Is AI a bubble? Do the arithmetic yourself
The honest answer is that nobody can tell you, because it depends on a number nobody has yet. What can be done is to work out what that number would have to be.
Key figures
Frontier AI revenue is real and compounding fast: one developer went from $9bn to $19bn annualised in about three months.
The open question is not whether AI earns money. It is whether it earns enough to cover what is being built for it.
At $600bn of annual capital spending written off over six years, the industry needs about $100bn a year of revenue simply to break even on depreciation.
The revenue side is not imaginary
One frontier developer's annualised revenue was reported at $9bn at the end of 2025, $14bn in February 2026 and $19bn in March 2026. In the autumn of 2025 the same company was projecting a rise from $4.7bn in 2025 to more than $15bn in 2026, and a competitor was projecting $13bn rising to $30bn.
Those are revenue run rates reported in the press, not audited accounts, and a run rate flatters a fast-growing business. But the direction is not seriously disputed by anyone, including the people who think this is a bubble. Demand exists.
The cost side is the question
What is disputed is the other side. The largest cloud and AI infrastructure companies have guided to capital spending in the hundreds of billions of dollars a year, and reported totals for 2026 vary between sources depending on which companies are counted and whether leases are included.
This page does not pick between those totals. It does something more useful: it shows what any of them would require.
The sum you can do yourself
Capital spending is not a loss the year it happens. It is written off over the life of the asset. So the test is not whether revenue covers capex in the same year, it is whether revenue covers the annual depreciation, plus the power, plus a return.
Six years is not an arbitrary choice. Microsoft extended the depreciable life of servers and network equipment from four years to six from fiscal 2023, and Alphabet made the same change in January 2023. The depreciation load is then capex divided by six:
| Annual capex | Depreciation over 6 years | Revenue needed to break even |
|---|---|---|
| $400bn | $67bn a year | $67bn a year, before power, staff or any profit |
| $500bn | $83bn a year | $83bn a year, before power, staff or any profit |
| $600bn | $100bn a year | $100bn a year, before power, staff or any profit |
| $700bn | $117bn a year | $117bn a year, before power, staff or any profit |
Pick whichever capex figure you believe and read across. That is the revenue the whole AI industry has to reach, every year, just to avoid destroying capital, before electricity bills and before anyone earns anything.
Two things this simplification leaves out, both of which cut against the optimistic reading. Capex that is still being built is not yet earning, so the revenue has to arrive later than the spending. And the six-year life is an accounting choice rather than a physical fact: Amazon shortened a subset of its servers back from six years to five with effect from January 2025, citing the pace of AI hardware change. Shorten the life and every number in that table gets larger.
What would actually settle it
A bubble is not high spending. It is spending that cannot be repaid. Railways, telecoms fibre and the dot-com build-outs all ended with infrastructure that was genuinely useful and investors who were genuinely wiped out, because the second thing does not require the first to be false.
So the question to watch is not whether AI is useful. It plainly is. It is whether revenue growth stays on a curve that reaches the break-even line in the table above before the assets bought in 2025 and 2026 are written down. On current published numbers that is not yet settled in either direction, and anyone telling you it is settled is selling something.
Sources and method
- It still doesn’t look like there’s an AI bubble, Timothy B. Lee, Understanding AI, 16 March 2026. Source of the $9bn, $14bn and $19bn annualised revenue figures and the autumn 2025 projections.
- Capital spending totals are deliberately not asserted here. Published 2026 estimates for the largest spenders vary with the companies included and the treatment of leases, so the table is given as a range for the reader to apply their own figure to.
- The depreciation arithmetic is our own and is stated in full: annual capex divided by a six-year asset life.
- Microsoft annual filings on EDGAR and the equivalent disclosures by the other operators. Microsoft extended servers and network equipment from four years to six from fiscal 2023, Alphabet made the same change in January 2023, and Amazon shortened a subset back from six years to five with effect from 1 January 2025, citing the pace of AI hardware change. Source of the six-year figure used in the table, and of the fact that one operator has already begun shortening it again.
- Energy and AI, International Energy Agency, 2025, for the scale of the electricity demand the spending is being built to serve: about 415 TWh in 2024 rising to roughly 945 TWh by 2030 in its base case.
- Every source here was opened and checked on 25 September 2026. Where two credible sources disagree, both are shown with their scope, rather than averaged into one number.