Minsky’s Vision and the Coming AI-Induced Crash

This is the text of a presentation delivered at the Levy Institute’s 40th Anniversary conference (https://www.levyinstitute.org/events/event/levy-economics-institute-anniversary-conference/); a much longer version contains the references: https://www.levyinstitute.org/publications/artificial-intelligence-friend-foe-fraud/

Bill Janeway nicely summarized Minsky’s two price approach used in his financial theory of investment and investment theory of the cycle that led to the well-known financial instability hypothesis. I want to step back and examine the vision of the economy that Minsky developed in his PhD dissertation of 1954 and guided his work up through the 1990s at the Levy Institute, which led to the annual Minsky conference that we are celebrating today.

The main themes are:

• the endogenous instability of market processes;
• rising financial instability over the course of a business cycle—with the possibility of a financial crash and Fisher debt deflation in the more severe downturns;
• the importance of stabilizing institutions to check deviation-amplifying processes;
• and the government as an important stabilizer, with the Fed acting as lender of last resort and the Treasury funding government as employer of last resort.

Whenever a financial crisis hits it is labeled a Minsky Moment, and most have in mind the financial instability hypothesis. But that is simplifying Minsky too much because none of the recent crises has really been about financing investment. Still, they have been consistent with Minsky’s vision of the role played by finance in an inherently unstable and unsustainable system. In his work at Levy, Minsky insisted we entered a new phase—Money Manager Capitalism—that crashed into the Global Financial Crisis (GFC). But the system was rebooted and we now stand on the precipice of another crash, possibly bigger, maybe even the “It” that Minsky warned about for half a century: a depression with a debt deflation.

This time the crash will be triggered by financial practices related to the tech bubble, much of it boosted by AI.

We see all the elements that led to the GFC: tremendous hype, soaring asset prices, and stratospheric debt that is mostly arranged in the shadows of banking with high leverage and little prospect that it can be validated by profit flows.

Let’s look first at the hype. AI forecasts are awe-inspiring. GDP growth rates of 30% per year once AI automates a third of all tasks. Tyler Cowen predicts that AI-run factories will produce goods that are “essentially free.” William Nordhaus believes this will bring on “singularity” when output “becomes infinite.” The economic problem of scarcity will finally be resolved, ending economics as a discipline. Don’t worry about job loss as the displaced workers will become owners. In the very near future, “people’s only source of income will be as rentiers—owners of capital.” True capitalism! The workerless economy is our ultimate destination! Lavish riches are on their way—no retirement saving necessary according to Musk! Keynes’s “Economic Possibilities for our Grandchildren” will be achieved. By 2030!

There are obvious parallels to the dot.com bubble—massive investments in start-ups, highly leveraged finance, booming stock prices, and hyper hype. Most of the dot.coms failed—Schumpeter’s creative destruction—even though the internet really did change life, for better and for worse. The bursting of that bubble left us with massive unused fiberoptic cable that—ironically—we are only now able to use because of the AI bubble.

However, as Robert Gordon famously argued, you can see the evidence of the internet everywhere you look except in the productivity data. The apparent lack of productivity growth is not because the internet failed us but rather because the measure we use requires nominal spending to grow. While there was an initial burst of spending, the simultaneous bursting of the bubbles in dot.coms, housing, and commodities led to slow growth of aggregate demand. That plus negative impacts of the internet on many sectors of the economy (online shopping replaced brick and mortar stores; online work reduced commuting and foot traffic at lunch time; streaming replaced movie night at the cinema, etc.)—all that led to low measured productivity growth.

And this is what the hype about AI boosting growth doesn’t understand. If AI produces infinite abundance, prices fall toward zero, and there’s no GDP or revenue to service debt. Not to mention the wholesale destruction of earth’s environment that makes life possible.

For the sake of argument, let’s assume the internet is our friend—the benefits of living online outweigh the costs of addiction to social media, identity theft and other internet fraud, and loss of in-person interaction. What does AI promise to add to our economy that is already online? First, there is investment in physical infrastructure—largely, building data centers as well as ramping up electricity generation—together that requires spending orders of magnitude greater than the internet boom. AI-related spending accounted for 40% of GDP growth in the first half of 2025. And 80% of the rise of the stock market last year was due to AI-related firms. Projections for the next few years imply that AI will continue to drive markets and the economy.

The claims that AI will replace most workers are surely overstated—but let’s presume there are significant impacts. What will that imply for profits and economic growth? Capitalism is—as explained by Marx, Veblen, Keynes, and Minsky—a monetary production economy where the goal is to end up with more money than you started with. The decision makers are the capitalists—those who have access to money capital. Minsky refines that by adding finance—those positions have to be financed, so the banker is the ephor. Today, that is the shadow banking ephor. Why does this matter? AI hype foresees a future in which either all of us become capitalists, or in which capitalists agree that workers displaced by AI get a share of output without contributing to production.

Let’s address the first scenario: everyone’s a capitalist. Minsky used the Kalecki equation to show that the wage bill in the investment sector generates profits for the consumption sector so long as those workers spend all their income. So, in the simple version, investment creates and is equal to profits. Joan Robinson asked: what happens if robots not only make the consumer goods, but also produce the robots (i.e., investment goods)? I answered that in an article over 30 years ago: Robots that make robots cannot produce any profits because they work for free. Minsky showed that in the expanded Kalecki equation capitalist consumption adds to profits. So, the only profits generated will be from capitalist consumption. If any capitalists decide to save, then profits fall by that amount. This is a strange form of capitalism, in which capitalists live only to consume—not to accumulate money wealth. Profits are maximized by maximizing capitalist consumption.

