I may not know everything about business, but I’m reasonably confident that it’s a bad sign when you write a seven-page letter about how your company is not Enron. That was the task of some poor soul at Nvidia over the Thanksgiving holiday, trying to calm market nerves put on edge by short sellers alleging financial improprieties. One sensational claim asserted that Nvidia was having trouble collecting billions of dollars in sales, with spikes in inventory and lower cash flow than its competitors. Michael Burry, last seen correctly predicting the housing bubble in The Big Short, also questioned how much cash Nvidia was leaking to investors and employees. And skepticism about Nvidia funding its own customers to bulk up sales has already been rampant.
Nvidia refuted much of this in its memo, but there was a quality to it best expressed by an internet meme. To sum up, Nvidia’s “We are not Enron” T-shirt has people asking a lot of questions already answered by the shirt.
A slap fight between a tech company and some short sellers would be less interesting if we all weren’t so reliant on the outcome. Nvidia’s graphics processing units (GPUs) are the guts of most of the data centers now proliferating throughout the country. Deals between computing providers, AI model makers, speculative data center builders, and private credit firms (read: shadow banks) lending into the build-out represent a historically large source of the nation’s industrial activity. It’s powering energy demand, stock valuations, and wild venture capital deals for companies with no products, just vague mumblings of “something something AI.”
In short, AI has eaten America. Whether it rests on a solid foundation or on quicksand will determine whether we have a future of prosperity or pain. And so when Nvidia starts wearing a “We are not Enron” T-shirt, that sinking feeling sets in.
But what are the actual vulnerabilities here? Why could AI as a business fail even if its models produce something useful in some form? The possibilities are so vast that it’s hard to summarize, but let me try.
AS OUTLINED IN THE LATEST EPISODE of the Organized Money podcast that I co-host (listen above), data centers are both infrastructure and technology, but they may be better thought of as real estate. They are giant boxes with tenants who need computing power, built with construction loans that roll over into traditional real estate loans and bundle into mortgage-backed securities, in case you wanted to be reassured that nothing untoward is going on. After all, when have mortgage-backed securities ever been a problem for the economy?
Usually an “anchor tenant”—a big AI firm like OpenAI or Meta or Amazon—guarantees payments of what amounts to rent. They’re paying for the build-out, but through complicated off-balance-sheet vehicles owned by lenders and underwriters, as I’ve talked about at more length elsewhere. But AI firms are also floating corporate bonds in unprecedented numbers, growing the debt load.
The money ultimately covering this debt financing is not coming from AI-related revenues, but cash flow that Big Tech companies earn from surveillance advertising and cloud computing and other business lines. “The big four in this space (Amazon, Meta, Google and Microsoft) … they’ve made about $35 billion in reported revenue from AI, but they have spent over $560 billion on the technology,” said Advait Arun, author of a great report on the financial structure of the AI industry for the Center for Public Enterprise called “Bubble or Nothing.” A JPMorgan report puts the need for AI revenues at $650 billion a year by the end of the decade in order to finance the current level of spending and make a small return.
Investment in new technology always needs startup financing, and if AI is truly God in the machine, wide adoption will pay off over time. But that isn’t materializing, at least not yet. As The Economist recently pointed out, Census Bureau research shows that the number of American workers using AI has either flatlined or fallen after three years of access to the technology, with the sharpest drops at large businesses. AI firms really need corporate revenue; high prices for people to send around cat videos doesn’t feel sustainable. Sarah Myers West, co-executive director of the AI Now Institute, noted on the podcast that companies are looking at cybersecurity risks and error rates from AI and making the determination that it’s not yet worth it to jump in.
That’s bad for model makers and for the companies one step removed from direct revenue. Oracle has borrowed $18 billion to build data centers, with rental profit of just $125 million on low margins. If the current hot demand cools because nobody’s buying AI, there’s little chance that Oracle can pay off the debt. (There’s a big jump in credit default swaps that are hedges against Oracle’s corporate debt: another financial crisis–era buzzword that is returning to the lexicon in uncomfortable ways.)
Perhaps a bigger problem is on the spending side, where the true cost may be underestimated. The GPUs inside the data center, where all the value is maintained, burn out over an uncertain time frame. Depending on the tasks given, they could lose their value over two to three years, even though companies are claiming a four-to-six-year life cycle. If the companies are wrong—or if they want the latest and greatest chips, if only to keep them away from their competitors—that’s an additional heavy expense. As Ed Zitron has pointed out, the newest Nvidia Blackwell chips might require entirely new data centers to handle their energy and cooling needs.
“The big worry here is that a lot of these tenants will have to be taking multiple cycles of really expensive capital expenditure within the lease term, which doesn’t even last the whole term of the debt of the facility itself,” Arun said. If the tenant can’t afford to essentially refurnish their house every year with GPUs, they’ll exit, and you’ll have vacancies leading to lower cash flow. That would also be the case if demand slows down and AI startups go out of business. Either way, the result sounds a lot like 2008: empty units and not enough money coming in to finance the debt.
These risks have put no real brake on the mania, however, and in fact there are accelerants. The largest companies making models are also the dominant companies selling cloud computing—Amazon Web Services, Microsoft’s Azure, and Google Cloud. So we end up with high-cost, low-efficiency versions of AI models, because the companies making them want to sell cloud services to themselves. “This version of building AI has come into vogue because it’s been shaped by the incentive structures of these hyperscalers,” West said.
All of this is separate from the real liabilities AI firms are incurring from copyright infringement and defamation and assisting suicides and self-harm, which have led insurance companies to separate their exposure to those firms. AI becoming uninsurable is yet another in a series of downside risks.
DOOMSAYERS ARE OFTEN DERIDED for predicting nine of the last five recessions. And the doomsaying here, as long as AI demand is insatiable and investors see the potential, will look foolish for a little while. But over the longer term, overinvestment with little sustainable revenue growth is just mathematically going to be a problem, as it has been in the past. Arun sees a key hinge point around 2028, when data center leases start to come due and get rolled over into long-term real estate loans only if they have enough tenants to justify them. If companies involved in buying GPUs with stock warrants or other noncash instruments ask for actual money instead, and there isn’t any money available to pay them, “that’s when the contracts start falling,” Arun said.
There’s also a risk to local governments that have offered tax breaks to attract data centers, on the theory that whatever the data centers do pay will be good for their economies. This doesn’t appear true: West cited a report that localities are losing between 52 and 70 cents for every dollar spent on data center tax exemptions. If the economics of data centers fall apart, that will create a real hole in municipal balance sheets. Not to mention the fact that a lot of the holders of this pile of debt are institutional investors like pension funds. Nobody is that safe from an AI financial crash.
When usually staid moderate Democrats are asking the Financial Stability Oversight Council to test AI for systemic economic risk, when the Organisation for Economic Co-operation and Development lists AI as a “key downside risk” for the U.S. economy, when the best that investors like Mohamed El-Erian can do is say “maybe the bubble is good,” and yes, when Nvidia is wearing a “We are not Enron” T-shirt, it’s just plain hard to shrug off concerns about disasters that could manifest in so many different ways.
There’s another aspect to this, which has pushed much of the public into a negative posture regarding AI. “What does the scenario look like if AI as a sector is wildly successful?” asked West. “It’s a vision of mass displacement, and people being out of jobs … it is the sublimation of the rest of our productive economy for the sake of this one sector, which profoundly benefits only a few companies.”
When your “win” is that bleak, you start to almost root for the loss, which would be catastrophic in its own right. That’s the dilemma policymakers are now backed into, with little interest as yet in getting themselves out of it.