Is AI a bubble? The evidence for and against, with the sources

What would make AI a bubble?

A bubble is a price that is higher than the future cash flows of an asset can justify. Nobody can measure those future cash flows today, so analysts look at five kinds of evidence instead.

1. Spending. How much is being invested, and how that compares with past booms.
2. Financing. Whether the spending is paid from cash or from debt and share sales.
3. Revenue. Whether customers pay enough to repay the spending.
4. Valuations. What investors pay for the earnings they expect.
5. Behavior. Whether suppliers fund their own customers, which can make demand look larger than it is.

The sections below take them in turn. Each shows evidence that supports the bubble case and evidence that does not.

How much is being spent?

A lot, and the central banks have said so. The four largest U.S. cloud builders spent $201.9 billion on property and equipment in the quarter to September 30, 2026. The detail is in Big Tech's AI spending in Q3 2026.

The Fed staff note adds a point on each side. Another boom is possible if the productivity gains spread. Overinvestment is also possible, and it concludes that investment should not necessarily be curbed in advance.

How much of U.S. economic growth comes from AI is disputed. Estimates range from 0.3 to 1.39 percentage points. They cover different periods, and the sources do not say how they treat imports or which categories they count. Capital Economics, reported by Fortune, put the 2025 effect at 0.3. The BIS said AI-related investment added roughly one point to 2025 growth, by some estimates. Jefferies, reported by ANI, said 1.39 points, or 67% of growth, in the four quarters to the third quarter of 2026. The Bureau of Economic Analysis reports that real GDP grew at a 3.4% annual rate in the third quarter and that, within equipment, the increase was led by information processing equipment. It does not publish an AI line.

How is the spending paid for?

Increasingly with borrowing. In 2026 the financing has changed from the cash-funded spending of earlier years.

• Bonds. U.S. investment-grade companies issued about $1.88 trillion through November 4, a record, and hyperscalers raised more than $370.7 billion of it, about a fifth, according to Reuters using LSEG data.
• Free cash flow. Alphabet's was negative $9.1 billion in the third quarter and Meta's negative $2.1 billion. Microsoft's was positive. Amazon reports only a trailing twelve months, and it was negative $42.4 billion.
• Share sales. Alphabet announced equity offerings of $80 billion in June, including a $10 billion private placement with a Berkshire Hathaway affiliate. About $30 billion of the $40 billion at-the-market program is meant to cover tax on employee stock awards. The rest is for general corporate purposes, including AI capital spending.
• Structures outside the balance sheet. Meta holds a 20% interest in the venture that builds its Louisiana data center and does not consolidate it. Its filing shows maximum residual value guarantees of $41.49 billion across its Louisiana and Texas ventures, and says payments are not probable.
• Credit signals. Reuters reported on July 29 that new hyperscaler bonds paid a median concession of 12 basis points so far in 2026, against 2.25 in 2025. On October 15 it reported that CoreWeave's credit default swap spread had passed 1,000 basis points.

The other side is also on the record. A fund manager told Reuters the market has absorbed the issuance, and the Bank of England said it found little evidence that AI borrowing is crowding out other borrowers. It also said the complexity and opacity of some debt structures could raise risks to financial stability.

Are the deals circular?

Some arrangements have a supplier funding or backing its own customer. NVIDIA addressed this itself on its August 26 call, saying that some will call its credit support circular financing and that it sees it differently.

The BIS called circular financing the most prominent of the private arrangements linking hyperscalers, chip makers and AI labs, and said the terms of such deals are typically poorly disclosed. The IMF listed circular financing among the factors that could worsen an AI-related correction. NVIDIA's reply is that its platform can be redeployed to other customers, and that the compute it ships is consumed by investment-grade customers or those backed by one.

Where is the revenue?

It is growing quickly, and it is hard to compare. The figures below measure different things.

• Anthropic. Revenue of $33.2 billion in the first nine months of 2026, against $2.5 billion a year earlier, and run-rate revenue of about $81 billion at October 31, 2026. It also reports a large net loss, most of it non-cash. These are in its November 10 filing.
• OpenAI. Its chief financial officer said on CNBC that the business grew 70% quarter over quarter in the third quarter, without naming the measure. Spezia found no audited revenue figure.
• Enterprise spending. Menlo Ventures estimates enterprise spending on deploying AI at $176.8 billion by September 2026, against $33 billion in 2025. The two periods are not clearly like for like, Menlo restated its 2025 figure, and the numbers are an investor's estimates.
• Adoption. The Census Bureau found 19.8% of U.S. businesses used AI as of May 3, 2026, and 24.6% in its October 5 to 18 survey, with the same wording. The question changed in November 2025, so earlier figures are not strictly comparable.
• Returns. A 2025 MIT NANDA report found 95% of organizations getting zero measured return. It is a preliminary study of 300 public initiatives, 52 interviews and 153 survey responses, and it defines success as moving past the pilot stage with measurable results.

The cloud companies report contracted backlog, which is revenue customers have agreed to pay later.

The three figures are defined differently, so they are not like for like. Backlog is a promise, not cash. Amazon's filing describes expanded commitments from OpenAI and Anthropic, two labs funded by investors and suppliers, but does not say how much of its backlog they make up. Google also said its model interfaces now process nearly 32 billion tokens a minute, more than four times a year earlier.

What do valuations say?

On the index as a whole, earnings have grown faster than prices this year. FactSet reports the S&P 500's forward price-to-earnings ratio at 17.7 on November 5, down from 22.0 at the end of 2025. The index price rose 7.4% so far this year while forward earnings estimates rose 33.6%. The ratio is below its 5-year average of 19.8 and its 10-year average of 19.1, and above its 15-, 20- and 25-year averages of 17.6, 16.5 and 16.4. Earnings growth for the third quarter is 31.4%, and the information technology sector's ratio of 17.3 compares with a 25-year average of 20.0.

The concentration is the concern on this side. As of May 18, Bloomberg Intelligence, reported by Bloomberg Linea, counted 44 AI-linked companies at about 45% of the index's market value. It said they accounted for 71.1% of the index's earnings growth, without giving the period. RBC Wealth Management put the ten largest stocks at nearly 41% of index weight at the end of 2025, against about 32% of expected earnings.

Private valuations are a separate matter. OpenAI's round is reported at $1.25 trillion, and Anthropic's May round was announced at $965 billion. Reuters reported on November 6 that some investors value Anthropic's planned listing at about $1.5 trillion, below earlier expectations of about $2 trillion.

What have central banks and analysts said?

None of these sources predicts a crash on a date. The warnings are about conditions that could make a correction worse. The most optimistic figure in the table, NVIDIA's 70% revenue growth, comes from a company that sells into the demand.

What would change the answer?

These are figures that bear on the question. Company figures arrive in quarterly filings, and spreads are market prices.

• NVIDIA's next report, on November 17, 2026, shows whether data center revenue keeps growing.
• Alphabet's fourth-quarter call is where it said it would give detail on its 2027 spending.
• Free cash flow at the cloud companies shows how much of the spending is borrowed.
• Spreads on hyperscaler bonds and on the most leveraged AI firms show how lenders are pricing the risk.
• Backlog turning into revenue shows whether contracted demand is being paid for.
• Lab revenue against commitments shows whether the customers can pay. Anthropic's filing gives audited figures for 2025 and unaudited figures for the nine months to September 30, 2026.

Spezia is a research product, not an adviser. Nothing here is investment advice or a forecast.