Turning to the second scenario, either all those who used to be workers starve to death, or we pay them for not working (this is Silicon Valley’s fantasy—the basic income guarantee [BIG]). It isn’t going to happen. The richest sliver already account for half of consumption and they want it all. As Kenneth Boulding explained a long time ago, surveys of the rich consistently show that their greed knows no bounds. But let’s suspend disbelief. As Minsky shows in his macro price equations, competition for consumer goods by those who don’t produce them causes the aggregate price level to rise proportionately—and inflation accelerates as capitalists compete with the unemployed, whose BIG gets bigger. To avoid that, we could just allocate a portion of output to them: the robots produce everything, and then output is divided among the humans according to some rule.

But, in either scenario, if—as Nordhaus says—output becomes infinite, everyone can have an infinite amount of output. There isn’t any reason to own capital (the robots), or to produce profits, or to use money. All we need is warehouses from which consumers order robots to deliver what they want. Apparently, Silicon Valley’s enthusiasts foresee the final destruction of capitalism as we understand it, at the hands of AI. (Surprisingly, that comes close to Schumpeter’s fear: capitalism is doomed because innovation stops when all the entrepreneurs are transformed into mere capitalists.)

Let’s turn to AI Financing. Firms are raising huge amounts of cash with no clear source of revenue to service debt—just like the dot.coms that drove NASDAQ to about 5000 in March 2000, then crashed—losing $5 trillion of value. Half of the dot.com firms failed by 2004. In that boom, 80% of venture investments went to dot.coms; through December 2025, 64% of venture capital went to AI startups. But the AI bubble is 17 times bigger than the dot.com bubble and four times the housing bubble. Nvidia was the first public corporation to reach a market value of $5 trillion. Adding Amazon, Google, Meta, and Open AI brings that to $17 trillion—more than the entire stock market value at the peak of the dot.com bubble.

Altman says he wants to spend trillions; Zuckerberg says hundreds of billions; Apple $500 billion. Where’s all that money coming from? The shadows. We are seeing the same practices that were rampant in financing and manipulation of stock prices in the bubbles that preceded the GFC. AI firms are, again, using “off balance sheet” finance through special purpose vehicles to hide their risk exposure. Risky debt is, again, securitized for sale to unwary pension funds and university endowments. Speculators are, again, using credit default swaps and other derivatives to bet on failure. Chip suppliers are taking positions in the firms that purchase their chips—circular finance. Banks are running on so-called PIK (payment in kind) in which they lend the money needed by debtors to make their payments—a pyramid scheme that Minsky called Ponzi, but PIK sounds fancier.

For example, Nvidia accepts shares of OpenAI as payment and owns about 5% of CoreWeave, whose biggest customer is Microsoft, which is an investor in, and shares revenue with OpenAI, which buys chips from Nvidia and has partnerships with AMD, which is a rival to Nvidia, and has issued warrants for OpenAI to buy 10% of AMD. If you can make sense of that daisy chain of a Gordian knot, you still have working neurons.

In a sign of the times, Alphabet announced it is selling 100-year bonds to finance investment in new data centers! The last time a tech firm did that was Motorola in 1997 during the dot.com bubble. Where’s Motorola now? The problem is that the GPUs —the collateral that stands behind AI debts—have a short life span as faster chips displace them. Furthermore, most of the datacenters are rented, many to startups that are going to fail. No one knows what a used GPU will be worth three years from now, when the lenders might want to collect on their defaulted loan. Those Alphabet bonds could outlive the investments by more than 95 years!

The dot.com investments were like the railroad investments of the 19th century. Cut-throat competition soon eliminated most of them. But the tracks and the locomotives survived—just as the internet infrastructure remained after the GFC. Likewise, creative destruction will eliminate the majority of AI firms, creating excess capacity in data centers. The chips will soon become obsolete, but I guess we can use the warehouses for Trump’s deportation centers. Nobody believes that AI is going to bring massive revenues within the next few years—not even the AI leaders. OpenAI projects losses will grow from $9 billion to $47 billion by 2028. The Techs are already begging for a government backstop against the expected tsunami of defaults. Admittedly, some of the biggest players generate lots of income—the Magnificent Seven are not the equivalent of Pets.com startups, as some have clear revenue sources. The question for them is whether the new earnings will cover the massive costs of investment.

It won’t take a crash for problems to appear—just a revision of expectations and slowing investment will cause problems in the financial sector, as well as in the whole economy, because growth of both has relied to a dangerous extent on AI. Hold on to your hats because the unwinding will be wild, especially with the level of incompetence that rules Washington. 2029 will be an interesting year—maybe as memorable as 1929, potentially with a third term Trump standing in for Hoover.

Let me close with a suggestion: I think most of the debate over AI’s impact on job loss, productivity gains, and abundance is misguided because it presumes that we—workers and employers, doctors and researchers, teachers and students—are the targeted consumers of AI. No. We are the prey. AI will deliver our eyeballs to the ultimate consumers: the scammers, grifters, hackers, advertisers, political parties, Crypto conveyers, and MAGA fascists, or Antifa anarchists. AI has proven useful for kidnappings and assassinations of foreign leaders and for supporting the domestic police state. The rest is mostly hype.