Is AI a Bubble? Triggers, Impacts, Timing and How to Prepare
Three years of warnings, $740 billion of spending a year, a growing pile of debt and a market that wobbled in October 2026. This explainer lays out what a bubble is, the full timeline, the numbers on spending and revenue, how circular financing and debt work, the honest case for and against, what could trigger a fall, what it would do, when to watch, and how founders, investors, workers and governments can prepare.
1. The question in one minute
- Is AI a bubble? The honest answer: the technology is not, the financing might be.
- Real money is being spent on real things (chips, power, buildings), and AI revenue is growing fast.
- But spending is growing faster than revenue, a rising share is borrowed, and some of it flows in circles between suppliers and customers.
- As of 3 November 2026 nothing has burst.
- October 2026 brought the sharpest stress so far: the Nasdaq fell about 10% in the month, Alphabet lost about 13% in a day, and the IMF warned that an AI correction could slow the world economy.
- Those are warnings and a correction in prices, not a crash.
- This explainer covers what a bubble is, the timeline, the numbers, who is lending, the case for and against, what could trigger a fall, what it would do, when to watch, and how to prepare.
Six numbers that frame the debate (Nov 2026)
Figures from the sources in this article; ⚠️ marks aggregator or secondary figures.
2. What is a bubble, exactly?
- A bubble is when prices run far ahead of the value the asset will ever earn.
- People buy because prices are rising, not because of the cash flows.
- It usually ends when new buyers or new credit run out.
- The classic pattern (Minsky, as told by Kindleberger).
- Displacement: a real new technology changes expectations.
- Boom: investment and prices rise on genuine early success.
- Euphoria: borrowing grows, prices detach from earnings, "this time is different".
- Distress: insiders sell, credit gets tighter, a surprise breaks confidence.
- Panic and revulsion: forced selling, prices overshoot down.
- Technology bubbles are special.
- Carlota Perez describes an "installation period" when finance floods into a new technology and overbuilds it, followed by a crash and a "deployment period" when society actually uses the infrastructure.
- Railways in the 1840s and fibre in 2000 both crashed and both became the backbone of what came next.
- So two different questions get mixed up.
- Is the technology overrated? Mostly no.
- Are prices and borrowing ahead of the money it will earn soon enough? That is the real debate.
Where the bubble cycle stands (our reading, not a forecast)
Minsky–Kindleberger stages; placing 2026 on them is our judgement.
3. The timeline: three years of warnings
The AI bubble debate, Sep 2023 → Nov 2026 (click a dot)
Each dot is a dated, sourced event; dashed outline = secondary source ⚠️. Toggle lanes to compare spending, financing, revenue, markets and official warnings.
- How to read it.
- The worry started with a simple sum (Sequoia, 2023–24): GPU spending implies revenue nobody yet earned.
- It became a market event in January 2025, when DeepSeek showed cheaper models were possible and Nvidia lost about $590B of value in one day.
- Financing became the story from late 2025: vendor investments, guarantees, private credit and record bond issuance.
- October 2026 is where every lane lit up at once: a revenue scare, bank pullback, the IMF warning and earnings selloffs.
4. How much is being spent
- The four biggest US platforms guided more than ~$740B of capital spending for 2026.
- Alphabet, Amazon, Meta and Microsoft combined; estimates range from about $650B to $825B depending on what is counted (finance leases, Oracle, neoclouds) ⚠️.
- Meta alone raised its 2026 range to $135–145B in October; Alphabet spent $51.4B in a single quarter.
- Several forecasts put 2027 above $1T.
- Why so much?
- Each generation of frontier model needs several times more compute.
- Serving agents and reasoning models consumes far more compute per user than chat did.
- Nobody wants to be the company that under-built (see Altman's view in our startup notes).
- The cash-flow squeeze.
- One analysis finds hyperscalers spending about 102% of cloud revenue on capex ⚠️; CreditSights reportedly found capex plus buybacks and dividends now exceeds projected cash flow for the big five.
- Depreciation lags: reported trailing depreciation (~$149B) is far below capex (~$434B), so the cost hits profits years later ⚠️.
Estimates of 2026 AI capex ($B, sources differ in scope)
Secondary compilations ⚠️; the range itself is the point: even definitions disagree.
5. The revenue gap
- The core arithmetic.
