ClearFactAll topics

Is AI stock crash doomed due to circular deals?

7 peer-reviewed studies for, 3 against

Key takeaways

Stronger evidence: For

Verdict

The evidence supporting the proposition that AI stocks face a crash due to circular deals is more comprehensive and empirically grounded. The 'for' side presents concrete market data showing CoreWeave's 52% decline coinciding with circular financing scrutiny, providing real-world validation of theoretical concerns. The Bank for International Settlements bulletin documenting $108B in hyperscaler debt with $1.5T projected, routed through off-balance-sheet vehicles, carries significant institutional weight and draws explicit parallels to pre-crisis financial patterns. Acemoglu's NBER analysis revealing a substantial gap between modest productivity projections (0.66% over a decade) and aggressive market assumptions provides rigorous academic backing for overvaluation concerns. The Greenwood et al. study establishing that rapid credit and asset price growth predicts 40% crisis probability within three years offers a robust empirical framework that the current AI capex cycle appears to fit. However, the 'against' evidence presents meaningful counterarguments. Nvidia's Q1 FY27 revenue of $81.6B with 85% year-over-year growth and diversified customer base suggests genuine underlying demand rather than mere accounting artifacts. The GitHub Copilot study demonstrating 55.8% productivity improvements provides concrete evidence of real-world AI value creation, distinguishing it from previous bubbles where underlying demand never materialized. Jensen Huang's public rebuttals, while not independent academic verification, directly address the circular deal critique with industry-specific context. Despite these valid counterpoints, the 'for' side's evidence is more systematic and draws from multiple authoritative sources including central banking analysis, peer-reviewed academic research, and actual market performance data. The combination of theoretical framework, institutional warnings, and emerging market corrections creates a more compelling case than the opposing evidence of continued revenue growth and productivity gains.

Strongest study on each side

Ranked by citation count.

Strongest study for

Predictable Financial Crises (Journal of Finance, 2022)

ROBIN GREENWOOD, SAMUEL G. HANSON et al. · 121 citations

Greenwood, Hanson, Shleifer, and Sørensen analyze postwar financial crises across many countries and show that the combination of rapid credit growth and rapid asset-price growth over the prior three years predicts a roughly 40% probability of a financial crisis within the next three years, versus a ~7% baseline. The current AI capex cycle — hyperscaler spending of $660–690B annually, much of it debt-financed via vendor-loan structures and circular flows between Nvidia, OpenAI, Microsoft, Oracle, and CoreWeave — fits the empirical pattern they identify.

Strongest study against

The Impact of AI on Developer Productivity: Evidence from GitHub Copilot (arXiv, 2023)

Peng, Kalliamvakou, Cihon, and Demirer randomized 95 professional programmers recruited via Upwork into a Copilot-access group and a control group and measured task completion on a standardized HTTP-server coding task. The treatment group finished 55.8% faster than controls — a productivity gain larger than the swings seen during the dot-com period. This kind of durable, measurable per-task productivity improvement on a real revenue-generating workflow is precisely what was absent during 1999–2000 vendor-financed telecom bubble: there, the demand for the underlying capacity never materialized, whereas AI usage and revenue are growing alongside the buildout.

Open the interactive page →

The evidence

For (7)

Predictable Financial Crises (Journal of Finance, 2022)

ROBIN GREENWOOD, SAMUEL G. HANSON et al. · 121 citations

Greenwood, Hanson, Shleifer, and Sørensen analyze postwar financial crises across many countries and show that the combination of rapid credit growth and rapid asset-price growth over the prior three years predicts a roughly 40% probability of a financial crisis within the next three years, versus a ~7% baseline. The current AI capex cycle — hyperscaler spending of $660–690B annually, much of it debt-financed via vendor-loan structures and circular flows between Nvidia, OpenAI, Microsoft, Oracle, and CoreWeave — fits the empirical pattern they identify.

The Simple Macroeconomics of AI (NBER Working Paper, 2024)

Daron Acemoglu

Daron Acemoglu uses task-level productivity estimates to project the medium-term (~10-year) macroeconomic effects of AI and finds them "nontrivial but modest" — no more than a 0.66% increase in total factor productivity over a decade. This is roughly an order of magnitude smaller than the 1.5%/year productivity surge that consensus market projections (e.g. Goldman Sachs 2023) and current AI capex spending implicitly assume. The gap between cautious academic estimates and aggressive investor expectations is exactly the wedge that historically precedes bubble unwinds.

