Investors can already bet against the AI boom. The harder question is whether the money being spent will earn its way back.
By Jeff FergusonOctober 27, 20268 minute read
In The Big Short, Michael Burry sees trouble in the mortgage market and goes looking for a way to profit from it. Recognizing the problem is only part of the job. He also needs a financial instrument that will pay off if his analysis is right.
Burry, played by Christian Bale in the film adapted from Michael Lewis’s book, works with Wall Street banks to arrange that trade. Watching the money pouring into artificial intelligence, I started wondering how much of that machinery already exists for an AI bust.
Quite a bit, as it turns out. There is even an institutional product built around the same broad type of derivative Burry used.
That deserves attention. It also requires care. The existence of a way to bet against a boom tells us that people want the exposure. It cannot tell us, by itself, whether the boom will end badly or when.
The question that interests me is whether AI can become a durable part of the economy while the investment boom around it ends in substantial losses. I think that is a plausible outcome. A technology can be useful without producing enough profit to justify everything investors have paid for it.
Burry’s original trade helps explain the distinction. In his own account at Vanderbilt University, he described buying credit-default swaps on selected portions of securities backed by subprime mortgages. His first trades were arranged with Deutsche Bank in 2005; Goldman Sachs and other banks also became counterparties. He wanted contracts he could trade between dealers, rather than being stuck with an agreement only one bank could price.
These swaps transferred specified credit risk. Burry paid to hold the protection, which could become more valuable as the underlying mortgage securities deteriorated. He was making a particular claim about particular debts, rather than simply announcing that housing looked expensive.
That level of specificity is useful when thinking about AI. A falling share price, an unprofitable business and a borrower unable to pay its debts are related possibilities. They are not interchangeable.
That is the amount of debt exposure referenced by the trade, not a $25 million upfront fee. It is an institutional arrangement tied to the credit of five large companies, each with businesses extending beyond AI. It is nevertheless a concrete example of Wall Street packaging a way to hedge the financing behind the buildout.
A credit-default swap lets a protection buyer pay a seller to take on defined credit risk. A specified credit event, such as failure to pay, can trigger settlement. The protection can also gain value before any default, allowing a buyer to potentially realize a gain by closing the position. Falling stock prices alone do not trigger a credit-event payout.
Stock-market bets are available through more familiar instruments. Investors can buy puts, short shares or use funds designed to move opposite a benchmark. ProShares’ PSQ targets the inverse of the Nasdaq-100’s daily return. Direxion’s SOXS targets three times the inverse of a semiconductor index’s daily return. Neither is a pure bet on AI, and the daily objective matters: a months-long holding does not simply produce the stated inverse of the market’s cumulative move.
Burry has also returned to the story. Reporting on his October 11 update said he had concentrated much of his bearish exposure in puts on the Nasdaq-100 and the VanEck Semiconductor ETF. That is a reported position at a particular date. His reputation adds interest, but it cannot substitute for examining the current evidence.
For EconRisk, the more consequential development is what the credit market is charging for protection.
Reuters reported on October 15 that CoreWeave’s five-year CDS spread had exceeded 1,100 basis points, while Oracle’s approached 300. Those were observations from that day, not live prices. They show investors assigning very different levels of risk to different participants in the AI economy.
A higher quoted spread means more expensive credit protection. It does not automatically mean a company is about to default. Heavy issuance, demand for hedges and uncertainty about future financing can all affect the price. J.P. Morgan’s October 23 assessment emphasized that distinction: protection on the strongest cash-generating companies still looked very different from protection on borrowers dependent on continued financing.
We should preserve that distinction. Putting every company with an AI connection into one bucket makes the story easier to tell and the financial analysis less useful.
Consider a hypothetical data-center operator that borrows to build capacity on the assumption that customers will pay attractive prices for years. Demand could grow and the project could still disappoint. Construction might cost more than expected. Competition might push prices down. The equipment might need replacement sooner. A large customer might buy less capacity than anticipated.
The technology would still work. The cash coming in might simply be insufficient for the costs and financing commitments already made.
The Bank for International Settlements has described another vulnerability: financial connections between chip suppliers, cloud providers, AI developers and infrastructure companies. Some suppliers invest in businesses that, in turn, commit to buying their products or services. Such arrangements can support real expansion. They also connect investment values, expected sales and financing in ways that can transmit problems between companies.
The relevant test is what happens when independently paying customers do not arrive at the pace assumed in those plans. We need to know whether contracts turn into cash, whether margins cover costs, and whether a borrower can withstand delays without repeatedly raising new money.
This is why I keep coming back to the dot-com comparison.
The internet did not disappear when the bubble burst. Its eventual importance did not protect every company built around it or every investor who bought into it. Continued adoption and poor investment returns can occur together.
In a September speech, the BIS placed today’s AI spending alongside earlier booms in canals, railways, electrification and the internet. These involved consequential technologies. They also attracted investment that eventual returns could not fully justify.
AI could follow a version of that pattern. Companies could continue using it to write software, analyze information and perform other work, while some suppliers fail and investors mark down the value of businesses that survive. Cheaper computing could even benefit customers while hurting the owners of equipment bought at boom-era prices. That is a possible scenario, not an established forecast.
