AI Bubble: Is AI a Bubble? The Numbers on Capex, Debt and Valuations, Tracked

Theme hub · AI bubble · Updated 21 NOVEMBER 2026

Not investment advice. This page is for information and education only; it is not a recommendation to buy or sell any security. Figures marked “our arithmetic” are our own calculations from the cited sources.

$202.0BBig-four capex in one quarter (Sep 2026, 10-Qs)
$380B+Asset-backed debt sold in 2026 (record; Bloomberg data)
>$4TBig tech off-balance-sheet commitments (ECB, Q2 2026)
−42.5%Chip index from its June high (20 Nov close)
10.7xNvidia, next-year earnings (10 Nov close; our arithmetic)
>3%of US GDP a year in AI investment needs (ECB; internet boom 1.10%)

Is AI a bubble? The short answer: the spending is real and still rising, but it is now running ahead of the cash that pays for it, and more of it is borrowed. The four biggest US hyperscalers spent $202.0 billion on property and equipment in the September 2026 quarter, more than their operating cash flow; asset-backed debt sales hit a post-crisis record of more than $380 billion; and the ECB counts more than USD 4 trillion of big tech commitments that sit off balance sheets. Chip stocks have already fallen 42.5% from their June high. On the other side, Nvidia sold $115.8 billion in a quarter and guides $130.0 billion, and trades at only 10.7 times next year’s earnings (10 November). Central bankers are split: the ECB and the Bank of England warn, Fed Governor Waller sees nothing bubble-like. This page tracks the numbers that will settle it, each one dated and sourced.

Latest

  • 20 Nov
    The chip index falls 17.1% in a week. The PHLX Semiconductor index fell 17.1% in a week to 8,408.59, 42.5% below its 22 June high close, and the Nasdaq-100 closed 19.8% below its 6 October high (Nasdaq closes; our arithmetic). CNBC reported that “the S&P 500 and Dow entered a correction, and the Nasdaq wiped out its 2026 gain”. Our story
  • 20 Nov
    A record in asset-backed debt. Asset-backed securities sales passed $380 billion in 2026, above the more than $367 billion of 2025 and a post-crisis record for a third straight year, on Bloomberg-compiled data; Bloomberg tied the run to “builders of artificial intelligence infrastructure”. The AI share is not given. Our story
  • 20 Nov
    All 16 AI stocks fell after Nvidia’s CEO exit. From the 11 November close to the morning of 20 November Nvidia fell 18.9%, from $173.73 to $140.87, about $792 billion of market cap; the median move across the 16 stocks we track was −15.6%, with CoreWeave −29.7% (Nasdaq data; our arithmetic). Our story
  • 19 Nov
    SpaceX borrows $32 billion off balance sheet. SpaceX sold $32 billion of bonds to buy AI chips at yields above 8%, part of a financing of more than $40 billion, in a deal Bloomberg said was designed to keep the borrowing off its balance sheet. Our story
AI investment needs of more than 3% of US GDP a year in 2025-2032, against about 2.24% for the Northern Pacific Railway in 1870-1890 and 1.10% for the internet boom in 1996-2003, per the ECB
The ECB’s November 2026 Financial Stability Review puts AI investment needs at more than 3% of US GDP a year between 2025 and 2032, more than 2.7 times the internet boom’s share and more than 1.3 times the Northern Pacific Railway’s (our arithmetic). History of bubbles: the dot-com bubble, tulip mania, the South Sea bubble.

The bull case vs the bear case, in one box

It is a bubble (or close)

  • Capex outran cash: the big four’s $202.0B quarter exceeded their operating cash flow; only Microsoft’s cash covered its capex. Alphabet spent 121% of its operating cash.
  • Free cash flow at Alphabet, Amazon, Meta, Microsoft and Oracle fell about 84% to $32.7B over four quarters (Hyundai Research Institute).
  • More is borrowed and kept off balance sheets: record $380B+ of ABS, SpaceX’s $32B above 8%, Oracle’s $288B of unstarted leases, >$4T of commitments (ECB).
  • Rich prices for new listings: Anthropic at 18.9 times its run-rate revenue at its first close.
  • Gensler: “It’s already a bubble.”