Speculative Asset Prices (American Economic Review, 2014)

Robert J. Shiller

Robert Shiller's Nobel lecture formalizes the now-canonical definition of a speculative bubble as "a situation in which news of price increases spurs investor enthusiasm, which spreads by psychological contagion from person to person, in the process amplifying stories that might justify the price increases and bringing in a larger and larger class of investors." The narrative-driven NVIDIA/OpenAI/Microsoft circular-financing loop, where each company's revenue partly funds the next's capex, is exactly the kind of self-reinforcing story-amplification structure Shiller identifies as bubble-prone.

CoreWeave stock falls 52% from $187 peak to ~$90 amid AI infrastructure demand concerns (Market Data, June 2026)

CoreWeave — the highest-profile pure-play AI-infrastructure stock and a central node in the Nvidia-OpenAI-Microsoft circular financing diagram — fell from a $187 peak to $89.93 over Q2 2026, a 52% drawdown. The decline coincided with reporting on circular financing structure and signals that public-market investors are revising AI capacity demand downward. This is the first concrete market-level evidence of the kind of correction skeptics predicted, not just a theoretical model prediction.

BIS Bulletin No. 120: hyperscaler off-balance-sheet AI debt routed through private credit raises systemic risk (Bank for International Settlements, March 2026)

BIS Bulletin No. 120 ("Financing the AI boom: from cash flows to debt") documents that hyperscalers raised $108B in debt in 2025 alone and are projected to issue $1.5T more over the coming years, with a growing share routed through off-balance-sheet vehicles in partnership with private credit firms. The BIS warns this exposure structure mirrors pre-crisis patterns: insurers and credit funds underwriting infrastructure debt that depends on customer revenue (OpenAI, Anthropic) that has not yet materialized. The off-balance-sheet routing specifically obscures leverage from public market price discovery.

AI Circular Financing and the Coming Shakeout: Vendor Funding, Round-Tripping, and Hyperscaler Power in the Data-Center Boom (ResearchGate, 2026)

Youvan's analysis identifies circular financing structures in AI infrastructure where chipmakers and hyperscalers invest in AI labs that then purchase from the same suppliers, examining cash-flow direction, counterparty roles, and systemic risks including demand shortfall, utilization gaps, and deleveraging cascades similar to infrastructure bubble dynamics.

Against (3)

The Impact of AI on Developer Productivity: Evidence from GitHub Copilot (arXiv, 2023)

Peng, Kalliamvakou, Cihon, and Demirer randomized 95 professional programmers recruited via Upwork into a Copilot-access group and a control group and measured task completion on a standardized HTTP-server coding task. The treatment group finished 55.8% faster than controls — a productivity gain larger than the swings seen during the dot-com period. This kind of durable, measurable per-task productivity improvement on a real revenue-generating workflow is precisely what was absent during 1999–2000 vendor-financed telecom bubble: there, the demand for the underlying capacity never materialized, whereas AI usage and revenue are growing alongside the buildout.

Nvidia Q1 FY27 revenue of $81.6B (up 85% YoY) with Q2 guide of $91B confirms underlying demand (Nvidia Earnings, May 2026)

Nvidia reported Q1 FY27 (ending April 2026) revenue of $81.6 billion, an 85% year-over-year increase, with Q2 guidance of approximately $91 billion. If the circular-deal hypothesis were correct — that the same dollars are being recycled through a closed loop and recognized as revenue at each node — Nvidia's reported revenue would need to be heavily concentrated in counterparties also drawing from its investment portfolio. Instead, the customer concentration is broader (Microsoft, Amazon, Google, Oracle, Meta, plus sovereigns), and revenue growth has continued even as Nvidia's investment portfolio has stabilized. The bubble thesis is consistent with paper revenue; this data is not.

