The industry is badly overleveraged and that's unlikely to change.
The AI boom is already the largest capital investment project in history, running a price tag greater than the Manhattan Project, Interstate Highway System, and Apollo program combined. Estimates of the total cost of the data center buildout run as high as $7 trillion.
The issue is not whether AI will transform our economy and society—maybe it will, maybe it won’t. But the conversation around AI’s long-term impact misses the more urgent question: whether AI companies can increase revenues and profits fast enough to pay their bills.
For all the excitement around the technology, AI companies are bringing in only a small fraction of the revenues needed to cover their spending—and the gap is widening, not narrowing. That has forced AI companies to turn to skyrocketing amounts of debt to pay for the chips and data centers needed to train and run generative AI models, including $570 billion in new debt this year alone. Those numbers—irrespective of any capabilities in the underlying technology—suggest a real risk of a financial crisis.
And the stakes there go well beyond a particular industry. AI-linked companies dominate stock and credit markets in a way no industry has since the Gilded Age. Unless the AI industry quickly finds a way to produce a return on its debt-fueled spending splurge, the industry’s increasingly fragile economics could unravel—and drag the rest of the financial system and even the economy down with it.
To properly understand the precarity of the AI companies, it is important to realize that their margins are being squeezed from several directions at once. For starters, the market for generative AI models is fiercely competitive. The leading developers—Alphabet, Anthropic, OpenAI, Meta, Microsoft, and xAI (now a part of SpaceX)—are among the world’s most valuable companies, and each offers models that are very similar in their capabilities and quality. Without the protective “moat” of patent protection, as is common with new products, each company operates with the knowledge that windows of opportunity to build a lead over its rivals open rarely and close quickly. Whenever an AI developer discovers a use case with even a whiff of profit potential, such as coding or military applications, its rivals quickly jump in to offer competing products.
Such dynamics are often a recipe for price wars that make profits difficult-to-impossible to generate. Sure enough, The Wall Street Journal and Bloomberg reported last month that OpenAI and Anthropic were preparing to cut prices in an effort to win customers ahead of their respective IPOs—a development that was largely overshadowed by SpaceX’s record-shattering stock market debut. In the weeks since, Silicon Valley’s AI heavyweights have started offering discounts of as much as 75% for access to their models.
Launching such a price war is, in effect, a bet that the fevered competition that follows will be temporary and that the survivor(s) will establish market dominance and rake in monopoly profits. That is, after all, what happened with Internet search (Google), cloud computing (Amazon and Microsoft), smartphone apps (Apple and Google), and social media (Facebook), among others. Placing a similar all-or-nothing bet on dominating the generative AI market might make sense if intense competition were the only obstacle to generative AI’s profitability. But there are other factors in play here.
Training and running generative AI models is enormously expensive, and it is not clear if (much less when) it will get cheaper. To win market share, developers have long offered access to their models below cost, but this has meant that AI companies lose money on virtually all of their users, including those with premium subscriptions. As a result, OpenAI and Anthropic, despite their lofty valuations, have burned through billions of dollars over the course of their existence. Thus far, they have stayed afloat because, for all of their struggles to earn money, they have had no trouble raising it from investors.
But these financial pressures will be unceasing. The physical assets at the center of previous private-sector infrastructure booms, like railroads and high-speed internet, have a useful life measured in decades. AI chips have a useful life of at most half a decade, and the data centers that house them are built for specific generations of chips and cannot easily be repurposed. These chips and data centers are tied to trillions of dollars in lease commitments, construction loans, and more complex forms of debt that drive up AI companies’ costs—even as their customers are becoming more cost-conscious.
Businesses have started to balk at AI companies’ efforts to juice more revenue out of them. In the early part of this year, model developers began moving from flat-fee to usage-based pricing for access to their flagship models, effectively raising prices for most enterprise customers. The move apparently succeeded in boosting revenues, at least temporarily, but the higher prices gave many customers sticker shock. One unnamed company reportedly racked up a $500 million Claude bill.
Unsurprisingly, many businesses are starting to question whether they are getting enough return to justify the technology’s costs, with AI FOMO giving way to economic reality and Uber’s Chief Operating Officer saying the company simply isn’t getting enough bang for the buck with its AI usage. Uber, Walmart, and other companies have responded to cost overruns not by increasing their AI budgets, but by tightening their limits on employees’ AI use, switching to open-source Chinese models, or even, like Microsoft, canceling third-party AI licenses altogether.
This summer’s price cuts may help stem some customers’ concerns, but it is unusual for companies to slash prices so soon after raising them. Perhaps Anthropic and OpenAI are simply struggling to find a sustainable price point for their products given the economic wind shear the industry is facing. Or perhaps the price hikes that preceded them were prompted by a desire to engineer a misleading one-time spike in revenues for the sake of the pre-IPO financial disclosures the leading model developers will soon have to file.
Either way, the industry’s brutal economics are not likely to improve anytime soon. Studies and surveys repeatedly show that the vast majority of businesses are not seeing any return on their investments in generative AI projects. Competition, both among the big U.S. companies and with their open-source Chinese rivals, will continue to be intense. Those forces will apply unrelenting downward pressure on prices even as high costs squeeze companies from the opposite direction. And the lease commitments for data centers remain, at least for the foreseeable future, a vast and immovable fixed cost.