- Data centres must earn back their cost plus a return before the chips are obsolete (four to six years).
- Sequoia's David Cahn: take GPU spending, double it for the rest of the data centre, then double again for the end customer's margin. That gave a "$200B question" in 2023 and a "$600B question" in 2024.
- Bain estimates the industry needs about $6T a year of AI revenue by 2031 if capex keeps rising, assuming capex runs at about 25% of revenue.
- What revenue exists?
- Estimates of AI cloud revenue are around $150B annualised ⚠️; Microsoft has cited an AI run-rate in the tens of billions.
- The big labs grew several-fold in a year, but their numbers are not comparable: Anthropic counts cloud-partner sales gross, OpenAI net, which partly explains why OpenAI's run-rate was reported about $20B lower than assumed in October.
- Disclosure: Claude, made by Anthropic, helps write this site.
- The honest summary.
- The gap is large, but it is a gap in time, not necessarily a permanent hole: revenue would need to grow very fast for years, and spending would need to slow.
How big is the gap? Different methods, different answers ($B a year)
All rough. The method matters more than the number: what share of revenue can go to capex, and how long chips last.
6. Follow the money: circular financing
- What "circular" means.
- A supplier invests in or lends to a customer, and the customer uses the money to buy the supplier's products.
- Nvidia has invested in OpenAI ($30B in March 2026, as reported) and in other labs and neoclouds, and agreed to back up to $105B of financing for an OpenAI data centre in Ohio.
- Why it worries people.
- It can make demand look stronger than it is: sales partly funded by the seller.
- It concentrates risk: if a customer struggles, the supplier loses the sale and the investment at the same time, which the IMF calls "wrong-way risk".
- Lucent and Nortel lent heavily to telecom customers in the late 1990s; when those customers failed, so did the loans.
- The counter-argument.
- Nvidia's CEO: "a little bit of money in, a lot of money comes back".
- Morgan Stanley notes that bringing in third-party asset managers means most capital is not Nvidia's own.
- Unlike Lucent, Nvidia is extremely profitable and its chips are in short supply.
The loop, in five steps
Simplified; real deals mix equity, guarantees, leases and private credit.
7. Debt enters the boom
- Early AI spending came out of cash flow. That is changing.
- Axios (Oct 2026): about a third of AI capex is now funded by debt, and banks are pulling back.
- Goldman Sachs expects debt to fund about a third of hyperscaler capex in 2026, peaking around 35% in 2027.
- One tally: Microsoft, Alphabet, Amazon, Meta and Oracle raised about $326B of debt in 2026, roughly triple the previous year ⚠️.
- The weakest links are the newest borrowers.
- Neoclouds (GPU rental companies) borrow against GPUs and customer contracts; CoreWeave's yields went above 15% in October before Nvidia's CEO spoke up for it.
- Oracle carries large debt at the lowest investment-grade rating (BBB-) ⚠️.
- Special-purpose vehicles and private-credit funds keep some borrowing off company balance sheets.
- Why debt changes everything.
- Equity losses hurt investors; debt defaults force sales, spread to lenders and can freeze credit. That is how a price correction becomes an economic problem.
Who funds the build-out (from safest to riskiest, our ordering)
The lower the layer, the faster stress shows up there.
8. October 2026: the first real stress test
- 8 October: the revenue scare.
- Reports that OpenAI's annualised revenue was about $50B, around $20B below what many assumed, sent Nvidia, Oracle and CoreWeave lower.
- 13 October: the IMF.
- Its Global Financial Stability Report said markets are increasingly exposed to an AI-driven correction if spending fails to produce expected profits, and modelled about −0.4% to world growth in a moderate case. Banks' capital looked resilient.
- 16 October: credit fatigue.
- Banks limit AI lending; neocloud yields jump.
- 21–30 October: earnings.
- Meta raised capex; Alphabet missed and fell about 13% (about $500B of value) despite Google Cloud more than doubling; Amazon fell about 7% with negative free cash flow; Microsoft's Azure grew 48%.
- Result.
- The Nasdaq fell about 10% in October. Markets did not reject AI; they started to punish spending without matching profit.
October 2026 market moves (as reported)
Single-day or monthly moves from news reports; check exchange data for exact closes.
9. The case that it is not a bubble
- Demand is real and measurable.
- Cloud AI revenue is growing at rates rarely seen at this scale (Azure 48%, Google Cloud more than doubling).