It is not (yet)

  • Demand is still rising: Nvidia’s revenue was $115.8B in its third quarter and it guides $130.0B.
  • Capacity is sold ahead: Micron’s CEO says 75% of its 2027 supply is sold; Jassy says AWS’s big chunks of capacity are “largely reserved already in 27”.
  • Cloud is growing fast: AWS over 39% ($184B run rate), Azure 48% to $33,165 million a quarter.
  • The leader is not expensive: Nvidia at 10.7x next year’s earnings (10 Nov) and 13.3x its next four quarters, below the S&P 500’s 18.3x (16 Nov).
  • Waller: nothing “bubble-like” in how AI is priced.
Where both sides meet: as our newsroom read the week Anthropic listed, the strain is in how AI is financed rather than in demand for its inputs. The bubble question has moved from the stock market to balance sheets. The 10 numbers from that week

What central bankers, regulators and investors say

Verbatim and dated, newest first. Each links to the newsroom piece that reports it and names its source.

  • Christine LagardeECB president, Frankfurt keynote, 20 Nov 2026Nuanced

    Wants public money, “including from the EU budget, taking the first losses and drawing in private capital” for chip plants and data centres. She cited a KIRA Center estimate (not an ECB figure) of around €1.3 trillion to raise Europe’s share of world AI computing capacity from around 5% to 15% by 2030. Source

  • Austan GoolsbeeChicago Fed president, per Reuters, 19 Nov 2026Warns

    “The AI boom is contributing to the potential of overheating in the economy.” “The Fed’s job is to prevent this thing from overheating the economy.” On Anthropic’s listing: “A $1.5 trillion IPO. The question of what people are going to do with the money now that they have some access to that liquidity is going to become a macro-relevant statistic.” Source

  • Christopher WallerFed governor, per FirstSquawk and Walter Bloomberg headlines, 18 Nov 2026Not a bubble

    There is nothing “bubble-like” in how AI is being priced. (He also urged a half-point rate hike in December, per Bloomberg.) Source

  • European Central BankFinancial Stability Review, 18 Nov 2026Warns

    “The rapid shift towards debt finance invokes historical parallels with the dotcom and subprime bubbles and could lead to spillovers affecting euro area financial stability.” And: “the maturity of hyperscalers’ outstanding debt exceeds the economic lifetimes of a substantial share of the physical capital funded by this debt, especially in the form of advanced chips.” Source

  • Gary Genslerformer SEC chair, to CNN, 16 Nov 2026Warns

    “It’s already a bubble.” “The only question is if it’s a soft recession or something more dramatic.” On who breaks first: “the major companies will survive. It’s these neocloud companies where we’re going to see some breakage and bankruptcies.” Source

  • Ed YardeniYardeni Research, to CNN, 16 Nov 2026Nuanced

    Sees “a self-correcting bubble” rather than one that bursts. Source

  • Boris VujčićECB vice-president, Euro Finance Week, 16 Nov 2026Warns

    IPO volumes are “now on track to exceed the dot-com bubble high, relative to US GDP”. Source

  • Andrew BaileyBank of England governor, Zurich speech, 9 Nov 2026Warns

    “Bubble spotting is an imprecise science. But, let’s recognise that such a risk exists, which we should do.” One of his four ways it could burst: “everyone is priced to be a winner, but history tends to tell us that this is not the case. Remember Netscape.” He added: “I am an optimist on AI.” Source

  • Brett RyanDeutsche Bank Securities economist, to Bloomberg, 5 Nov 2026Nuanced

    “If some of these moratoriums or regulatory pressures slow some of this buildout, it almost prevents you from blowing too big of a bubble.” Source

  • David WoodcockSEC enforcement director, to Reuters, 30 Oct 2026Nuanced

    As the SEC examines AI companies’ disclosures: “Our principal concern is that people are making accurate disclosures about material information so that investors know what they’re getting into.” Source

1. Capex vs revenue: is the spending paying for itself?

Amazon, Alphabet, Microsoft and Meta spent $202.0 billion on property and equipment in their September 2026 quarters, from their own 10-Qs: about $808 billion a year at that pace (our arithmetic). Amazon’s quarter was 2.2 times its level a year earlier. Not all of it is for AI; the filings do not split it out. With SpaceX, the five companies’ own 2026 outlooks add up to about $803 billion to $828 billion.

Capital spending in the September 2026 quarter: Amazon $77.3B, Alphabet $51.4B, Microsoft $39.4B, Meta $33.9B, $202.0B in total
Against about $175.5 billion of operating cash flow in the same quarter, capex was higher; only Microsoft’s operating cash covered its capex. Alphabet’s capex was 121% of its operating cash flow (49% a year earlier); Microsoft’s 68% (43%). Running tally: Big Tech AI capex and cash tracker.

