Yet More AI Bubble Worries, Now on Debt Side, with Wall Street Journal Featuring AI Datacenter Borrowing “Frenzy”

Yours truly is overdue on an overview how private-debt-backed AI datacenter deals work and why companies like Meta who are perfectly capable of borrowing in their own name paying 200-300% basis points extra not to do so is not a good sign. But the Wall Street Journal just published a very detailed story on the general outlines of this debt binge, which has gotten more attention in the popular press as borrowing levels have skyrocketed this year:

Big Tech borrowing for AI data centers:

2015-2024 average: $32B/year
Sept-Oct 2025 alone: $75B

> Meta borrowed $30B
> Oracle borrowed$18B
> $META also did $27B off-balance sheet with Blue Owl

AI companies now 14% of IG index

The “money-printing” tech companies are…… pic.twitter.com/IuhBU0LS4M

— junkbondinvestor (@junkbondinvest) November 5, 2025

And from the new Wall Street Journal story:

What the Journal calls a “frenzy” in Wall Street Blows Past Bubble Worries to Supercharge AI Spending Frenzy is awfully reminiscent of the toxic phase of the subprime lending binge, when originators were so desperate for product, as the jibe then went, they’d fund borrower who could fog a mirror. It was years later that we were able, with the key elements including some insiders plus a remarkable pre-crisis analysis by (of all things) an equity market analyst, Henry Maxey of Ruffler Investment, to piece together how leverage-on-leverage created an even-then, well-reported “wall of liquidity”. It was also the leverage on leverage (and that systemically important yet fragile financial institutions were heavily exposed) that made the bubble unwind so catastrophic. We have warned that the unwind of an, erm, traditional credit bubble, even if very large (see Japan and for a lesser but still nasty version, the S&L crisis) typically produces at worst very deep and protracted recessions and zombification (the runup to the 1929 US crash also featured CDO-like leverage on leverage, see Frank Partnoy’s The Match King for details), as opposed to banking system near-or-actual failures.

So again, we have yet to see evidence of that meteor-hitting-the-financial-system event being in the offing.

But absence of evidence does not amount to evidence of absence.

And perhaps finance historians can correct me, but I do not recall a historical instance of a massive equity bubble (without 1929-style heavy borrowings directly against those equities) accompanied by so many red flags, particularly operating and financial leverage, in the underlying commercial activity. That included recursive deals among key companies and as we’ll discuss a bit in this and more in posts to come, overly-clever borrowing structures that make sense only to achieve higher levels of leverage than could be achieved by traditional means. The Journal points out in passing that one of the mega-deals will pay more than 2% that a plainer-vanilla offering would. In a fine overview of in June on how these financings work, Paul Kedrosky similarly said the premium was 200 to 300 basis points. That’s an awful lot to pay for opacity and supposed balance sheet remoteness.¹

As a brief introduction to causes for concern about the datacenter boom, see the executive summary from Bubble or Nothing from the Center for Public Enterprise (hat tip Matt Stoller):

● Cash flow uncertainty persists as the cost of providing AI inference services continues to rise. Leading AI inference service providers are not particularly differentiated from one another; this competitive market structure suppresses market
participants’ pricing power and prevents them from recovering rising costs.

● The collateral value of a graphical processing unit (GPU), the sector’s keystone asset, looks poised to fall in the near-term. The value of chips fluctuates depending on uncertain user demand as well as the supply dynamics and technical specifications of new GPUs, now released yearly. The cash flow that GPU collateral can demand is suppressed due to the sector’s competitive market structure and the uncertain depreciation schedule of existing GPUs.

● Data center tenants will undertake multiple cycles of intense and increasingly expensive capital expenditure within a single lease term, posing considerable tenant churn risks to data center developers. This asset-liability mismatch between data center developers and their tenants will strain developers’ creditworthiness without guarantees from market-leading tech companies.

● Circular financing, or “roundabouting,” among so-called hyperscaler tenants—the leading tech companies and AI service providers—create an interlocking liability structure across the sector. These tenants comprise an incredibly large share of the
market and are financing each others’ expansion, creating concentration risks for lenders and shareholders.

● Debt is playing an increasingly large role in the financing of data centers. While debt is a quotidian aspect of project finance, and while it seems like hyperscaler tech companies can self-finance their growth through equity and cash, the lack of transparency in some recent debt-financed transactions and the interlocked liability structure of the sector are cause for concern.

The first two points alone, the fact that inference costs are not only not falling but still rising, and near-term downside in GPU prices ought to be fatal, or at least detrimental to debt funding.

Yet as you’ll see, even though the Journal raises concerns, and even adds to this list by describing how AI “hyperscalers” are putting out duplicate orders for the datacenter capacity, virtually assuring whipsaw, it’s not as sobering as the list above. But it includes signs that some dogs are turning their noses up at the dogfood:

Stock prices normally go up when a company reports record revenue but after Meta did just that on Oct. 29, its shares plummeted 11% instead. The reason: Zuckerberg disclosed he will “aggressively” increase capital spending on AI, drawing questions from analysts about how the company plans to actually make money off the new technology.