- AI lab revenue rose several-fold within a year; coding agents are among the fastest-adopted software products ever.
- The spenders are rich.
- Most capex still comes from some of the most profitable companies in history, not from start-ups with no revenue, as in 2000.
- Scarcity, not glut, so far.
- Chips, power and memory have been the constraint; utilisation of new capacity has stayed high.
- Mainstream analysts are divided, not alarmed.
- Goldman: the rally has been driven by fundamentals, though some patterns "rhyme" with past bubbles.
- Julius Baer reads chip weakness as a crowded trade resetting; Ashmore expects slower growth rather than collapse ⚠️.
- Timothy B. Lee (Understanding AI) argues it still does not look like a bubble on revenue trends.
- Even if it is, the assets remain.
- Fibre laid in 2000 carried the 2010s internet. Data centres and power built now will be used, at some price.
10. History rhymes: 2000 versus now
| Dot-com and telecom, 1995–2002 | AI, 2023–2026 | |
|---|---|---|
| The technology | Internet, real and transformative | Generative AI and agents, real and transformative |
| Who spent | Telecoms and start-ups, heavily debt-funded | Mostly cash-rich platforms; debt share rising |
| Revenue at the time | Many companies with almost none | Large and fast-growing, but behind capex |
| Vendor financing | Lucent, Nortel lent to customers | Nvidia invests in and backs customers |
| What was overbuilt | Fibre: most of it unused for years | Possibly data centres and GPUs; not yet visible |
| Peak to trough | Nasdaq fell about 78% (2000–02) | Nasdaq −10% in Oct 2026, so far |
| What survived | Amazon, Google, the fibre itself | To be seen |
- The biggest differences.
- Profits of the spenders, and the fact that today's constraint has been supply, not demand.
- The biggest similarities.
- Spending justified by future revenue, vendor financing, record debt issuance and "this time is different" language.
11. What could trigger a fall
Triggers, from most to least discussed in the sources
Our ranking from how often each appears in the sources cited here.
- Triggers rarely act alone.
- A bubble usually ends when two hit together: for example a revenue miss during a period of tight credit. That is why October 2026 mattered: both appeared in the same month.
12. What a correction would do
| Scenario (our framing) | What happens | Who is hurt | Who benefits |
|---|---|---|---|
| 🟢 Soft landing | Capex growth slows, revenue catches up over 3–5 years, stocks drift sideways | High-valuation stocks; late investors | Steady builders; users get cheaper AI |
| 🟠 Correction | AI stocks fall 30–50%, neocloud defaults, capex cut for a year or two | Neoclouds, leveraged developers, private credit, chip suppliers | Cash-rich labs and start-ups: GPUs get cheap |
| 🔴 Crash with credit stress | Credit freezes, broad market falls, IMF-style growth hit or worse | Pension and retail investors, regions built around data centres | Few; long-term buyers of assets at the bottom |
- Economy-wide.
- The IMF's moderate case: about −0.4% to world growth relative to baseline; larger if credit tightens at the same time.
- AI capex has been a meaningful share of US growth recently, so a pause would show up in GDP.
- Inside the AI industry.
- GPU rental prices fall: good for researchers and start-ups (Gopala's compute budget in the Maestro series would shrink).
- Consolidation: weaker labs and neoclouds are bought; power projects are delayed or cancelled.
- Research and adoption continue: the 2000 crash killed companies, not the internet.
- For people.
- Losses in retirement accounts as private-credit and index funds hold AI exposure; hiring freezes in AI infrastructure; but tools keep getting cheaper.
13. Try it: does the revenue catch up?
Does the revenue catch up? (our model)
━ capex · ━ AI revenue · ┅ revenue needed to justify the capex (log scale)
- How it works.
- Spending and revenue grow at the rates you choose. The red line is the revenue needed for the spending to be sustainable (capex ÷ the ratio you choose; Bain uses about 25%).
- Defaults are rough 2026 values from this article; the model is ours and deliberately simple.
- Things to try.
- With the default inputs revenue catches up only in the late 2030s: the build-out is a very long bet.
- Set revenue growth to 25%: it never catches up within 15 years.
- Cut capex growth to 0%: revenue catches up around 2034. This is why markets react to capex guidance.
- Raise the ratio to 40% (more capex per dollar of revenue tolerated): the gap closes years earlier.