The revenue side. Amazon’s CEO Andy Jassy said on 17 November that Amazon’s AI revenue is “over $25 billion” three years into the generative AI wave; he did not say how it is counted. Amazon’s own 2026 cash capex guide of about $220 billion is 8.8 times that. AWS grew over 39% in the third quarter to a $184 billion run rate, and Microsoft’s Azure grew 48% to $33,165 million in the quarter. The clearest AI revenue line is the seller of the chips: Nvidia reported $115.8 billion of revenue in its third fiscal quarter and guided $130.0 billion for the fourth.

Amazon: 2026 cash capex guide about $220B, AWS annualized run rate $184B, AI revenue over $25B per CEO Andy Jassy
The $25 billion is about 13.6% of AWS’s $184 billion run rate (our arithmetic). The Hyundai Research Institute, a Seoul think tank, says free cash flow at Alphabet, Amazon, Meta, Microsoft and Oracle fell to $32.7 billion over four quarters from $201.1 billion, projects their AI capex rising from $789.6 billion this year to $1.3079 trillion in 2029, and sees a possible correction around 2028.

2. Debt: how much of the build-out is borrowed

The ECB says cumulative capex by US big tech “is forecast to reach about USD 5.5 trillion by 2030, although estimates vary widely”, and that free cash flows “turned negative for most US hyperscalers in 2026”. The gap is being filled with bonds, asset-backed debt and leases.

Asset-backed securities sales: more than $337B in 2024, more than $367B in 2025, more than $380B in 2026 to 20 November
Bloomberg-compiled data as reported by Bloomberg on 20 November 2026; the bonds are backed by assets including data center leases. KBRA projects data center ABS to grow 29.4% in 2027 against 3.2% for the US ABS market. Deal by deal: AI data center debt tracker.
  • SpaceX: $32 billion of bonds to buy AI chips, yields above 8%, part of a financing of more than $40 billion, designed to keep the borrowing off its balance sheet (Bloomberg). Our story
  • Oracle: $288 billion of data center leases signed but not yet started, for 15 to 19 years, not yet on its balance sheet: about 4.6 times the $63.2 billion of lease payments already on its books (10-Q). Its capex was $28.5 billion in the quarter to 31 August, against $8.5 billion a year earlier. Oracle stock hub
  • Bonds: Alphabet, Amazon, Meta, Microsoft and Oracle issued $182 billion of corporate bonds from January to October, about double the $93.25 billion of all of 2025 (Hyundai Research Institute).
  • Off balance sheet: the ECB puts big tech’s commitments at more than USD 4 trillion as of the second quarter of 2026, “debt-like in economic terms” but not liabilities under US accounting. A Fed report cited by Governor Barr puts the seven largest AI firms’ potential maximum off-balance-sheet obligations above $3 trillion; a report by Americans for Financial Reform and the Open Markets Institute puts nine tech giants’ at least $3.1 trillion (citing a Wall Street Journal calculation), against $2.3 trillion for subprime at its 2007 peak. The estimates cover different groups of firms and are not added. How Big Tech finances data centres off balance sheet
  • The pipeline: the Bank of England’s Financial Policy Committee cites a JP Morgan estimate of about $4.1 trillion of debt-financed AI capex between 2026 and 2030, about $820 billion a year, and Morgan Stanley’s estimate of about $450 billion of AI-related debt issued by early September.

3. Valuations: expensive or not?

Multiples depend on which earnings they are set against, so every figure here is dated. On the 10 November close and Nasdaq’s consensus for each company’s next fiscal year, Nvidia traded at 10.7 times, Broadcom at 19.0 times and AMD at 32.1 times (our arithmetic). FactSet’s next-12-month multiple for Nvidia was about 11.7 times, its lowest since October 2011, as TradingKey reported it.

Forward price-to-earnings on the 10 Nov 2026 close and next fiscal year consensus: Nvidia 10.7x, Broadcom 19.0x, AMD 32.1x
On 16 November Nvidia traded at 13.3 times the sum of its next four quarters of consensus earnings, below the S&P 500’s 18.3 times (FactSet; the index’s 5-year average is 19.8). After its results, at $161.44 premarket on 18 November, it was 15.1 times its third-quarter earnings annualised, against 31.8 times a year earlier (our arithmetic). Method: forward P/E explained.