If these combined pressures culminate in an AI market collapse, the fallout would likely be severe. AI-focused companies now comprise 45% of S&P 500 market cap and loans to the AI industry have become “all-encompassing” for credit markets. “The numbers are like nothing any of us who have been in this business for 25 years have seen,” a senior executive at Bank of America Corp told Bloomberg in reference to AI’s domination of available lending. The last time an industry held such sway over the nation’s economy was the 19th century, when railroads reigned supreme. When the railroad bubble burst in 1873, it led to the longest depression in U.S. history.
We are deeply familiar with the United States’ recent financial crises—one of us was Vice Chair of the Congressional Oversight Panel for TARP, the 2008 bank bailout. Certain patterns are noticeable. Before crises happen, those whose behavior brings them on always deny anything terrible is possible. When crisis nonetheless comes, the same people deny anyone could have foreseen it, and the general public suffers in ways that they claimed were unimaginable. Then, of course, the culprits demand bailouts, and the public’s confidence in democracy collapses.
Financial crises often unfold in fits and starts, and there is typically a lag before a crash spirals into a calamity. The subprime mortgage bubble burst in the summer of 2007, more than a year before it morphed into the worst financial and economic crisis in 80 years. By the time that full-blown financial meltdown hit, a number of financial institutions had already failed, and the situation was well beyond the ability of ordinary regulatory tools to fix. Fixes need to occur in advance of a crash, not in emergency fashion like in 2008.
It is not too late for regulators to act to protect the U.S. economy against the risk of the AI bubble bursting. It is beyond the power of any regulator to stop people who want to indulge in magical thinking from doing so, but thanks to legislation passed after 2008, regulators have tools to stop such delusions from endangering our entire economy and society, including through stronger oversight of AI-linked companies, financial auditing by the Public Company Accounting Oversight Board, inquiring into banks’ exposure to an AI meltdown, and conducting stress tests to see how balance sheets would respond if data center debt became toxic.
Regulators need to use those tools now to alert market participants to AI-related risks and prevent those risks from concentrating further. Additional legislation may ultimately be needed to stop a crash from spiraling and certainly to prevent it from recurring, such as stronger barriers between private credit markets and commercial banks and stronger measures to deter the types of financial chicanery that AI companies have used to create a misleading picture of the industry’s finances.
As it stands, however, the risks are increasing every day. Unless either regulators act or the AI industry can find a way to pay its fast-mounting bills, the future could well be bleak—not just for the AI industry, but for the financial system and, from there, the economy as a whole.
Damon Silvers is Visiting Professor of Practice at University College London’s Institute for Innovation and Public Purpose.
Matthew Scherer is a fellow at the Open Markets Institute and columnist for Hard Reset Media.
Follow Persuasion on X, Instagram, LinkedIn, and YouTube to keep up with our latest articles, podcasts, and events, as well as updates from excellent writers across our network.
And, to receive pieces like this in your inbox and support our work, subscribe below:
Counterpoint: The article identifies real risks, but it repeatedly jumps from “AI economics are challenging” to “a systemic financial crisis is likely.” That’s a much bigger claim than the evidence supports.
First, falling token prices are treated as proof AI is becoming uneconomic. History suggests the opposite is often true. Compute, storage, bandwidth and cloud all became dramatically cheaper—and usage exploded. The key question isn’t whether prices fall; it’s whether demand grows faster than prices decline. So far, inference demand appears to be doing exactly that.
Second, it conflates frontier model developers with the entire AI ecosystem. Even if OpenAI or Anthropic struggle to earn attractive margins, it doesn’t follow that NVIDIA, TSMC, Broadcom, Micron, power suppliers, networking vendors or enterprise software companies face the same economics. Infrastructure providers and application developers have never captured value equally.
Third, the comparison to 2008 is weak. The mortgage crisis was driven by hidden leverage, opaque securities and fragile bank funding. Today’s AI capex is largely being funded by some of the world’s strongest balance sheets—Microsoft, Alphabet, Amazon and Meta. A capex slowdown is not the same thing as a banking crisis.
Finally, the piece largely ignores supply constraints. Much of today’s spending reflects shortages in HBM, advanced packaging, power and transmission—not speculative overbuilding alone.
There are legitimate questions about frontier-lab profitability and capital efficiency. But the more plausible downside is a multi-year capex normalization and industry consolidation—not a Lehman-style collapse. History shows transformative technologies often experience investment bubbles while still permanently changing the economy. Hat trick AI
It is obvious that printing presses with movable type created reproductions of written materials much more efficiently than scribes churning out handwritten copies. However, the question is whether there was a pre-existing market for printed materials from a literate public or did the availability of printed materials make literacy a more valuable skill? (If there is little available to read, why learn to read?)
The same questions could be asked of AI. Is AI meeting (or will AI meet) an existing market need more efficiently than current methodology or will the availability of AI create new markets/uses? If it's the latter, buckle up, hold on to your money, there will be some turbulence.