Nobody Warned Them: How AI Bets Torched $35 Billion Overnight
In this article
Bottom line: A single trading session last week wiped roughly $35 billion off AI-adjacent stocks after a mid-tier cloud provider disclosed it was carrying GPU-backed debt it couldn't service without a fresh infusion of hyperscaler cash.
The selloff wasn't about AI failing to work — it was about the financing structure underneath the AI buildout: circular vendor deals, leveraged data-center loans, and revenue that depends on the next round closing.
If your paycheck or portfolio touches AI infrastructure, the lesson isn't "AI is a bubble" — it's that the compute layer and the model layer are now two very different risk profiles, and most people are still pricing them as one.
I watched $35 billion evaporate from the AI sector in a single trading session, and the strangest part wasn't the number.
It was how many people in my Slack channels acted like they'd seen it coming all along.
Nobody had actually done the math. They just felt, correctly, that something underneath the AI story didn't add up — they just couldn't tell you what.
I can. I spent the last three years building infrastructure for companies riding this wave, and I've watched the financing get weirder every quarter while the actual technology kept getting better.
Those two facts coexisting is exactly why so many smart people got blindsided this month.
The Setup: A Boom Built on IOUs
Here's the scenario, stripped of ticker symbols. A cloud provider — call it a GPU landlord — signs a multi-year compute deal with a foundation model lab.
To build the data centers that deal requires, the landlord borrows heavily against future revenue from that same contract.
The chipmaker, meanwhile, has taken an equity stake in the lab that's renting the compute, and extends favorable financing terms to the landlord buying its chips.
Money moves in a circle: chipmaker invests in lab, lab pays landlord, landlord buys chips from chipmaker, chipmaker's stock goes up because landlord revenue looks like real demand.
Everyone in the loop reports growth. Nobody in the loop has taken in a dollar from outside the loop.
I'd been tracking this pattern since early 2025, mostly because it kept showing up in vendor contracts I was reviewing for infrastructure decisions.
It's not fraud — it's leverage, dressed up as partnership. And leverage is fine right up until one node in the chain has a bad quarter and can't make a debt payment.
That's what happened. The mid-tier landlord in this story missed a covenant on GPU-backed debt. Credit default swap spreads on data-center debt widened within hours.
Every stock with exposure to that lending chain — chipmakers, hyperscalers, the "AI infrastructure" ETFs retail investors had been piling into since 2024 — got repriced by algorithms that don't care about your roadmap.
$35 billion, gone before lunch.
The Core Insight: Two Different Bets Got Confused
The thing I need developers and tech professionals to internalize is this: there are two separate bets happening under the single label "AI," and last week's crash only broke one of them.
Bet One: Does the Model Work?
This bet is doing fine. I use Claude 4.6 for architecture review and code migration work daily. I run ChatGPT 5 for rubber-duck debugging when I'm stuck on something gnarly.
Gemini 2.5 is my default for anything involving long-context document analysis. These tools are materially better than what I had eighteen months ago, and they save me real hours every week.
None of that changed last week. Not one token of inference got worse. The technology bet — does the model produce useful output — has been answered, and the answer is yes.
Bet Two: Does the Financing Survive?
This is the bet that broke.
It has almost nothing to do with model quality and everything to do with whether the companies building data centers can service the debt they took on to build them, on a timeline that assumes demand keeps compounding at 2024-2025 rates forever.
That assumption was always the fragile part. Compute demand is real, but it's not infinite, and it's definitely not evenly distributed.
When one landlord's utilization numbers came in soft enough to spook a lender, the market didn't just reprice that landlord — it repriced every company whose growth story depended on the same assumption holding everywhere else too.
Why the Confusion Happened
Most retail investors, and honestly most engineers I talk to who dabble in the market, were pricing chipmakers and cloud landlords as if they were the same trade as "AI is useful." They're not.
A company can make a genuinely transformative product and still go bankrupt from bad balance sheet decisions made while scaling it.
That's not a new story — it's the dot-com playbook, just with GPUs instead of fiber.
The Reality Check: Where the Hype Actually Breaks
I want to be careful here, because it's easy to slide from "the financing is fragile" into "therefore none of this matters," and that's just as wrong as the hype it's reacting against.
What's real: AI coding assistants, agentic workflows, and model-driven automation are delivering measurable productivity gains right now, today, in production systems.
I've shipped things this year I genuinely couldn't have shipped as fast three years ago. That's not marketing copy — it's my own git history.
What's fragile: the capital structure paying for the data centers those models run on.
Debt-financed infrastructure only works if utilization stays high enough to cover interest payments, and the industry has been building capacity on the bet that demand grows in a straight line.
It doesn't.
It never has, for any infrastructure buildout in history — rail, fiber, cloud, or otherwise. There's always a digestion period, and digestion periods are exactly when overleveraged players get exposed.
What gets missed: this doesn't mean the labs themselves are in danger. OpenAI, Anthropic, and Google DeepMind aren't the ones carrying GPU-backed construction loans.
It's the landlords and the mid-tier neoclouds sitting between the chip and the lab that are exposed, and most people can't tell those companies apart from a ticker symbol.
The Practical Takeaway: What to Actually Do With This
If you're a developer, don't let this shake your confidence in the tools themselves.
Keep using what works. Nothing about last week's selloff changes whether Claude writes a cleaner refactor than you'd write at 6pm on a Friday.
If your job depends on an employer whose revenue is itself downstream of one of these vendor-financing chains — a startup reselling GPU capacity, a company whose main customer is a single hyperscaler's AI division — it's worth actually reading their most recent funding announcement for the word "partnership" doing double duty as "customer." That's the tell.
Real customer revenue doesn't usually come with an equity stake attached.
If you're investing, separate the two bets explicitly.
Betting on model capability improving is a very different risk than betting on a specific data-center operator's balance sheet surviving a demand digestion period.
Conflating them is how you end up owning something that drops 20% in an afternoon for reasons that have nothing to do with whether the product works.
And if you're building anything that depends on cheap, abundant compute continuing indefinitely — stress-test that assumption now, while it's cheap to do so, instead of during the next liquidity crunch when everyone else is doing the same math at the same time.
Where This Goes Next
I don't think this is the last $35 billion afternoon we'll see.
The circular financing problem doesn't get solved by one bad week — it gets solved by the industry actually diversifying revenue away from vendor-to-vendor deals, which takes years, not quarters.
Expect more of these jolts as utilization numbers get reported and some of them disappoint.
The technology isn't the bubble. The balance sheets underneath it might be.
Did you feel the selloff coming, or did it catch you off guard like it caught my Slack channel? What's your read on where the next crack shows up?
