A "Clever" Way to Crash the Economy Again
The AI Data Center Debt Bubble Is 2008 With Better PR
I’ve been reading about how companies raising the tens (and in some cases hundreds) of billions of dollars needed to build AI data centers have landed on a financing strategy that gets described as clever: borrow the money, then bundle that debt and sell it off to institutional investors in pieces.
It isn’t clever. It’s a rerun.
We’ve seen this movie before
That exact pattern — pooling a large number of loans and slicing them into tranches to sell to institutional investors — is what caused the 2008 mortgage crisis. Banks bundled home loans, marketed them as high-quality debt, and sold them off. The flaw was simple: the whole scheme assumed housing prices in different regions wouldn’t all fall at once. They did. Supposedly nobody saw it coming. Everybody saw it coming.
The data center version isn’t hypothetical or something I’m speculating might start — it’s already well underway. Data center loans are increasingly being securitized through commercial mortgage-backed securities (CMBS) and asset-backed securities (ABS), the same structures used for mortgage bonds. JPMorgan projects annual data center securitization issuance could reach $30 billion to $40 billion in both 2026 and 2027, representing 7% to 10% of combined CMBS and ABS issuance in those years, up from about $27 billion in 2025.[1] That’s on top of a much larger wave of AI-related corporate debt and private credit — AI-related companies and projects tapped debt markets for at least $200 billion in 2025 alone, and Morgan Stanley expects $250 billion to $300 billion of issuance in 2026 from hyperscalers like Microsoft and Meta.[1]
It’s not just journalists and bloggers drawing the 2008 comparison. The Bank for International Settlements warned in its annual report that an AI investment bust could hit credit markets with disruption comparable to the 2008 financial crisis, specifically flagging what it called “circular financing” — chipmakers and hyperscalers taking equity stakes in AI labs or neocloud providers who in turn commit to multi-year purchases of chips or computing power from those same investors, with deal terms that are typically poorly disclosed.[4] Lawyers who worked structured-finance litigation after 2008 are using the same word to describe what they’re seeing now: one attorney who worked on housing-crisis litigation said tracking developments in AI data center financing feels like “deja vu,” noting we’re talking about trillions of dollars with almost no transparency about the financing structures.[3]
The scale is the scary part
In 2008, the underlying problem wasn’t just that some mortgages were bad — it was how enormous the mortgage market was, and how much of it had been securitized and distributed into “safe” institutional portfolios. Data center financing is being talked about in the same order of magnitude. Global data center spending could reach $7 trillion by 2030 according to McKinsey estimates,[3] and that money increasingly can’t come just from hyperscalers writing checks out of cash flow — it’s being raised through corporate bonds, private credit, and securitization, layered on top of each other in ways that are hard for any single investor to fully see through.
Will the demand actually be there?
The 2008 crisis depended on a bet that home prices could never fall everywhere at once. The AI data center bet depends on demand for AI compute continuing to grow at something close to its current pace, forever, to justify the debt taken on to build it.
I’ll admit the data here is messier than I expected going in. Per-token prices for AI models have actually been falling sharply, not rising, thanks to competition and efficiency gains — and so far that hasn’t suppressed demand; if anything, usage has grown even faster than prices have dropped. Anthropic’s annualized first-quarter 2026 revenue rose fivefold to $45 billion, and token consumption is forecast to keep climbing even as chipmakers lower annual token costs by 60% to 70%.[5] So the simple story of “prices are rising, so demand must be cooling” doesn’t hold up cleanly.
But that doesn’t mean the underlying concern is wrong — it just points at a different pressure point. The real open question is whether AI usage translates into revenue that can service all this debt. Enterprises are already pushing back on cost, with many companies reporting minimal measurable profit from generative AI pilots,[5] and Goldman Sachs Research has found that gains from the AI buildout have concentrated heavily in infrastructure stocks — semiconductors, hyperscalers, data center operators — with equity gains far outpacing growth in actual forward earnings, a gap that has made analysts more cautious about how long capital spending can run ahead of visible monetization.[6] If growth in paying, profitable AI usage doesn’t keep outrunning the buildout — whether because adoption plateaus, competition erodes pricing power faster than volume grows, or the ROI just isn’t there for enough customers — you get a glut of very expensive, purpose-built infrastructure with no way to pay off the debt behind it. And these facilities aren’t easily repurposed: they’re built around GPU racks, power density, and cooling systems tailored to one job.
