AI bubble

Together the four biggest AI spenders now borrow to build: their capex reached 115% of operating cash in the third quarter of 2026, up from 61% a year earlier.
Alphabet, Amazon, Microsoft and Meta spent $201.9 billion on property and equipment in the three months to 30 September 2026, $26.4 billion more than their operations brought in. A year earlier the same four companies had $61.7 billion left over after paying for their build-out.
That swing is the clearest answer yet to the question behind every AI bubble argument. High prices alone do not make a bubble. A bubble forms when spending runs ahead of the cash it can ever earn and the gap is filled with other people's money. For three years the largest buyers of AI chips paid for their data centers out of profits. In the third quarter of 2026 they stopped doing so, and the difference is now borrowed.
The Federal Reserve's financial stability report, published on 13 November, describes the same shift from the lenders' side. AI was the risk most often named by the 23 market professionals surveyed for the report in September and October, ahead of a correction in risk assets and higher long-term interest rates. Debt raised to finance advanced computing reached more than $680 billion in the first ten months of 2026, about five times the $130 billion raised in all of 2023.
What makes a boom a bubble
Every investment boom overbuilds. Railways in the nineteenth century, fiber-optic networks in the 1990s and data centers today follow the same arc: a real technology, a race to own capacity, and a stretch when capacity runs ahead of demand. What separates a boom that leaves useful assets behind from a bubble that destroys capital is who carries the bill when revenue arrives late.
One ratio captures it. Divide what a company spends on property and equipment by the cash its operations generate. Below 100%, the build-out pays for itself. Above 100%, someone else is financing it: bondholders, banks, landlords on long leases, or shareholders through dilution. Call it the cash-cover line.
For the four largest AI spenders together, capital spending took 61% of operating cash in the third quarter of 2025. A year later it took 115%. Spending on property and equipment roughly doubled, while cash from operations rose by about a tenth.
Crossing the line is not a crisis for companies of this size. They hold large cash balances, earn some of the highest profits in the world and can borrow cheaply. What changes is the chain of risk. As long as the platforms paid in cash, a disappointment in AI revenue would have hit their shareholders and stopped there. Once the build-out runs on debt, leases and guarantees, a disappointment travels to whoever holds those claims.
How much the biggest AI spenders are investing
| Company | Capex 2025 | Capex 2026 | Cash from operations 2026 | Left after capex 2026 |
|---|---|---|---|---|
| Alphabet | 24.0 | 51.4 | 42.3 | −9.1 |
| Amazon | 35.1 | 77.3 | 42.4 | −34.8 |
| Microsoft | 19.4 | 39.4 | 58.0 | 18.6 |
| Meta | 18.8 | 34.0 | 32.8 | −1.2 |
| Total | 97.3 | 201.9 | 175.5 | −26.4 |
As of 30 September 2026, from each company's quarterly filing. Capex is cash paid for property and equipment; Microsoft's figure leaves out data centers taken on finance leases, and the July to September quarter is its fiscal first quarter. Left after capex is our calculation. Totals differ from the sum of the rows because of rounding.
Amazon's shortfall alone is larger than the gap for all four companies combined. Microsoft is the only one of the four still covering its spending from operations, and its true commitment is larger than the table shows because part of its capacity arrives through finance leases rather than purchases. Meta crossed the line in the summer, a turn that became visible when Meta's spending began to outrun its cash.
Over the first nine months of 2026 the four companies spent $496.7 billion on property and equipment, against $257.4 billion in the same period of 2025. The pace is still rising: the four spent $36.9 billion more in the third quarter than in the second.
Where the money for the build-out comes from now
The Federal Reserve's report shows how fast the financing has changed. Borrowing tied to advanced computing made up less than 6% of investment-grade bond issuance by non-financial companies in 2023 and about 31% in 2026. Outstanding public debt linked to the build-out exceeded $1.1 trillion in October 2026, and sizable private credit deals sit outside that figure, where the total is hard to measure.
The form of the debt matters as much as the amount. The mix is shifting from plain corporate bonds toward private debt, structured credit and project finance, often through special purpose vehicles, companies set up to own a single data center or a fleet of chips and to borrow against them. The platforms are also signing long leases, guaranteeing the debts of their partners and taking on other obligations that do not show up as debt on their balance sheets. The riskiest slices of these deals end up with high-yield bond funds, life insurers and private debt funds, and the largest US banks have close to $1 trillion of exposure to large technology firms, data center developers, the vehicles that finance them and the utilities that power them.