There are reasons the outcome could be less damaging. Several major participants have substantial businesses and cash flows outside their newest AI projects. Revenue might grow into today’s investment. Spending could slow before debt becomes unmanageable. We need to take those possibilities seriously, too.
The IMF’s October financial stability assessment recognizes strong corporate earnings while identifying an AI-related correction as a risk that leverage and concentrated exposures could amplify. That is a more defensible position than assuming either permanent prosperity or an unavoidable collapse.
I also would not assume that a technology-led bust must be safely contained. A stock-market repricing could leave many operating businesses intact. A spending slowdown could hurt construction firms, equipment suppliers and local economies. Defaults could carry losses into lenders and investment funds. These outcomes can overlap, and the extent of the damage depends on where the obligations sit and who can absorb them.
The dot-com analogy helps explain how a useful technology can support a bad investment. The housing-crisis analogy directs our attention to the financing and the consequences of connected losses. Neither gives us a ready-made forecast.
For someone tempted to turn that forecast into a trade, there is another part of The Big Short worth remembering: waiting was costly.
Burry’s account of the period before the payoff includes angry investors, disputed valuations and pressure to liquidate. Knowing how the story ended makes it easy to underestimate how difficult the position was to maintain.
Being right about a vulnerability does not guarantee that a particular trade will make money. The instrument has to match the event, the timing and the price paid.
Our job at EconRisk is to examine the vulnerability. I would put more weight on a combination of worsening financing conditions, weak cash generation and spreading losses than on a famous investor’s prediction. I would also look for evidence that challenges the bearish case: customers renewing at profitable prices, cash flow improving, and businesses financing expansion without increasing financial strain.
The same discipline applies to rising demand for protection. An investor may be insuring a large existing investment rather than expecting the entire sector to fail. We should compare AI-related credit conditions with the wider market and look for confirmation in company results.
The Big Short makes financial skepticism feel cinematic. Its more useful lesson is how much depends on examining the obligations underneath a persuasive story.
AI may become part of everyday economic life, just as the internet did. Every company financing that future still has to meet its own costs, and every borrower still has to pay. A convincing demonstration of what the technology can do will not settle those bills.
Commentary reflects the author’s interpretation of the evidence. AI assisted the research and drafting. Sources are linked throughout. Read our AI disclosure.
Commentary · AI Investment & Credit
AI Already Has Its Big Short
Investors can already bet against the AI boom. The harder question is whether the money being spent will earn its way back.
In The Big Short, Michael Burry sees trouble in the mortgage market and goes looking for a way to profit from it. Recognizing the problem is only part of the job. He also needs a financial instrument that will pay off if his analysis is right.
Burry, played by Christian Bale in the film adapted from Michael Lewis’s book, works with Wall Street banks to arrange that trade. Watching the money pouring into artificial intelligence, I started wondering how much of that machinery already exists for an AI bust.
Quite a bit, as it turns out. There is even an institutional product built around the same broad type of derivative Burry used.
That deserves attention. It also requires care. The existence of a way to bet against a boom tells us that people want the exposure. It cannot tell us, by itself, whether the boom will end badly or when.
The question that interests me is whether AI can become a durable part of the economy while the investment boom around it ends in substantial losses. I think that is a plausible outcome. A technology can be useful without producing enough profit to justify everything investors have paid for it.
Burry’s original trade helps explain the distinction. In his own account at Vanderbilt University, he described buying credit-default swaps on selected portions of securities backed by subprime mortgages. His first trades were arranged with Deutsche Bank in 2005; Goldman Sachs and other banks also became counterparties. He wanted contracts he could trade between dealers, rather than being stuck with an agreement only one bank could price.
These swaps transferred specified credit risk. Burry paid to hold the protection, which could become more valuable as the underlying mortgage securities deteriorated. He was making a particular claim about particular debts, rather than simply announcing that housing looked expensive.
That level of specificity is useful when thinking about AI. A falling share price, an unprofitable business and a borrower unable to pay its debts are related possibilities. They are not interchangeable.
In March, Bloomberg reported that JPMorgan had introduced a basket of credit-default swaps covering Alphabet, Amazon, Meta, Microsoft and Oracle. The report, citing people familiar with the arrangement, said trades occurred in $25 million increments, with $5 million of exposure to each company.
That is the amount of debt exposure referenced by the trade, not a $25 million upfront fee. It is an institutional arrangement tied to the credit of five large companies, each with businesses extending beyond AI. It is nevertheless a concrete example of Wall Street packaging a way to hedge the financing behind the buildout.
A credit-default swap lets a protection buyer pay a seller to take on defined credit risk. A specified credit event, such as failure to pay, can trigger settlement. The protection can also gain value before any default, allowing a buyer to potentially realize a gain by closing the position. Falling stock prices alone do not trigger a credit-event payout.
Stock-market bets are available through more familiar instruments. Investors can buy puts, short shares or use funds designed to move opposite a benchmark. ProShares’ PSQ targets the inverse of the Nasdaq-100’s daily return. Direxion’s SOXS targets three times the inverse of a semiconductor index’s daily return. Neither is a pure bet on AI, and the daily objective matters: a months-long holding does not simply produce the stated inverse of the market’s cumulative move.
