One sees stress emerging in credit markets. The other sees the familiar financial cycle around a transformative technology. Both warn that AI’s success may not justify every investment made in its name.
Michael Burry and Ray Dalio are looking at the AI boom from different vantage points, but they arrive at the same central warning: a technology can change the world and still leave its financiers with losses. Burry is watching the loans, bonds and guarantees supporting today’s buildout for signs that lenders are losing confidence. Dalio is looking at the longer historical pattern in which a genuine breakthrough draws in capital, encourages overbuilding and creates debt that eventual returns cannot support.
Neither argument requires AI to fail as a technology. Their difference is less about the destination than about where we are on the road to it. Burry believes the credit market may already be showing the first cracks. Dalio’s framework explains why those cracks could emerge even if AI ultimately delivers substantial productivity gains. The question is whether financial pressure remains a repricing of risk—or begins to stop projects from being built.
What Burry sees in credit
In comments made October 1 and reported the following day, Burry said the infrastructure financing for AI was “crumbling,” pointing to widening credit-default-swap prices and weakening or illiquid loans and bonds. He compared the apparent gap between credit-market concern and equity-market confidence with 2006, before the housing crisis became unmistakable. This is Burry’s reported assessment, not an established diagnosis of the AI sector.
The analogy is useful only if its limits remain clear. Nvidia GPUs are not subprime mortgages, and the largest technology companies are not the weak borrowers at the center of the last credit crisis. They have major operating businesses, substantial cash flow and customers still seeking computing capacity. There is no broad wave of AI-infrastructure defaults or verified collapse in equipment orders.
But Burry’s point is about the order in which a problem might become visible. Equity investors can value a company for the revenue it may produce years from now. Lenders must also ask who makes the next interest payment, what can be recovered from the collateral, how a loan will be refinanced and who actually stands behind a guarantee. If those answers become less convincing, the terms of financing can deteriorate before a company’s share price or equipment orders turn.
That is why the warning deserves testing rather than automatic acceptance or dismissal. Widening spreads, discounted project debt and tougher collateral assumptions would be early evidence. The stronger confirmation would be behavioral: developers postponing viable projects because financing costs rose, lenders requiring enough additional equity to change project economics, or customers reducing equipment commitments. Until that happens broadly, credit repricing is not the same thing as a broken buildout.
What Dalio sees in history
Dalio’s warning begins earlier in the cycle. In a World Economic Forum interview, he argued that transformative technologies tend to inspire investment booms and that the technology’s eventual usefulness does not guarantee good returns for those who financed its expansion. The railroad and internet booms changed economic life while also leaving substantial investor losses. To Dalio, the prospect of AI-driven productivity and the risk of an investment bubble can both be true.
The distinction is between economic value and the price paid to capture it. Companies may eventually use AI to produce more with less. Yet investors can still finance too much capacity, at too high a valuation or with too much debt, before that benefit is reflected in cash flow. The more convinced competitors become that AI will matter, the more urgently they may spend to avoid being left behind. A sound technological thesis can become a poor financing decision if too many projects assume exceptional utilization and returns.
Dalio is not identifying a particular bond as the first domino. His perspective is a longer-cycle question: can the revenues and productivity gains arrive quickly enough, and be distributed widely enough, to justify the capital already committed? Burry’s perspective is narrower and more immediate: are lenders beginning to answer no for some of the assets now being financed?
The evidence between them
In the previous Macro Record article, I described early repricing in several AI-linked financings while emphasizing that the physical buildout had not broadly reversed. Subsequent developments sharpen the question without settling it. Amazon is reportedly exploring a vehicle that would hold roughly $8 billion of Nvidia chips and lease them back. That is a proposed transaction, not evidence that Amazon cannot finance its own spending. It shows how even a strong borrower may seek outside capital for a rapidly expanding asset base.
Nvidia, meanwhile, has announced partnerships intended to mobilize more than $500 billion of third-party capital over time. The company describes the arrangements as financing platforms for AI compute; they remain subject to definitive agreements. The announcement is evidence of an ambition to develop this market, not proof that the full amount has been raised.
The relationships among participants also matter. A Bank for International Settlements study found that 55.2% of incoming investment in AI firms in its 2021–25 sample came from other AI firms. Some investors and recipients also had commercial supply-chain relationships. Such arrangements can help secure capacity and share risk, but they complicate the task of separating independently financed customer demand from demand supported by companies that benefit from the spending. The historical sample is not a measure of today’s entire AI-financing market.
None of these structures is inherently unsound. Project vehicles, leases and guarantees are ordinary ways to finance productive assets. Their terms determine who absorbs a shortfall if a data center is delayed, compute prices fall or chips lose value faster than expected. Complexity does not prove a bust. It does make the buildout harder to judge from headline capital-spending figures alone.
One warning, two clocks
Burry and Dalio therefore reach the same broad conclusion by different routes: the financial returns on AI’s buildout may fail to justify the debt and investment committed to it, even if the technology itself succeeds. Burry believes credit markets may already be registering that danger. Dalio sees a recurring pattern in which the promise of a breakthrough invites more capital than the eventual economics can reward. Neither perspective establishes when, or whether, a sector-wide reversal will occur.
The test is whether changing financial terms begin to alter physical decisions. If spreads stabilize, lenders learn to price chip collateral, and new projects continue to earn acceptable returns, Burry’s near-term reading weakens. If AI produces enough durable revenue to support the assets built for it, Dalio’s bubble warning may prove too severe for this cycle. If instead expensive financing leads repeatedly to postponed data centers, reduced equipment orders and lower expected earnings, the warning will have moved from financial markets into the real economy.
For now, the evidence supports a more limited conclusion: early financing stress, not a mature bust. What matters next is not another prediction about AI’s ultimate importance. It is whether the price of money starts changing what gets built.