What the Journal describes parallels the late-stage subprime lending frenzy. In 2007, CEO Chuck Prince of later-big-bailout recipient Citigroup famously remarked:

When the music stops, in terms of liquidity, things will be complicated. “But as long as the music is playing, you’ve got to get up and dance. We’re still dancing….The depth of the pools of liquidity is so much larger than it used to be that a disruptive event now needs to be much more disruptive than it used to be.

Consider from the Journal the, erm, enthusiasm of lenders and fears of becoming wall flowers in this party:

Silicon Valley’s biggest players are flush with cash and were able to fund much of the initial AI build-out from their own coffers. As the dollar figures climb ever higher, they are turning to debt and private equity—spreading the risks and potential rewards more broadly across the economy.

Some of the financing is coming from plain-vanilla corporate bond sales, but financiers are making far bigger fees off giant private deals. Virtually every Wall Street player is angling to get a piece of the action, from banks such as JPMorgan Chase and Morgan Stanley to traditional asset managers such as BlackRock.

Investor appetite for data-center debt is so strong that some money managers have booked billion-dollar gains in a matter of days, even before construction of the facilities they are financing is complete.

Still, the longer-term performance is hardly assured. Big tech companies are expected to spend nearly $3 trillion on AI through 2028 but only generate enough cash to cover half that tab, according to analysts at Morgan Stanley.

Big names in the financial world, such as Goldman Sachs CEO David Solomon, are warning about AI-fueled froth in the markets and in capital spending.

At the same time, the fear of missing out is real. Days after Solomon voiced his concerns to analysts, Goldman formed a new team in its banking and markets group focused on AI infrastructure financing.

Later from the Journal:

Funds that invest in AI deals say they carry little risk, because tech companies with deep pockets have ironclad leases that will generate the money to pay investors back. Microsoft has a higher credit rating than the U.S. government, and it told investors on Oct. 29 that it would double its total data-center footprint in the next two years.

Perhaps I am too old, but I recall that IBM and GE were once AAA rated too, and that by 2000 (for IBM) and 2010 (for GE) their luster had taken quite a turn. And remember the pre-crisis pattern that housing prices had never fallen nationwide, only regionally? These supposedly blue-chip tech companies are placing monster bets on AI, so their historical solidity isn’t as germane as it might seem, unless they back off if the fundamental performance of large language models continues to fall well short of promises. The reaction of the Meta stockholders to Zuckerberg’s promise of even more big AI spending is confirmation.

Back to the Journal:

Tech executives see more risk in underbuilding than overbuilding…

But some tech companies are weaker financially than others. Oracle…needs to borrow billions more for its spending spree, prompting Moody’s Ratings and S&P Global Ratings to edge closer to reclassifying Oracle’s bonds as junk debt. In recent weeks, the company’s stock price has fallen 32% and its bonds have lost about 7%.

There’s also the risk that the chips tech firms are borrowing to buy could be obsolete in a few year…

The last time Wall Street went all-in on an industry was the fracking boom—then bust—over a decade ago. This time, financiers are marshaling even larger sums.

The article continues with a breathless account of the boomtown effect that these datacenter buildouts are generating. Matt Stoller flagged the concern of growing real-economy dependence on massive spend in what ought to be a niche:

A few months ago, I asked why our economy, despite steady growth in official numbers, feels so creepy and unstable. My conclusion was that the U.S. is in a “Chinese finger trap” economy. We are dependent for growth on monopolies and an AI bubble, which juices the all-important stock market. Trying to grow an economy in a more stable way could lower stock prices which would paradoxically lead to a downturn. So we’re stuck, until some outside event occurs….

Data center developers are now approaching multiple utilities with proposals for the same project, leading to “phantom” forecasts of demand that isn’t actually there. Essentially, it’s hoarding.

Hoarding is what happen in overheated markets….But this can lead to something called the “bullwhip effect.” Buyers overstate how much they want to buy….all of a sudden the demand evaporates because it was never real in the first place. This dynamic can throw an economy into an overheated state, and then a depression; that’s what happened globally after World War I, leading to, among other things, Mussolini’s takeover of Italy….

At this point, data centers are pretty much what’s growing in America. I recently had a conversation with an elected official who told me that data center construction is a huge construction jobs boost in rust belt areas….He posited a tension between political support for the new temporary jobs and political anger over higher electricity prices.

We have a one-legged economy, with the AI build-out serving as a driver of real estate values, the stock market, and GDP growth

This is a bet-the-economy scheme on models where China has much better mousetraps. There is no way this will end well.

____

¹ The crisis demonstrated that theory and practice can be two different things. Banks had long offloaded credit card receivables to investors in supposedly off-balance sheet deals. When losses on them rose to previously unthinkable levels, the investors successfully revolted and made the banks eat some of those costs. The reason was the bank credit card businesses depended on being able to keep using other people’s credit. I am not sure whether or not these AI datacenter borrowers won’t wind up in an analogous position, of being so dependent on ongoing lending that lenders won’t let them walk away from outsized credit losses.