The AI Bubble Is Bigger Than You Think - The American Prospect

One thing I’ve been tracking this year is the areas where Wall Street and Silicon Valley are going to war. Tech firms clearly want to become banking apps and receive special charters, private equity and crypto are jostling for position in worker 401(k) plans, and the tech right in general wants to supplant big banks as the go-to director of conservative business policy.

That’s all still going on. But in one area, Silicon Valley and Wall Street are in sync: conjuring up sketchy credit deals that are pointing us toward another financial crash.

More from David Dayen

Last month, the big focus was “round-tripping,” the way that sundry AI and tech companies were investing in their own customers—with Nvidia giving AI companies the investment necessary to buy their graphics processing units (GPUs), and so on. But there’s a lot more to this story, tangled up with yet another rebrand by the former masters of the universe, from “shadow banks” without proper regulation into something boring and neutral-sounding: private credit. Ever since the advent of financial regulation, there have been companies that have attempted to evade the rules with creative branding. Private credit companies are non-banks that are trying to rebrand into a name that doesn’t tell everyone they are unregulated lending vehicles.

The speculative financing of the artificial intelligence buildout is happening mostly in private credit, where assets under management hit $1.6 trillion in February and are likely higher today. The deals being made are perverse and irrational; there are huge mismatches between the life cycle of the assets being funded and the amount of time it will take to pay them off. Experts have been doing everything but picketing the stock exchange with signs that say “BUBBLE IN PROGRESS,” yet the country has so much sunk cost in the AI boom that the pathway feels inevitable. “We have sealed the deal on another financial crisis—the question is size,” said one former congressional staffer.

I WILL TRY TO EXPLAIN THIS as simply as I can. The build-out of computing power for AI needs about $2 trillion in annual revenue by the end of the decade to justify the current and planned investment. It’s an insane amount of money, nobody has it—nobody may ever have it—and so everything being constructed now, from the GPUs needed to train AI models to the data centers housing them to the energy supplying those data centers, needs some creative financing.

Big Tech companies have large cash flows, but the trillions needed are too steep even for them. Venture capital isn’t that interested in funding infrastructure either; while the God in the machine that may result from the AI build-out has potential for supercharged returns, data centers and power plants, by themselves, won’t get you there.

Related: The AI ouroboros

But investors are dying for this kind of investment opportunity, believing that a Big Tech–directed industrial policy will endure for years. Big Tech firms have historically not borrowed much, however, and though companies have issued twice as many bonds this year than at any time in the recent past, they would like to maintain strong credit ratings by limiting debt exposure. Smaller AI companies, meanwhile, don’t have the confidence of the markets to borrow hundreds of billions in capital.

Enter the “special purpose vehicle” (SPV). A company pops up to build a data center, and they get an “anchor tenant” that’s a Big Tech firm. (Data centers rent their space to tenants; they are infrastructure, but they’re also like apartment buildings.) It finances the data center through debt sales, implicitly (or explicitly) promising that Big Tech firm payments will pass through to investors.

This is the case with an SPV for Meta’s $30 billion Hyperion data center in Louisiana. Blue Owl, a private credit fund, owns the majority stake in the SPV; it put in a small amount of equity, and the SPV carries the debt. Meta has another 20 percent, which gives investors the confidence that they’ll get repaid without Meta having to put debt on its balance sheet. This has a side effect of highly rating these debt instruments, with “specialized” rating agencies doing much of the work, in ways that are more than suspicious.

Blue Owl is a private credit fund, which again is an end run around financial regulations. These companies suck up money from investors, well over a trillion dollars according to The Wall Street Journal. Blue Owl has $295 billion on its own.

It’s pitched as a bulletproof trade; Big Tech firms are locked into leases, and their payments will pay back the debt. But if that’s the case, why is Blue Owl, in the aftermath of a merger of two of its private credit funds, blocking redemptions in a way that will automatically give those investors 20 percent losses? As a traditional bank, that move would be akin to a bank run; this is just pissing off rich people who got into an arrangement with limited rights. But the warning signs are unmistakable. And it’s not limited to Blue Owl; covenants on corporate debt are gradually being rewritten to protect private credit in the event of a fallout.

