The AI Bubble, March to October 2026: How the Debate Changed

In March, Google searches for “AI bubble” hit an all-time high. The debate back then was mostly theoretical: bears pointed at the gap between infrastructure spending and revenue, bulls pointed at adoption curves, and nobody had much new data.

Seven months later, we have plenty of data. Revenue grew faster than almost anyone predicted. Chip stocks lost about a trillion dollars anyway. The Fed raised rates. And today the IMF joined the list of institutions warning that an AI bust could hit the wider economy.

The debate didn’t get settled. It changed shape. Here’s how.

Seven Months of the AI Bubble Debate

MarPeak "is it a bubble?" searchesMixed
OpenAI closes a round at an $852B valuation. Anthropic run-rate revenue roughly doubles to ~$20B.
Apr–MayRevenue answers backBullish
Anthropic run-rate reported at ~$30B (Apr), then ~$47B (May). Fed survey: AI climbs to a top-3 financial stability risk.
JunFirst real crackBearish
Tech leads the sharpest sell-off in over a year. Bears start going public against parts of the AI trade.
JulChips give back $1TBearish
Nvidia sheds roughly $1T in market value in under two months. Big Tech still guides ~$700B+ of 2026 capex.
AugDebt becomes the storyBearish
Hyperscalers tap external financing as capex outruns cash flow. CoreWeave borrows at ~10.4%.
SepRate hike, then recordsMixed
Fed hikes for the first time in 3+ years (3.75–4%). AI stocks drop, then the Nasdaq hits a new high on Sep 21.
OctThe institutions weigh inBearish
IMF, Australia's central bank and Ray Dalio warn of a bust. OpenAI run-rate ~$50B disappoints. Anthropic IPO slips to November.

March: The Question Everyone Was Asking

The starting point was Sequoia’s “$600 billion question”: the gap between what was being spent on AI infrastructure and what the AI ecosystem actually earned. Paul Kedrosky called it “one of the probably five largest CapEx bubbles in history”, next to canals, railroads and fiber optics.

On the other side, the demand looked real. OpenAI closed a funding round at an $852B valuation, and Anthropic’s annualized revenue run-rate roughly doubled to about $20B in a matter of months.

At that point, both sides were arguing about the future. That changed quickly.

April–May: Revenue Answers Back

The bull case had its best stretch in spring. Anthropic’s run-rate was reported at ~$30B in April and ~$47B in May. Nvidia kept posting record quarters.

The bull argument in spring: a bubble is spending without demand. Revenue run-rates more than doubling in a quarter don’t look like that.

But regulators were already paying attention. In the Federal Reserve’s spring Financial Stability Report, the share of surveyed market participants citing AI as a salient risk reportedly jumped to about 50%, up from 30% the previous fall. Private credit climbed the same list, partly because software companies, the ones most exposed to AI disruption, are a big share of private credit borrowers.

June–July: The First Real Crack

In June, tech led the sharpest sell-off in more than a year, and investors who had stayed quiet started taking public positions against parts of the AI trade.

July made it concrete. Nvidia gave back roughly $1 trillion in market value in under two months, and the sell-off spread to memory, optics and leveraged capacity builders like CoreWeave. The worries piled up:

OpenAI revenue coming in below expectations
Custom chips from the hyperscalers threatening Nvidia’s margins
High rates making long-duration bets more expensive

And yet, in the same weeks, Big Tech guided to roughly $700–725B of combined 2026 capex, up about 75% from 2025. The companies with the best view of demand kept spending more, not less. That tension, falling stocks and rising spending, is the core of the whole debate.

August: Debt Becomes the Story

This is where the conversation really shifted. In March, the bear case was “the spending can’t be justified.” By August it was “the spending is now being financed with debt.”

The new bear argument: hyperscalers are putting most of their operating cash flow into capex and tapping external financing to cover the rest. Equity-funded bubbles hurt shareholders. Debt-funded bubbles can hurt everyone else too.

