Executive Summary
This week delivered the clearest signs yet that the financing engine inflating the AI bubble is sputtering.
We aren’t seeing it in equity prices, or in earnings.
It’s more subtle than that. Beneath the surface, the debt structures, private credit funds, and off-balance-sheet entities quietly underwriting the AI arms race are beginning to fail.
Here’s what’s happening – and why it matters – entirely through the lens of credit, liquidity, and systemic fragility.
This is a story about the infrastructure behind markets and securities.
I. AI: “artificial intelligence,” but also “ain’t inexpensive”
The AI boom isn’t just a technology story. It’s a financing story. Just like railroads and telecoms, AI requires massive investment in hardware, datacenters and the electricity to keep them running.
AI infrastructure requires unprecedented levels of capital expenditure (capex) – like $2 trillion over the next two years.
Where does the money come from?
- Publicly-traded corporate bonds
- Private credit loans
- Special-purpose financing vehicles (SPVs)
- Multi-year lease-back agreements (that look like 1990s telecom vendor financing)
Whether AI continues its explosive growth will depend less on innovation and more on the stability of the credit channels funding it.
This is where cracks are now emerging.
II. Private credit, public problems
Private credit has grown from a niche category into a $3 trillion shadow banking system.(“Private credit” means non-publicly-traded loans made to businesses.)
This has been the quiet workhorse behind the AI build-out – especially data centers and GPU financing. The average AI datacenter is a 200 megawatt facility costing about $2 billion, but prices range from $5 million to $60 billion.
This week delivered three major warning signs:
1. Redemption freezes
Blue Owl – one of the biggest private credit firms – halted withdrawals from a major fund as redemption requests surged.
This is the first clear sign of liquidity mismatch in the system. When investors want to pull their money out and they can’t? Not a good sign.
2. High-profile borrower failures
Recent collapses of borrowers (Tricolor, First Brands, Renovo Home Partners) forced lenders to mark loans down to zero overnight.
This is a reminder of how illiquid and opaque private credit pricing really is. As Jeff Gundlach recently observed, “There’s only two prices for private credit – 100 or zero.”
3. Credit vehicles failing internal tests
A BlackRock-managed CLO of private loans breached over-collateralization tests multiple times – rare for a highly rated structure. Interestingly, as far as I can tell none of the problematic lending had to do with AI capex.
Still, this indicates:
- Deterioration in underlying loan values
- Rising expected losses
- Overstated asset valuations in private portfolios (100 rather than zero)
In other words: The pipes that carry credit into the real economy are getting clogged.
III. AI runs on credit
AI infrastructure spending is not funded by cash flow or profits. It is being financed.
Heavily.
The industry’s capital needs are enormous:
- ~$350 billion in data-center spending this year
- $400 billion projected next year
- and $5 trillion in total projected over the decade (conservative estimate)
Instead of taking on debt directly, tech heavyweights are using less traditional methods of financing these purchases.
- SPVs – separate companies established specifically for the purpose of taking out loans. This keeps debt off the primary company’s balance sheet
- Circular financing arrangements where AI companies invest in one another and then use that financing to purchase GPUs.
If the circular financing thing sounds fishy to you, that’s because it is. Bloomberg called it “a web of circular deals” while The American Prospect chose a more ominous metaphor, The AI Ouroboros.
If credit tightens, even slightly, there are severe consequences.
The entire AI build-out slows down – not for technological reasons, but for funding reasons. I stress this because we saw this exact pattern in prior cycles:
- 1999–2001 telecom build-out
- 2005–2007 subprime mortgage expansion
- 2020–2021 SPAC boom
Each of the above ended not because the underlying technologies or companies were “bad,” but because the financing mechanism funding the sector broke.
IV. Games accountants play
Several developments highlight the fragility:
1. Asset depreciation is accelerating
AI chips have short useful lives (just 1-3 years, according to Princeton’s Center for Information Technology Policy).
However, accounting rules allow companies to depreciate those chips over a longer period (5-6 years). That makes their revenues and profit margins look better.
If you never lived with a CPA, here’s an analogy: You want to lease a Lamborghini for 3 years – but you can’t afford it, so you pay over 6 years instead. Lowers your monthly payments! But it also means you spend three years paying for a car you can no longer drive.
Generally speaking, this is NOT a good idea! But it’s both perfectly legal and makes the company’s books look more attractive. The Economist calls it the “$4trn accounting puzzle at the heart of the AI cloud.”
2. Off-balance-sheet financing is masking leverage
Tech giants are playing a shell game. By moving capex into SPVs and lease-back structures, they avoid showing billions in debt on their books. Again, perfectly legal – but misleading in the sense that risk is shifted into less-visible corners of the market.
The early-2000s energy and telecom companies (notably Enron and WorldCom) did the same thing. Letting a corporate parent book profits while a separate entity recorded the debt isn’t always fraudulent. But it’s suspicious.
3. Some financiers are stepping away
Fortress and Bain Capital publicly announced they’re opting out of next-generation data center financing. They see the AI build-out as an “arms race” with uncertain winners.
This is stunning because Fortress specializes in distressed and “special situations.” Bain is usually an early adopter in thematic infrastructure lending. When experienced, risk-loving lenders step back, everyone should pay attention.
V. The music’s still playing
I’m not calling a bubble peak. I’m not forecasting imminent catastrophe. Because, right now, we are not seeing:
- A broad credit freeze
- Widespread defaults
- Systemic panic
But we are seeing:
- The first liquidity gates (private credit redemption freeze)
- Early loan impairments (not yet AI related)
- Widening credit spreads
- Hesitation from major lenders
- The first financing failures tied (directly or indirectly) to AI-related debt
This isn’t a crisis – not yet. It’s phase one of a tightening cycle, where credit availability begins to shrink at the margins.
Historically, that is where major capex-heavy speculative frenzies begin to slow, before rolling over.
VI. What I’m watching
The key takeaway is simple:
The AI industry’s biggest vulnerability isn’t technology – it’s financing.
As long as:
- credit stays cheap,
- lenders continue taking risk,
- and private credit vehicles retain investor confidence,
Then the AI infrastructure race will continue at full throttle.
But the moment:
- liquidity falls,
- funding models stumble,
- or refinancing windows close,
Then AI infrastructure spending – the backbone of today’s tech boom – will buckle.
That’s the real systemic risk to watch. And it matters more than you’d think. In early October, the Financial Times warned “America is now one big bet on AI.” In the first half of 2025, AI capex contributed more to U.S. GDP growth than consumer spending (second chart below).
In case you’ve forgotten, consumer spending is typically 2/3 of GDP.
All this to say, what happens in the tech sector’s credit market will affect AI capex, which will have broad economic consequences.
VII. The bottom line
We are entering a phase where credit quality, liquidity, and funding mechanisms tell a truer story than day-to-day headlines or earnings releases.
But it’s a story that’s much harder to tell – full of forgettable characters and confusing acronyms – so you won’t see a lot of mainstream reporting on it.
The issues discussed above do not signal an immediate downturn. But it does show that the foundation of the AI boom – the financing engine – is overheating, and desperately needs an oil change.
If nothing else, please understand that right now, the credit-fueled AI boom has massive and underappreciated macroeconomic effects.
I’ll be watching further developments closely.
Update:
- Dr. John P. Hussman recently provided a very useful overview that incorporates current AI capex in his post An Unstable Equilibrium. And, unlike me, he’s actually qualified to give investing advice!
- David Dayen beat me to the punch in The American Prospect with his article The AI Bubble Is Bigger Than You Think. Honestly, his version of the story is both more narrow and more compelling than my own; strongly recommended.