Artificial intelligence firms initially justified their extreme capital investment—the four largest tech companies expect to spend more than $750 billion for AI infrastructure just this year—by saying that the technology would replace all human workers. They’ve since recognized what an unbelievably bad PR pitch that was, and have pivoted to promote a sunnier scenario where “we’re going to be able to keep people at the center of everything,” as OpenAI’s Sam Altman said in May.
Read more Not All Nepo Babies Are Trumps
More from David Dayen
But there’s a sobering reality underneath the rhetorical shift: AI is turning out to be more expensive for businesses than paying their workers. And that could be one of the many triggers that collapses the fragile economic edifice that the dreams of AI are propping up.
The news has mostly been relegated to the business pages, but AI firms repriced their product for business customers in recent months. Instead of a subscription fee to use OpenAI’s ChatGPT or Microsoft’s Copilot or Anthropic’s Claude, they now use token-based billing. Every time the model is queried, a small fee is charged. This is a very common technique, hooking users on a product and then charging more. But it’s thrown corporate planning for incorporating AI completely out of whack.
Companies that previously told workers to use AI in every facet of their job are now seeing how that affects the bottom line. One unnamed company reportedly spent half a billion dollars on Claude in a single month. Part of this is because AI is being used on mundane tasks like making PowerPoint presentations to “prove” rapid take-up of the technology, something highly prized on Wall Street.
Now, though, we’re seeing the snapback. Companies like Uber and Tesla and Meta and Microsoft are capping worker token usage. (For Tesla, Elon Musk’s in-house product, Grok, is exempted from the cap.) Palantir CEO Alex Karp said bluntly earlier this month that “something has gone completely wrong” with the billing model. These are some of the biggest evangelists for AI adoption out there, and some of them have AI businesses themselves; if they’re scrambling, then imagine what even somewhat more skeptical firms are thinking.
This cost conundrum is separate from other enterprise concerns about AI. Ford Motor Company rehired hundreds of skilled engineers; losing their institutional knowledge and subsequently failing to catch AI mistakes proved to be incredibly expensive, costing billions of dollars. Meta’s Mark Zuckerberg has said something similar, conceding in internal town halls that AI development wasn’t as rapid as he’d expected. But AI errors are the kind of thing that future improvements could theoretically weed out over the years. Simple cost-benefit analysis that makes AI too expensive cannot be remedied by a faster or smarter tool.
That’s a serious hurdle for the economics of AI, which is really another way of saying it’s a serious hurdle for the U.S. economy.
The repricing was necessitated by persistent money-losing from the big AI modelers. OpenAI and Anthropic are simply not making the kind of revenue to keep up with their compute spending, which one estimate puts at 70 percent of entire industry revenues. They raised prices because they had to show at least something approaching an acceptable balance sheet in advance of planned IPOs.
Read more Commies EVERYWHERE!
If companies that have adopted AI are now rebelling against the repricing, then the AI business model has a fatal corruption. The market valuations of these companies, and the “pick-and-shovel” companies that feed them computer chips and data centers to house them, are based on perpetual growth. Capped usage of AI does not fit with that vision.
The stock market depends on not only promises of growth but the infrastructure spending that reinforces it. The firestorm of protest over data centers is slowing down deployment—the data center pipeline was cut in half as far back as the end of last year—but delays are one thing, while excess capacity because nobody wants to pay for the end product is something else, something a lot worse.
Meta’s lunge to develop a cloud computing business may be the first signal of overcapacity. Zuckerberg doesn’t really know what to do with a company led by a dying social network at the end of its growth cycle, and he’s casting about for anything to excite investors, from virtual reality to prediction market apps to smart glasses with paywalled subscription fees. The cloud compute idea is part of this flailing, but the fact that Meta wants to rent out some of its existing compute capacity to serve the new business line is an acknowledgment that it has that capacity available.
In other words, Meta spent more than it needs to, and may have to scale back or find an alternative use. That is a horrifying thought for investors. And there’s more: After a monster IPO, SpaceX has flattened out, and even bonds on its debt aren’t selling. We’ve seen tech stock pullbacks over the last month that have been explained away as just technical, but they’re happening with more frequency. Certain gauges show the market more overvalued than it was on the eve of the 1929 collapse. And companies like Nvidia supplying the AI build-out would have no cover if there were a prolonged stumble.
A leaked draft report at the Treasury Department has likened AI risks to investors to the early-2000s dot-com bubble, but with a bigger effect on the broader economy because of the deeper impact of AI. If anything, the report is undershooting it. As I have written, the economics inside data center construction has more similarities to the housing bubble, complete with sketchy lending and buildups of debt. Despite these warnings, large investors continue to plow money into the sector and the private credit apparatus underlying it.
Where the leaked Treasury report is correct is its claim that any change in conditions could bring about crisis, and that companies rolling back their AI usage en masse would certainly qualify. Treasury has said the report came from a “low-level staffer” and, at most, will talk about data center delays as a systemic risk. But that’s hardly the biggest risk here; in fact, if it limits overbuilding that the economic fundamentals cannot rationalize, those delays may be a saving grace.
So maybe we should see Sam Altman offering the U.S. government 5 percent of OpenAI as a scheme not to enrich the public but to dump something on us that isn’t as lucrative as anybody thinks.
Read more The Meat Industry’s Tireless Effort to Keep Live Pigs in Coffins