A great paper from the Center for Public Enterprise lays out additional alarms. Data centers are not lasting infrastructure, or at least the guts inside them aren’t. If the GPUs are working overtime to compute AI models, they may not last more than two years, and high-end AI firms will always want to upgrade to the latest version anyway. (By the way, our friends at Blue Owl are lending money to Elon Musk’s xAI to buy Nvidia GPUs.)

But AI firms are extending their depreciation schedules for the GPUs; they’re saying they will last much longer than they likely will. That means overstated revenues, as companies have to purchase far more GPUs than they are admitting publicly, and thence possible financial disaster. Indeed, some of the smaller companies are using loans backed by their GPUs to purchase other GPUs.

It also means that these hulking buildings, and the dedicated power plants some of them are constructing in the vicinity, are likely to become stranded assets when the site outlives its usefulness.

Did I mention the securitization? Data centers roll over from construction loans into asset-backed securities, where investors can pick a tranche based on their risk comfort. The CPE report says that 61 percent of the applicable securitizations in this market are coming from data centers. But again, these securitizations happen years into a rapidly depreciating asset.

Everyone’s all in on data centers now; we’re in bubble inflation territory, and it’s unclear when it will pop. Of course, the big problem at the end of this road is that there just isn’t enough expected cash flow to pay any of this off. OpenAI lost more than $11.5 billion last quarter, and wants to spend far more in future years. The potential revenue to make up the gap is not anywhere on the horizon.

If the Chinese AI models that are more efficient at training—in part because it is possible to reconstruct LLM datasets by purchasing their output—break ahead of the U.S. counterparts, there’s not going to be any way to finance the growing pile of infrastructure and real estate assets being built. But the big AI firms have no interest in efficiency here. The same companies making the models are the ones who own the cloud computing firms that are pushing more compute to everyone. In fact, the money made from cloud compute is subsidizing the model training at Amazon and Microsoft and Google. Everyone is financing the expansion of everyone else. That’s a classic description of a bubble.

There’s a whole middleman aspect to this as well, the “neoclouds” that build data centers on spec and rent them out. (CoreWeave is a terrifying debt-ridden entry in this genre.) But I think you get the picture. We have a 2000s housing bubble level of financial engineering on top of a 1920s level of private unregulated lending on top of something bigger than a 1990s internet (or 1870s railroad) level of technology and infrastructure build-out. It’s one bubble to rule them all.

WHILE THIS PILE MAY BE CENTERED IN PRIVATE CREDIT, private equity funds, real estate investment trusts, and other vehicles have pieces of it. A Moody’s report last month said that banks have at least $300 billion in loans to private credit. Blue Owl has BlackRock, Invesco, and PIMCO selling the bonds financing the Hyperion data center.

Because this is the only thing moving in this economy, everyone on Wall Street wants some of the action, even those who are wary of it. Deutsche Bank, for example, is talking about reducing its exposure to data center risks, including by shorting AI stocks. But within that, the bank has acknowledged that some of its traders had “gone heavy” on financing for data centers.

That puts the effort by Trump’s financial regulators to deregulate traditional banking, including the Federal Reserve slashing bank supervision, in a different light. It may be hard for a private credit firm to get bailed out by the government; for the banks, it’s easier. If banks soak up and carry over the debt from the Blue Owls of the world, they have better opportunities to have it effectively extinguished.

So while the biggest names swap money back and forth to buy each other’s stuff, these special purpose vehicles and their investors could be left holding the bag. That increasingly could be ordinary retail investors with 401(k) plans. Financial policymakers I’ve talked to are perhaps most alarmed by this part.

There are enough people in the know, like Peter Thiel, quietly sidling away from this investment/fiasco that nerves should be pretty frayed if you have any concern about the U.S. economy. Blue Owl’s stock is way down this year, as skepticism with private credit abounds; it fell 6 percent just on Monday. It’s not exactly comforting when Google’s CEO says “no company is going to be immune, including us” if the AI bubble bursts.