B, 25 September 2025. Used for the $22.4B contract total. Nvidia, Anthropic, Meta and press sources for the remaining rows are listed in the table note.

The weakest part of this source base: the IMF does not name its 14 transactions or show how it priced commitments stated in gigawatts, so every deal-level match in this post is our inference. Rows 5, 7, 10 and 13 rely partly on press reports rather than company documents.

RR
Ritu Raj
Writes for Zan Digital about AI market structure, the financing behind the infrastructure build and what company filings say about it.

Related reading