Where the high multiples are. At its first close of $97.17, Anthropic was worth $1.28 trillion, 18.9 times its third-quarter revenue run-rate (our arithmetic), against an $83 IPO price. Anthropic stock hub · Nvidia stock hub

4. The market: what has already fallen

Drawdowns from the 2026 high close on 20 Nov 2026: PHLX Semiconductor index 42.5%, Nasdaq-100 19.8%, SPDR S&P 500 ETF 11.4%, SPDR Dow ETF 11.4%
The chip index fell 17.1% in the week to 20 November, from 10,146.32 to 8,408.59, and is 42.5% below its 22 June high close of 14,634.72; it is still 18.7% above its 31 December 2025 close of 7,083.13. The S&P 500 and Dow are measured through the SPDR ETFs ($690.43 and $480.86 at the close). Daily prices: AI stocks market cap tracker.

Since the last close before Nvidia disclosed its chief executive’s resignation (11 November), Nvidia fell 18.9% to $140.87 by the morning of 20 November, about $792 billion of market cap ($4.19 trillion to $3.39 trillion; our arithmetic on its 10-Q share count). All 16 AI stocks we track fell, median −15.6%, led by CoreWeave at −29.7%. At the 19 November close, 8 of 18 AI-trade stocks were more than 50% below their 52-week highs (our count of Nasdaq data). CoreWeave stock hub

What would change the answer

  1. Capex against operating cash. In the September quarter the big four’s capex ran ahead of about $175.5 billion of operating cash flow. Their own outlooks point higher: Amazon, Alphabet, Microsoft, Meta and SpaceX guided $241.6 billion to $270.0 billion for the December quarter, against $228.3 billion in September.
  2. Whether debt still places. SpaceX’s $32 billion priced above 8%. KBRA expects data center ABS to grow 29.4% in 2027 against 3.2% for the whole market. A deal that fails to place, or prices much wider, would be the first crack in the financing.
  3. 1 December: House Democrats’ deadline for the Commerce Department to answer on Nvidia contacts in the Megaspeed probe, per Reuters.
  4. 8-9 December: the Fed meeting. Reuters says a rate hike is widely expected; Waller wants a half point. Higher rates raise the cost of every borrowed dollar in the build-out.
  5. 15 December: up to 303 million Anthropic shares can become eligible for sale under its 424B4 lock-up schedule: the first test of whether the newly listed AI wealth Goolsbee is watching meets buyers.
  6. Nvidia’s $130.0 billion. Its fourth-quarter revenue guide, after $115.8 billion in the third quarter. Demand guides being met is the bull case in one number.
  7. Oracle’s leases. $288 billion of signed data center leases start between Oracle’s second fiscal quarter of 2027 and fiscal 2029, for 15 to 19 years each.
  8. Polymarket’s trigger list. Its “AI industry downturn by 30 June 2027” market (26.5% on 7 November) needs three of six events in one 90-day window, such as Nvidia closing 50% below its all-time high.

Dates from Reuters and Anthropic’s 424B4 as reported by our newsroom on 20 November 2026; guides from company filings. We update this list as each date passes.

AI bubble FAQ

Is AI a bubble?

Nobody can prove it either way yet, and the people paid to watch it disagree. Former SEC chair Gary Gensler told CNN on 16 November 2026: “It’s already a bubble.” Fed Governor Christopher Waller said on 18 November there is nothing “bubble-like” in how AI is being priced. The numbers on this page are the test: spending is running ahead of cash (the four biggest US hyperscalers spent $202.0 billion on capex in one quarter, more than their operating cash flow), more of it is borrowed, and chip stocks have fallen hard, while demand numbers such as Nvidia’s $115.8 billion quarter are still rising.

What did Bank of England Governor Andrew Bailey say about the AI bubble?

In a speech in Zurich published on 9 November 2026, Bailey said: “Bubble spotting is an imprecise science. But, let’s recognise that such a risk exists, which we should do”. He set out four ways a bubble could burst, including that “everyone is priced to be a winner, but history tends to tell us that this is not the case. Remember Netscape”. He also said “I am an optimist on AI”. The published text does not use the word “crash”.

What did the ECB say about AI debt?