Burry has also returned to the story. Reporting on his October 11 update said he had concentrated much of his bearish exposure in puts on the Nasdaq-100 and the VanEck Semiconductor ETF. That is a reported position at a particular date. His reputation adds interest, but it cannot substitute for examining the current evidence.
For EconRisk, the more consequential development is what the credit market is charging for protection.
Reuters reported on October 15 that CoreWeave’s five-year CDS spread had exceeded 1,100 basis points, while Oracle’s approached 300. Those were observations from that day, not live prices. They show investors assigning very different levels of risk to different participants in the AI economy.
A higher quoted spread means more expensive credit protection. It does not automatically mean a company is about to default. Heavy issuance, demand for hedges and uncertainty about future financing can all affect the price. J.P. Morgan’s October 23 assessment emphasized that distinction: protection on the strongest cash-generating companies still looked very different from protection on borrowers dependent on continued financing.
We should preserve that distinction. Putting every company with an AI connection into one bucket makes the story easier to tell and the financial analysis less useful.
Consider a hypothetical data-center operator that borrows to build capacity on the assumption that customers will pay attractive prices for years. Demand could grow and the project could still disappoint. Construction might cost more than expected. Competition might push prices down. The equipment might need replacement sooner. A large customer might buy less capacity than anticipated.
The technology would still work. The cash coming in might simply be insufficient for the costs and financing commitments already made.
The Bank for International Settlements has described another vulnerability: financial connections between chip suppliers, cloud providers, AI developers and infrastructure companies. Some suppliers invest in businesses that, in turn, commit to buying their products or services. Such arrangements can support real expansion. They also connect investment values, expected sales and financing in ways that can transmit problems between companies.
The relevant test is what happens when independently paying customers do not arrive at the pace assumed in those plans. We need to know whether contracts turn into cash, whether margins cover costs, and whether a borrower can withstand delays without repeatedly raising new money.
This is why I keep coming back to the dot-com comparison.
The internet did not disappear when the bubble burst. Its eventual importance did not protect every company built around it or every investor who bought into it. Continued adoption and poor investment returns can occur together.
In a September speech, the BIS placed today’s AI spending alongside earlier booms in canals, railways, electrification and the internet. These involved consequential technologies. They also attracted investment that eventual returns could not fully justify.
AI could follow a version of that pattern. Companies could continue using it to write software, analyze information and perform other work, while some suppliers fail and investors mark down the value of businesses that survive. Cheaper computing could even benefit customers while hurting the owners of equipment bought at boom-era prices. That is a possible scenario, not an established forecast.
There are reasons the outcome could be less damaging. Several major participants have substantial businesses and cash flows outside their newest AI projects. Revenue might grow into today’s investment. Spending could slow before debt becomes unmanageable. We need to take those possibilities seriously, too.
The IMF’s October financial stability assessment recognizes strong corporate earnings while identifying an AI-related correction as a risk that leverage and concentrated exposures could amplify. That is a more defensible position than assuming either permanent prosperity or an unavoidable collapse.
I also would not assume that a technology-led bust must be safely contained. A stock-market repricing could leave many operating businesses intact. A spending slowdown could hurt construction firms, equipment suppliers and local economies. Defaults could carry losses into lenders and investment funds. These outcomes can overlap, and the extent of the damage depends on where the obligations sit and who can absorb them.
The dot-com analogy helps explain how a useful technology can support a bad investment. The housing-crisis analogy directs our attention to the financing and the consequences of connected losses. Neither gives us a ready-made forecast.
For someone tempted to turn that forecast into a trade, there is another part of The Big Short worth remembering: waiting was costly.
Burry’s account of the period before the payoff includes angry investors, disputed valuations and pressure to liquidate. Knowing how the story ended makes it easy to underestimate how difficult the position was to maintain.
Today’s instruments have their own constraints. A purchased put can expire worthless before the expected decline arrives. Short-selling shares can generate theoretically unlimited losses if the price keeps rising. Credit protection also has a cost, and a well-publicized concern may already be expensive to hedge.
Being right about a vulnerability does not guarantee that a particular trade will make money. The instrument has to match the event, the timing and the price paid.
Our job at EconRisk is to examine the vulnerability. I would put more weight on a combination of worsening financing conditions, weak cash generation and spreading losses than on a famous investor’s prediction. I would also look for evidence that challenges the bearish case: customers renewing at profitable prices, cash flow improving, and businesses financing expansion without increasing financial strain.
The same discipline applies to rising demand for protection. An investor may be insuring a large existing investment rather than expecting the entire sector to fail. We should compare AI-related credit conditions with the wider market and look for confirmation in company results.
The Big Short makes financial skepticism feel cinematic. Its more useful lesson is how much depends on examining the obligations underneath a persuasive story.
AI may become part of everyday economic life, just as the internet did. Every company financing that future still has to meet its own costs, and every borrower still has to pay. A convincing demonstration of what the technology can do will not settle those bills.
← More commentary