There’s also a real question, separate from the financing mechanics, about what all this compute is actually for. It’s not always clear whether the buildout is meant for bigger models, more inference to serve existing demand, or a more speculative bet that enough compute gets you a qualitative leap toward general intelligence. It’s probably some mix of all three. But the dollar figures being thrown around are what make the uncertainty concerning — you don’t want to find out after the fact which bet you were actually financing.
Nobody is stepping in
In 2008, the mortgage crisis eventually forced a regulatory response, but not before triggering a global financial crisis. What’s needed now is the thing the U.S. is generally bad at: getting ahead of it. That means regulators actually looking at how these data center financing deals are structured and passing rules that prevent debt from being oversubscribed or repackaged past the point where anyone can trace the real risk.
There are early signs some people in government are paying attention. Four U.S. senators sent a letter this year calling on the government to investigate how Big Tech is turning to complex and opaque debt markets to borrow staggering sums of cash, warning that massive debt loads could cause destabilizing losses for financial institutions and trigger a broader financial crisis.[3] Whether that turns into actual legislation before something breaks is another matter.
To be fair, there are real differences from 2008 worth naming. Data center tenants are predominantly investment-grade hyperscalers with enormous cash reserves, the underlying demand for AI compute is demonstrably real and growing, and the physical assets have genuine utility that residential mortgage-backed securities lacked when housing demand collapsed.[2] That’s a meaningfully different starting point than subprime borrowers with no ability to pay. But “the borrowers are richer this time” was also roughly the argument made about corporate debt before plenty of past blowups, and it doesn’t address the structural issue: opaque, layered securitization that lets risk pile up in places nobody’s really pricing correctly, at a scale — trillions of dollars — that rhymes uncomfortably with the last time this went wrong.
A side note on Andrew Yang’s tax-AI argument
Andrew Yang has been making a related but distinct case: that the U.S. should tax AI itself rather than labor, on the theory that AI is going to displace workers at scale, and the tax code shouldn’t keep penalizing employers for hiring humans while automation goes untaxed.[7] His argument is built on the same underlying number that worries me here — the trillion-plus dollars going into AI infrastructure — but he uses it to argue corporations need to generate enough revenue from AI to justify those valuations, and that revenue has to come from somewhere, likely payroll.
Taxing the construction and operation of data centers directly could be another route to some of what Yang wants, even if it’s not exactly what he’s proposed. A tax on data center capex or on-site power draw would land squarely on the physical infrastructure driving both the job-displacement story and the financing bubble described above, and it would generate revenue precisely at the moment resources are being committed — rather than waiting to tax AI revenue or job losses after the fact, when the money may already be spoken for by debt service. It also sidesteps some of the “where does AI stop and the human interpretation start” line-drawing problems that critics of a compute or automation tax raise, since a data center is a discrete, physical, easy-to-measure thing to tax. Whether that’s better or worse policy than taxing compute or automation directly is a separate question, but it’s worth having on the table alongside whatever regulatory response eventually comes to the financing side.
Sources
The $3 Trillion AI Data Center Build-Out Becomes All-Consuming For Debt Markets — EnergyNow, Feb. 2026
AI Data Center Financing Hits $200B: CRE Impact 2026 — The AI Consulting Network, Apr. 2026
AI data center boom ‘stress tests’ insurers as private capital floods in — CNBC, Apr. 2026
The BIS warns an AI bust could hit credit markets as hard as the 2008 financial crisis — The Next Web, June 2026
The AI Bubble Is Stable As A Price War Forces A New Reality — Forbes, June 2026
How AI Token Supply Shapes Prices and Investor Returns — The AI Insider, July 2026
Tax AI, Not Humans, Again — Andrew Yang, June 2026