FactSet documented hyperscalers turning to external financing as capex outran cash flow, with some pausing buybacks. Lending got more expensive at the edges: CoreWeave’s August loan reportedly priced 100–125 basis points above initial talk, at around 10.4%. And Nvidia drew scrutiny for backing financing deals for its own customers, a version of the “circular funding” worry that keeps coming up.

September: Rate Hike, Then Record Highs

On September 16 the Fed raised rates by 25 basis points to 3.75–4%, its first hike in more than three years. For a trade built on cheap capital and long payback periods, that’s the worst kind of news. AI and chip stocks fell hard.

Five days later, the Nasdaq hit a new record high on an AI chip rally.

That whiplash sums up September. Capital Economics found most of its bubble indicators at or near levels that preceded past market peaks. Meanwhile UBS estimated that AI investment, plus the wealth effect from AI stocks, accounted for 80% or more of US economic growth. Bears read that as dangerous concentration. Bulls read it as proof that AI is the economy now.

October: The Institutions Weigh In

Which brings us to this week:

Ray Dalio called AI a “classic bubble” approaching its bursting point, pointing at debt-funded spending and rising rates.
A strategist at Panmure Liberum warned the AI trade could end in the worst S&P 500 crash since 2008.
The Reserve Bank of Australia flagged the AI boom as a financial stability risk.
The IMF, in today’s Global Financial Stability Report, warned that an AI bust could spill well beyond the tech sector.
OpenAI’s run-rate was reported at about $50B, below expectations, and the stock market took it badly.

And the big test is a few weeks away. Anthropic’s IPO, reportedly pushed from October to November, could raise up to $100B at a valuation approaching $2 trillion. A leaked prospectus, reported by Reuters, showed $4.6B of 2025 revenue against a $42B net loss. The company has reportedly told investors it expects a second straight profitable quarter, and its run-rate passed $65B in late July. OpenAI, meanwhile, has ruled out an IPO before 2027.

For the first time, the public market will get to price a frontier AI lab directly.

What Actually Changed Since March

Question	March 2026	October 2026
Is demand real?	Probably, early signs	✅ Yes. Lab revenue grew several times over
Is spending slowing?	No	❌ No. ~$700B+ capex guided for 2026
How is it funded?	Mostly cash flow and equity	⚠️ Increasingly debt
What do rates look like?	Expected to fall	⚠️ Rising again (first hike in 3+ years)
Have markets corrected?	No	⚠️ Partially. Chips lost ~$1T, then rebounded
Who’s worried?	Investors and commentators	⚠️ Central banks, the Fed and the IMF
Has it burst?	No	No

The pattern is clear. The demand side of the debate mostly went to the bulls. The financing side moved toward the bears. In March the argument was about whether AI would make money. In October it’s about whether the money will arrive fast enough to pay back the debt that’s funding the buildout, now that rates are going up.

My Take

In March I would have said: not a bubble, but a correction is coming. Looking back, that was half right. We got a correction in chips, and the companies with real revenue proved the demand was there.

What I didn’t see coming was how fast the risk would move from valuations to balance sheets. A stock market correction hurts investors. A debt problem spreads, which is exactly why the IMF and central banks are the ones talking now, not just analysts.

Here’s what I’m watching for the rest of the year:

The Anthropic IPO. Real numbers, audited, priced by the public market. If it goes well, it validates the business model. If it struggles, it’ll reset private valuations across the industry.
Q4 capex guidance. If hyperscalers keep raising capex while borrowing more, the debt question gets louder.
Rates. One more hike changes the math for every long-payback data center project.
Credit spreads for AI infrastructure. CoreWeave-style lenders are the canary. If their borrowing costs keep rising, that’s the early signal.

Whether the technology works isn’t up for debate anymore. I use it every day, and so does every team I know. What’s in doubt is whether the way it’s being financed is sustainable, and that’s the question that matters now.

What do you think: is this a debt problem, or just the cost of building something big?

Resources

Follow Up

I’ll revisit this after the Anthropic IPO. Drop a comment if there’s a signal you think I’m missing.