# Is AI a financial bubble? What each side of the debate actually argues

Date: 2026-10-30 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/is-ai-a-bubble-debate/

**Answer:** Nobody can prove it yet, because a bubble is usually confirmed only after prices fall. Those who see one point to debt-funded spending, modest productivity gains so far and negative free cash flow, while those who do not point to fast-growing revenue, huge contracted backlogs and companies that say demand still exceeds supply.

A financial bubble is when prices rise far above what future earnings can justify, sustained by the expectation that prices will keep rising. The word does not mean the underlying technology is fake. Railways and the internet both changed the world, and both also ruined many investors. So "is AI useful?" and "are AI prices too high?" are separate questions.

The question is unusually live. On October 30, Intel fell 25%, Nvidia 4%, and Alphabet had just had what CNBC called its worst trading day ever, while the Financial Times reported that OpenAI had cut its funding target. Prices falling for a few days does not settle anything, in either direction.

## The case that it is a bubble

**The debt is growing quickly.** The Bank of England says AI companies passed a turning point in 2025, when investment needs outran internal cash, and that the shift to outside finance sped up in 2026. The stock of debt began the year modest. The Bank says that this protection is eroding fast. The IMF says hyperscalers now make up 10% of global investment-grade bond issuance, against under 1% in 2024.

**The cash is going the wrong way.** Alphabet, Meta and Oracle all reported negative free cash flow in their latest quarters, and Amazon's reached −$42.4 billion over 12 months.

**Productivity gains look small so far.** A survey of nearly 9,000 executives in seven countries, published through the National Bureau of Economic Research, found AI added about 0.2% a year to labour productivity growth over three years. Executives expect 0.6% a year over the next three. Spending is highly concentrated: 2% of firms make up over half of it.

**Hardware may age faster than the books assume.** The investor Michael Burry argues that AI chips stay useful for two to three years, not the five or more used in hyperscaler accounts. If he is right, profits are overstated. The hyperscalers keep their longer estimates, and the argument is open.

**The IMF sees some warning signs.** Its October report cites "growing evidence" that AI capabilities may need less investment than planned, or that the economic gains may be smaller.

Even insiders have used the word. Sam Altman, OpenAI's chief executive, told The Verge in 2025 that investors as a whole were overexcited about AI, and that someone would lose a phenomenal amount of money.

## The case that it is not

**Sales are real and rising fast.** Nvidia's quarterly revenue doubled from $46.7 billion to $96.2 billion in a year, to late July 2026. Microsoft reported Azure growth of 48%. Anthropic's revenue went from $1.2 billion to $16.3 billion between the first halves of 2025 and 2026, according to its filing, where half-year figures are unaudited.

**Customers have signed contracts.** Backlog, formally "remaining performance obligations," is revenue promised under contract but not yet recognized.

| Company | Backlog | Latest date |
|---|---|---|
| Microsoft (commercial) | $677 billion | Sept 30, 2026 |
| Oracle | $664 billion | Aug 31, 2026 |
| Google Cloud | $560.8 billion | Sept 30, 2026 |

**Capacity, not demand, is the stated limit.** Microsoft says delays at third-party GPU suppliers are pushing capacity into 2027, as reported by CNBC, and Nvidia raised its supply commitments from $119 billion to $279 billion in a single quarter.

**Most spenders can absorb a hit.** These are very profitable companies, unlike many dot-com firms of the late 1990s. The IMF's stress test found bank capital broadly resilient in a hypothetical AI slowdown, though banks holding 22% of global assets ended up with weak capital.

**Some overbuilding can still leave value.** Jeff Bezos said in 2025 that this looks like an "industrial" bubble, not a banking one, and that society can gain from such episodes when the dust settles. That argument concedes a bubble but says the damage would be limited.

## Where the two sides read the same facts differently

| Fact | Optimistic reading | Pessimistic reading |
|---|---|---|
| Huge backlogs | Demand is contracted | Much may rest on a few labs that have not yet proved they can pay |
| Vendor investments | Partners have conviction | [Sales may be partly self-funded](/en/learn/circular-financing-spv-ai-explained/) |
| Fast valuation rises | Revenue is growing even faster | Prices assume decades of success |
| Rising capex forecasts | Supply is short | Everyone is building for the same future at once |

The backlog point deserves a closer look. Oracle's filing attributes its rise to "certain significant cloud contracts" and says only 13% will become revenue over the next 12 months. Meanwhile OpenAI, a major buyer of compute, is itself funded by its investors, and the Financial Times reports projections of about $278 billion in negative free cash flow from 2026 to 2030, according to TNW's summary. Those are projections, not audited results.

## What a crash could look like

The Bank of England ran a hypothetical, not a forecast: a 45% fall in US stocks over six quarters, a 3.5 percentage point rise in corporate borrowing spreads and a weaker dollar. UK output fell by 2.2 percentage points, driven mainly by financial channels. It judged that UK banks could withstand a shock of that size. So a bust would be painful without necessarily being a banking crisis, unless debt keeps growing, which is exactly what the Bank says it is watching.

> **Try it:** Pick one claim from a headline about the AI boom, then write the strongest reply from the opposite side, using one number from a filing.

## What evidence would settle the argument?

Nothing settles it in advance, but a few indicators would shift the balance. Toward "bubble": customers cancelling or renegotiating contracts, resale prices of older chips collapsing, a default on data centre debt, or lenders refusing to fund neoclouds, the smaller GPU-rental firms. CNBC reports some lenders are already pulling back from that group, and Gartner predicts that 80% will be absorbed or close by 2030.

Toward "not a bubble": revenue at labs and clouds outpacing depreciation, free cash flow turning positive as buildouts mature, and productivity numbers rising outside the tech sector, in surveys and national statistics.

This article sets out the arguments. It is not investment advice, and it does not suggest buying or selling anything.

## Sources

- Financial Stability Report, July 2026: https://www.bankofengland.co.uk/financial-stability-report/2026/july-2026
- IMF cites AI slowdown, spiking yields as looming global market risks: https://www.zawya.com/en/capital-markets/bonds/imf-cites-ai-slowdown-spiking-yields-as-looming-global-market-risks-1607687
- Firm Data on AI (NBER Working Paper 34836): https://www.nber.org/papers/w34836

---

From the Guide to Artificial Intelligence, by Robert Scales, published by Holistic Hybrid Inc. Share and adapt it under CC BY-NC-SA 4.0 (https://creativecommons.org/licenses/by-nc-sa/4.0/): credit the guide, link to https://guidetoartificialintelligence.com/en/learn/is-ai-a-bubble-debate/, no commercial use without permission, and keep the same licence.