Is it really an AI bubble? The case for, the case against, and what would settle it

A bubble is not simply a market that falls. It is a period when prices rise well beyond what the underlying cash flows can plausibly justify, held up by the expectation that someone will pay more later. The test is therefore not “is the technology real?” but “can the money coming in cover the money going out, at today's prices?” A technology can be transformative and its stocks can still be a bubble. That distinction runs through everything below.

• Spending is enormous and rising. Big tech capital spending is on track for roughly $800 billion this year, with analysts expecting $1.2 trillion next year, per Fortune. AI-related capex has grown at about 26% a year in real terms since the start of last year, per Bank of America as cited by CNBC. A roundup of analyst estimates cites Goldman Sachs as saying roughly $300 billion a year of AI revenue would be needed to justify about $800 billion of spending; we have not seen the original.
• Bank of America's own analyst says so. Aditya Bhave called such a front-loaded capex wave “symptomatic of over-investment”, per CNBC.
• Cash flow is already strained. Alphabet reported negative free cash flow of $9.07 billion in the quarter, against positive $2.7 billion expected, and fell 12.6% in a day; see that story. Microsoft's free cash flow fell 27%, though it stayed positive.
• Borrowing is rising. Evercore's Julian Emanuel said hyperscalers, equipment providers and AI companies sought hundreds of billions in financing over the past two weeks, putting the cost of capital and dilution in focus. BofA economists linked some of the rise in interest rates to borrowing for the build-out.
• Some deals loop. Critics point to arrangements where a supplier invests in a customer who then buys the supplier's products, such as Nvidia's announced investment of up to $100 billion in OpenAI alongside OpenAI's purchases of Nvidia chips. This is a widely reported structure and a commonly cited concern; it is not evidence of wrongdoing, and the terms are in the companies' own statements.
• Accounting for chips is disputed. Investor Michael Burry has argued that AI chips' useful lives are closer to two and a half years than the five to six that hyperscalers use, which would mean they are earning less than they appear to. Nvidia disputed his maths in a memo to analysts. If the shorter life is right, costs are being spread too thinly.
• The weakest links are thin. Gartner predicts 80% of neoclouds could cease independent operations by the end of 2030, and Starwood's Barry Sternlicht said some could fail, per CNBC.

• The big spenders are profitable. Alphabet's search advertising still grew 12% to $63.35 billion in the quarter, and Microsoft's free cash flow was a positive $18.63 billion, per CNBC. These are not unprofitable start-ups; they are among the most profitable companies ever built.
• Use is real and measurable. Goldman Sachs finds 24.6% of US businesses use AI in at least one function, and that adoption shows up in hiring, per Fortune. A bubble usually has more promise than use.
• Suppliers say demand is there. AMD's Lisa Su said she does not see AI chip demand slowing soon. That is a company with an interest in saying so, and it is the supplier's view, not independent evidence.
• A pause can be healthy. RSM's Joseph Brusuelas said AI investment “needs to slow to a more sustainable pace” and that a gradual deflation would serve the AI trade. On that view, the Thursday fall is a correction inside a still-valid story.

The usual comparison is the dot-com bubble. The Nasdaq peaked in March 2000 and fell roughly three-quarters by October 2002, according to widely published index history, and the internet went on to transform the economy anyway. The parallel the bears draw is that the fibre and servers built in the boom were far ahead of demand, and many of the companies that financed them disappeared. The difference the bulls draw is that today's biggest spenders have profits and cash, while many 1999 darlings had neither.

Our reading is that both comparisons are partly right, and they apply to different layers. The deepest-pocketed layer, the hyperscalers, looks more like the survivors of 2000. The debt-funded layer, neoclouds and some suppliers, looks more like the telecom builders of 2000. A bubble need not burst evenly, and the first thing to give is usually the most leveraged part.

1. 01Revenue against capex. If AI revenue growth closes the gap with spending, the bubble case weakens; if capex guidance keeps rising while revenue disappoints, as at Alphabet, it strengthens.
2. 02The cost of money. Rising borrowing costs hit the most leveraged builders first. Watch credit spreads and bond issuance for neoclouds and large AI developers.
3. 03Chip utilisation and pricing. Whether older chips still earn their keep decides the depreciation dispute.
4. 04Nvidia's next quarterly report, which one market roundup says is due on 17 November. Check the date with the company.
5. 05Policy. How far data-center limits actually go, which is the subject of part 6.

On the evidence in these sources, “AI is a bubble” and “AI is not a bubble” are both too simple. A fairer statement is that the technology is real and widely used, that spending has run well ahead of proven revenue, and that the financing structure of the weakest layer looks fragile. The Thursday sell-off is consistent with a market that has started to price that difference. It is not proof that a larger collapse is coming, and it does not prove it is not. The questions readers ask most are answered in part 11.

This is one of eleven connected pieces on the November 2026 election and market shock. Read them in order, or jump to the one you need.

1. 01The crash: what happened on 5 November
2. 02How $2.7 trillion was wiped out in a day, and whether that number is real
3. 03From rally to rout: why stocks rose on the Democratic win and crashed a day later
4. 04The Democratic win: what the midterm result changes
5. 05What a Democratic Congress means for the world economy
6. 06The data-center backlash and what it means for AI
7. 07Is it really an AI bubble? The evidence on both sides (this article)
8. 08Gold after the midterms
9. 09Bitcoin and crypto after the midterms and the chip crash
10. 10What happens next: the dates, scenarios and signals to watch
11. 11The discussion: the hardest questions about the crash and the AI bubble, answered

• BofA warns of AI correction risk as data-center backlash hits the 2026 midterms — Fortune
• Democrats may go after data centers. That's not a reason to dump all your AI stocks — CNBC
• S&P 500 Wipeout Hits $2.7T as Hedge Funds Dump Hyperscalers and Chip Index Plunges 11% — HuggingNews
• Alphabet loses $500B in one day — Payaries
• AI Bubble: Hidden Risks and Opportunities — Man Group