14. When would it happen?
- Nobody can time a bubble.
- The sources disagree; anyone giving a date is guessing.
- The window most often mentioned is late 2026 to 2027.
- Debt's share of capex is expected to peak in 2027 (Goldman).
- Depreciation from 2024–26 purchases reaches profit statements.
- 2027 spending plans, above $1T by several forecasts, must be justified.
- Bank of America flagged an "air pocket" for data-centre debt in 2026.
- A watch list with dates.
- Big-tech earnings in late January 2027 and their 2027 capex guidance.
- Nvidia's next quarterly results and any change in its financing commitments.
- Neocloud bond yields and credit-default spreads.
- The IMF April 2027 report and central-bank stability reports.
- Lab revenue disclosures, especially any IPO filings (prospectuses force comparable numbers).
The indicator dashboard to check every quarter
Our checklist; thresholds are rules of thumb, not tested predictors.
15. Preparing for it
- Founders and start-ups.
- Raise while money is available, and keep 24 months or more of runway.
- Rent compute; avoid long take-or-pay contracts; use two providers.
- Price on value delivered, not on model capability; that is what survives a price war (see The Maestro, Season 2, Part 3).
- Plan for GPU prices halving: if you are not locked in, a correction is a discount.
- Investors.
- Separate cash-rich platforms from leveraged infrastructure: they will behave very differently in a fall.
- Watch debt and customer concentration more than headline growth.
- Keep dry powder: the best buying in past bubbles came after the crash.
- Workers and students.
- Skills in using, evaluating and checking AI keep their value either way.
- Roles that depend on one heavily funded employer carry more risk than the skills themselves.
- Companies adopting AI.
- Avoid locking into one vendor's pricing; measure real productivity gains (the 2025 MIT report found most pilots showed no profit impact).
- Policymakers.
- Require disclosure of vendor financing and off-balance-sheet AI debt.
- Stress-test banks and private credit with an AI scenario, as the IMF did.
- Plan grid investment on realistic demand, with contracts that protect ratepayers if data centres are cancelled.
- Treat AI like the railways and fibre: transformative, probably overbuilt in places, and financed in ways that can break. Plan for a correction without betting on one.
16. What would change the picture
| Signal | Points to "not a bubble" | Points to "bubble" |
|---|---|---|
| AI revenue growth | Stays above 40–50% a year for several years | Slows sharply while capex keeps rising |
| Financing | Debt share stabilises; vendor financing shrinks | More guarantees, SPVs and private credit |
| Prices | Token prices fall slower than usage grows | Price wars outrun volume |
| Utilisation | New capacity stays full | Idle GPUs, falling rental prices |
| Disclosure | Labs report comparable, audited revenue | Revenue restated or definitions shift |
17. Reading the numbers
- What is solid.
- Company results and filings as reported by major outlets (CNBC, Fortune, Axios, Bloomberg), the Sequoia essays, Nvidia's announcement.
- What is softer ⚠️.
- Aggregate capex totals, revenue-gap estimates and debt tallies come from trackers and blogs that disagree with each other; we show ranges, not single numbers.
- IMF and Bank of England findings here are from press coverage of their reports; read the original reports before quoting them.
- What is ours.
- The bubble-stage placement, trigger ranking, scenarios, dashboard thresholds and the calculator are our models.
- Disclosure.
- Claude, made by Anthropic, helps write this site. Anthropic is one of the AI labs whose revenue and financing are discussed here.
- Sources
- Sequoia: AI's $600B question
- Axios: banks pull back from AI debt
- The Register: IMF warning
- Startup Fortune: IMF report
- TechRadar: Bank of England
- Fortune: earnings report card
- Fortune: BofA air pocket
- CNBC: Nvidia $105B Ohio
- Axios: circular concerns
- phys.org: the circular money machine
- Fortune: Nvidia and private capital
- CNBC: OpenAI revenue report
- Axios: OpenAI ARR
- Bloomberg: revenue accounting
- Understanding AI: still not a bubble
- Forbes: the gap widens
- Yahoo Finance: 102% of cloud revenue ⚠️
- Statista: big-tech capex
- CNBC: DeepSeek and Nvidia
- Fortune: MIT gen-AI pilots
- Wikipedia: AI bubble ⚠️
- Seeking Alpha: warnings intensify ⚠️