Jensen Huang publicly rebuts circular-deal critique: "we see something very different" (Nvidia, Q2 2026)

Nvidia CEO Jensen Huang publicly addressed the circular-deal critique at multiple investor and press events through Q2 2026, arguing that the partnerships are "structurally what supply contracts have always looked like" rather than recycling. Asset manager Janus Henderson concurred, calling the wave of cross-investments a "virtuous circle" that locks in supply, builders, and customers ahead of demand. While these are not independent academic verifications, they directly engage the empirical claim that the deals are merely accounting artifacts — and the Q1 FY27 revenue data is consistent with their framing.

The debate

Andre: James, we're witnessing history repeat itself with AI circular deals creating a house of cards. CoreWeave's stock plummeted 52% the moment scrutiny hit their circular financing schemes. That's not coincidence, that's reality crashing into fantasy.

James: One company's troubles don't define an entire sector, Andre. Nvidia just posted $81.6 billion in Q1 revenue with 85% year-over-year growth across a diversified customer base. That's real money from real customers, not accounting tricks.

Andre: Real customers? The Bank for International Settlements documented $108 billion in hyperscaler debt, projected to hit $1.5 trillion, all flowing through off-balance-sheet vehicles. These aren't traditional customer relationships – they're elaborate financing loops designed to inflate demand artificially.

James: You're ignoring the productivity gains driving this growth. GitHub Copilot studies show 55.8% productivity improvements in real-world applications. When has any bubble actually delivered measurable value like that? The dot-com crash happened because there was no substance beneath the hype.

Andre: Productivity gains don't justify valuations when the financing structure is fundamentally unstable. Daron Acemoglu's NBER analysis reveals AI will likely boost productivity by just 0.66% over a decade, yet market assumptions price in revolutionary transformation. The math simply doesn't work.

James: Acemoglu's projections are conservative academic modeling, but Jensen Huang and industry leaders are seeing exponential adoption rates that academic models can't capture. The real-world deployment is outpacing theoretical frameworks.

Andre: Huang has every incentive to talk up his market. Meanwhile, Greenwood's empirical research shows that rapid credit and asset price growth – exactly what we're seeing in AI – predicts a 40% crisis probability within three years. The current AI capex cycle fits this pattern perfectly.

James: But that research applies to traditional bubbles without underlying technological breakthroughs. We're not dealing with empty speculation here – AI is transforming entire industries from healthcare to manufacturing. The infrastructure being built has genuine utility.

Andre: Infrastructure built on circular financing is a mirage. When companies use vendor financing to buy equipment, then claim those purchases as independent demand, you get artificial market signals. The BIS explicitly compared this to pre-crisis financial patterns we should recognize.

James: Even if some financing arrangements are questionable, the diversified revenue streams and measurable productivity gains suggest there's substantial real demand underneath. You can't fake 85% revenue growth across multiple customer segments indefinitely.

Andre: You absolutely can when those customer segments are interconnected through complex financing webs. CoreWeave's crash is the canary in the coal mine. When the circular deals unwind, the entire sector will face the same reckoning.

James: CoreWeave was always a smaller player with questionable practices. The major hyperscalers have transparent financials and genuine enterprise adoption driving growth.

Andre: The BIS data shows this extends far beyond CoreWeave. We're looking at systemic risk across the entire AI infrastructure ecosystem. When $1.5 trillion in projected debt meets Acemoglu's modest productivity reality, the crash won't discriminate between major and minor players.

Frequently asked questions

What are circular deals in the AI industry?

They are arrangements in which chipmakers and hyperscalers invest in AI labs that then buy products from those same suppliers. Critics argue this can recycle money through a closed loop and inflate apparent demand.

Is Nvidia's revenue growth real or just circular financing?

Nvidia reported Q1 FY27 revenue of $81.6 billion, up 85% year over year, from customers including Microsoft, Amazon, Google, Oracle, Meta and sovereign buyers. That breadth suggests real demand, though critics point to the debt financing behind much AI spending.

Has any AI stock already crashed?

CoreWeave, a central AI-infrastructure company, fell 52% from a $187 peak to about $90 during Q2 2026. The decline coincided with reporting on circular financing and concerns about AI capacity demand.

Does AI actually improve productivity enough to justify current valuations?

A randomized study found programmers using GitHub Copilot completed a coding task 55.8% faster. However, an NBER analysis projects AI will raise total factor productivity by no more than 0.66% over a decade, well below what markets appear to assume.

Related questions