Private markets add equity on top. AI companies took 82.7% of all US venture capital invested in the first nine months of 2026, the highest share on record, as record sums went to fewer hands. When one theme absorbs four of every five venture dollars, the venture market stops spreading risk and starts concentrating it. A large share of that money flows straight back to the build-out, because the biggest AI model developers rent their computing power from the same platforms.
How the AI bubble compares with the dot-com crash
The comparison everyone reaches for is 2000. The Nasdaq Composite closed at 5,048.62 on 10 March 2000 and fell to 1,114.11 on 9 October 2002, a loss of 77.9%. It took until April 2015 to regain its peak. The cyclically adjusted price-to-earnings ratio of the S&P 500, which divides prices by average inflation-adjusted earnings over the previous ten years, stood just under 40 on 13 November 2026, against its record of 44.2 in December 1999.
Valuations rhyme, but the spenders do not. In 2000 many of the most expensive companies had no profits, and the money they burned came from investors. The four platforms in the table are among the most profitable businesses ever built, and until this year they funded their own expansion.
The closer parallel is the telecom boom that ran alongside the dot-com mania. Carriers borrowed heavily to lay fiber in anticipation of traffic that arrived years later than planned. When the cash flow failed to come in time, the bust was a credit event: Global Crossing filed for bankruptcy in January 2002 and WorldCom in July 2002. The Federal Reserve now compares the two cycles directly. Measured against the size of the economy, the cumulative capital spending of the largest cloud companies since 2020 has tracked the telecom industry's spending after 1995 closely, and it has been accelerating, which puts it on course to overtake the telecom cycle.
The fiber did not disappear. It carried the next decade of internet growth at a fraction of its original cost, and the companies that built on cheap bandwidth created more value than the carriers that laid it. That is the most useful lesson for this cycle. The data centers will be used. The open question is who owns them, and at what price, when demand catches up.
Read together, the numbers say AI is a boom in demand and a bubble in financing. If it bursts, it will burst in credit, not in the share prices of the four platforms. The weak links are the borrowers without a platform's cash flow behind them: specialist cloud providers that rent out chips, a model public investors first priced in a cut-price listing in 2025, data center developers and the vehicles that lend against hardware whose value falls with every new chip generation.
What would show the AI bubble is bursting
Four signals are worth watching, and each will be visible within a year.
- The cash-cover line. If capital spending at the four platforms stays above operating cash flow through mid-2027 while their cloud revenue growth slows, the build-out is running on credit rather than on demand.
- Lending terms for data centers. Wider spreads on bonds and loans backed by data centers and chips would show lenders demanding more for the same risk before the stock market reacts.
- Rental prices for AI chips. Falling prices for renting computing power by the hour would mean supply has overtaken demand, and the hardware pledged as collateral is worth less.
- Leadership at the core supplier. Nvidia said on 12 November that Jensen Huang, who built the company over 33 years, had resigned as chief executive with immediate effect. Chief financial officer Colette Kress is running it on an interim basis while the board looks for a permanent successor, at the moment when its customers are committing record sums.
The nearest test is Anthropic's stock market listing. The company has set a range of $70 to $78 a share for its offering on Nasdaq under the ticker ANT, and the price public investors accept will show whether they will pay the multiples private investors paid for a frontier model developer. The range itself already sits below the last private multiple.
What the AI bubble means for investors in private companies
Investors in private AI companies do not own the platforms. They own the layers above and below them, and the financing shift splits those layers in two. One group depends on the build-out continuing: companies that sell to data centers, rent out chips or borrow against them. If credit tightens, they lose customers and collateral at the same time. The other group uses the capacity the build-out creates: companies whose customers pay for a finished product out of their own operating budgets. If credit tightens, their main input gets cheaper.
Fanaura invests in AI-native companies with its limited partners, and it applies to a private company the same test this piece applies to the platforms: does the customer pay from cash it earns, or from money raised for the build-out? That test, more than any multiple, separates the companies that will outlast a correction from those that depend on the cycle staying hot. It sits at the center of how Fanaura invests.
The four platforms report their results for the last quarter of 2026 in late January and early February 2027. If cash from operations closes the gap, the AI boom will have paid for itself, as the cloud boom of the last decade did. If the gap widens, the AI bubble question will be settled where the telecom one was, in the bond market rather than the stock market.
Invest alongside Fanaura as a limited partner. →
More from Fanaura
Invest alongside Fanaura as a limited partner
Fanaura is a private family office. We invest our own capital and the capital of our limited partners (LPs), usually through a separate SPV for each deal, in AI-native companies, games, technology and real estate across the US, the EU and Latin America. Accredited investors can talk to us about joining as an LP.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.