The European Central Bank’s November 2026 Financial Stability Review said: “The rapid shift towards debt finance invokes historical parallels with the dotcom and subprime bubbles and could lead to spillovers affecting euro area financial stability”. It estimates big tech’s off-balance-sheet commitments at more than USD 4 trillion as of the second quarter of 2026, and puts AI investment needs at more than 3% of US GDP a year between 2025 and 2032, against 1.10% for the internet boom.

What did the Fed’s Goolsbee say about AI?

Chicago Fed President Austan Goolsbee said on 19 November 2026, per Reuters: “The AI boom is contributing to the potential of overheating in the economy.” and “The Fed’s job is to prevent this thing from overheating the economy”. He did not cite a capex figure; by our tally of their 10-Qs the four biggest US hyperscalers spent $202.0 billion on property and equipment in the September quarter.

How much is Big Tech spending on AI?

Amazon, Alphabet, Microsoft and Meta spent $202.0 billion on property and equipment in the September 2026 quarter ($77.3 billion, $51.4 billion, $39.4 billion and $33.9 billion), about $808 billion a year at that pace (our arithmetic); the filings do not split out how much is AI. With SpaceX, the five companies’ own 2026 outlooks add up to about $803 billion to $828 billion. The ECB cites a forecast of about USD 5.5 trillion of cumulative US big tech capex by 2030.

Is AI spending paying for itself?

Not yet in cash. In the quarter to 30 September 2026 the big four’s capex exceeded their operating cash flow of about $175.5 billion, and only Microsoft’s cash covered its spending. Alphabet spent 121% of its operating cash flow on capex. Amazon’s CEO Andy Jassy puts its AI revenue at “over $25 billion”; its 2026 capex guide of about $220 billion is 8.8 times that (our arithmetic).

How is the AI build-out being financed?

Increasingly with debt and leases. Asset-backed securities sales passed $380 billion in 2026 (Bloomberg-compiled data), a post-crisis record for a third straight year that Bloomberg tied partly to AI infrastructure builders. SpaceX sold $32 billion of chip-backed bonds above 8%, structured off its balance sheet. Oracle has $288 billion of data center leases signed but not yet started. A JP Morgan estimate cited by the Bank of England puts debt-financed AI capex at about $4.1 trillion between 2026 and 2030.

Are AI stocks overvalued?

On forward earnings the biggest chip stock is not expensive. On the 10 November close Nvidia traded at 10.7 times next fiscal year’s consensus earnings, Broadcom at 19.0 times and AMD at 32.1 times (our arithmetic on Nasdaq data); FactSet put the S&P 500 at 18.3 times on 13 November. The richest multiples are in newly listed and private names: at its first close Anthropic was worth 18.9 times its third-quarter revenue run-rate.

How far have AI and chip stocks fallen?

The PHLX Semiconductor index closed at 8,408.59 on 20 November 2026, down 17.1% in a week and 42.5% below its 22 June high close; the Nasdaq-100 was 19.8% below its 6 October high. From the 11 November close to the morning of 20 November Nvidia fell 18.9%, about $792 billion of market cap, and all 16 AI stocks we track fell (median −15.6%).

What would show the AI bubble is bursting?

Bailey’s four routes are a useful checklist: markets overvaluing good returns, returns arriving later than priced, too many companies priced as winners, and a loss of public trust. In numbers, watch whether capex keeps outrunning operating cash, whether debt deals start failing to place, whether lock-ups such as the up to 303 million Anthropic shares eligible on 15 December meet buyers, and whether demand guides such as Nvidia’s $130.0 billion for its fourth quarter are met.

All our AI bubble stories

Every piece our newsroom has published on the question, by theme, newest first. New pieces on the AI bubble link back here.

Related hubs and trackers

Sources and method. Every figure and quote on this page comes from a FourWeekMBA newsroom piece, linked, that names its own primary source: company 10-Qs and earnings releases (Amazon, Alphabet, Microsoft, Meta, Oracle, Nvidia), Nasdaq closing prices and analyst consensus, the ECB Financial Stability Review (November 2026) and speeches, the Bank of England’s speech and FPC record, Federal Reserve officials as reported by Reuters, Bloomberg and CNBC, Bloomberg-compiled ABS data, KBRA, the Hyundai Research Institute, CNN, FactSet and Polymarket. Quotes are verbatim as published by those sources and dated. Figures marked as ours are our arithmetic. Estimates by different bodies cover different groups of firms and are never added together. This page is not investment advice.

Scroll to Top
FourWeekMBA