There have been some serious tremors in the AI firmament. The stock of CoreWeave, the largest of the neoclouds, lost 11.7% of its value last week after losing more than 25% the prior week. The interest rate on its long-term debt now exceeds 16%. The other neo-clouds are facing similar problems.

Among the hyperscalers, both Google and Meta had bad weeks, but Oracle is the one on life support. Its stock fell 1.6% last week, putting it down 12.7% over the last month, and 55.0% over the last year. The yield on its long-term debt is approaching 9.0%. This puts it well into junk bond territory, which is usually defined as above 8.0%.

The yields on debt matter hugely for the neoclouds and the hyperscalers because they are enormous borrowers. The trillions of dollars needed for building out the AI infrastructure is coming largely from debt issues. This is especially the case for the neoclouds, which don’t have big profits to rely upon, but it is also increasingly true for the hyperscalers. Companies like Alphabet and Microsoft have massive profits from their other business lines, but even these are not large enough to cover the costs of their data center construction plans. Oracle is especially vulnerable because it doesn’t have a massively profitable business to fall back on.

But, What’s the Use?

The difficulties of many of the AI companies are becoming increasingly evident. But the underlying issue is what will be the ultimate demand that can justify the massive infrastructure buildout we are seeing now, and the even larger amounts projected for the next three years?

There continues to be substantial increases in AI adoption, but it is overwhelmingly for low-cost AI. Meta was reportedly thrilled that its newly introduced Muse app was downloaded almost 11 million times in the first 40 days after it was introduced. However, the app is free. Meta does plan to sell advertising, but how much revenue will this raise, and of that revenue, how much will simply be pulling advertising away from its Facebook or Instagram platforms? Some number of users will surely upgrade to paying versions, but that still is not likely to get very far.

As many in the industry are saying, the low-cost models produced by China are good enough for most users. And even if users opt for the somewhat pricier U.S. versions, they are still not likely paying huge sums.

I have a lawyer friend who pays $20 a month for an Anthropic system. He finds it very useful for his work. That produces some revenue for Anthropic, but not the sort of money that will allow it to meet its lease commitments. His $20 monthly fee means $240 a year for Anthropic. There are somewhat more than 1 million lawyers in the country. Suppose 1 million buy the Anthropic system. That gets the company $240 million a year in revenue. That is a long way from meeting hundreds of billions in annual lease commitments.

While people in many other professions may also find it useful to spend $240 a year or so on AI, that is not going to produce anything like the bonanza that the industry is anticipating. As virtually every analysis to date has shown, spending on AI is hugely concentrated, with the vast majority of businesses spending very little. As an analysis from the Investment firm Apollo showed, there is still no evidence of a return from AI outside of the tech sector.

According to Ramp, just 1% of firms account for 80% of the spending at OpenAI and Anthropic. By itself, this should cause people to question the AI bonanza story, but it gets worse. It seems the vast majority of these heavy users are other AI companies. Their end use is developing other AI. Furthermore, they are heavily dependent on venture capital; they don’t currently have a profitable business model. Since venture capital firms will not be throwing money at upstart AI companies indefinitely, there is a real risk that some of the current heavy spenders may not be long for the world.

The story that the AI industry will be generating multiple trillions in revenue in the next couple of years is looking ever less plausible. There are already investors backing away from AI and stories of data centers being cancelled, ostensibly due to political opposition and difficulty arranging power. It sounds much better to say the local political opposition stopped the data center than that the demand doesn’t justify building it.

The cancellations are promising news. They may lead to many weeks for the AI industry like what the neoclouds have been enjoying.

Dean Baker is a senior economist at the Center for Economic and Policy Research and the author of the 2016 book Rigged: How Globalization and the Rules of the Modern Economy Were Structured to Make the Rich Richer. Please consider subscribing to his Substack, from which this is reprinted with permission.