Global · AI economics
The AI revenue gap: what a year of capex needs to earn, and what the labs earn
Microsoft, Amazon, Alphabet, Meta and Oracle have guided about $825bn of capex for 2026 (sum of midpoints, company guidance to 2026-10-11). On Sequoia partner David Cahn's arithmetic (capex ÷ (1 − 50% margin)), and if that spending recurs each year, end customers need to pay $1.65tn a year. The latest reported run-rates of the AI labs we track sum to about $131bn a year (press and company figures, unaudited), 8% of it (our arithmetic).
Updated 11 October 2026 · data build 2026-10-10 · how we know · corrections
Lab revenue covers 8% of what a year of capex needs on Cahn's arithmetic
Default settings: 2026 capex guidance, all of it treated as AI, spending repeated every year, a 50% gross margin for the sellers. Our reading
Change the assumptions
Each step uses our own sourced numbers. Move the settings to see how the requirement and the gap change. Scenario
Capex base
Is the spending repeated? If capex stays at this level every year, each year's revenue has to cover a year of capex, however long the assets last (Cahn's view). If this were a one-off, its cost could be spread over the servers' 4–6 year life.
- $825bn2026 capex guidance, five companies Reported
Sum of guidance midpoints: Alphabet $200bn, Amazon $220bn, Meta $137.5bn, Microsoft $175bn, Oracle $92.5bn - $825bnOf which for AI Scenario
× 100%. Amazon's capex includes warehouses; all five also buy non-AI servers. - $825bnCost to earn back each year Scenario
÷ 1: the same spending recurs every year - $1.65tnRevenue needed each year Our reading
÷ (1 − 50% gross margin) - $131bnAI-lab revenue today Reported
Latest run-rates: US $125.2bn, China $4.4bn, Europe $1bn, Canada $0.2bn - 7.9%Share covered Our reading
step 5 ÷ step 4 - $1.52tnGap Our reading
step 4 − step 5 - 3.7Years to close at this growth Scenario
log(step 4 ÷ step 5) ÷ log(1 + 100%), holding capex flat
Source: Hyperscaler capex guidance and SEC capex (global_capex) · AI-lab run-rates (Epoch AI, CC BY; SpaceX 10-Q) · as of 2026-10-11 · Compiled by StackPath
What this arithmetic leaves in and out
The test is deliberately crude. Here is how each choice moves the answer. Editorial judgement
| Choice | Effect on the gap | Why |
|---|---|---|
| One year of capex vs one year of revenue | Right if spending recurs; too large if it stops | The five depreciate servers over roughly 5–6 years and buildings over longer (general knowledge, unchecked against each 10-K), so a single year's capex is earned back over several years. But if the same capex is repeated every year, the revenue needed each year equals Cahn's figure whatever the asset life. Cahn calls his number the lifetime revenue needed to justify one year of capex. |
| Lab revenue only | Makes the gap larger | Hyperscalers also sell AI directly (Microsoft Copilot, Google Gemini, Meta's ad tools) and do not report that revenue separately. |
| No power, staff or interest costs | Makes the gap smaller | Running the capacity costs more than the capex alone; Cahn's original version doubled GPU spend to cover the rest of the data centre and its running costs. |
| Capex flat when closing the gap | Makes it look closable sooner | The 5 companies' guidance sums to $825bn, against $566bn paid in the last four quarters (SEC); the requirement has risen each year. |
Source: StackPath method, after Cahn (Sequoia) · as of 2026-10-11 · Compiled by StackPath
Our reading. On Cahn's own settings lab revenue covers about 8% of the requirement. Only if 2026 were a one-off, spread over five years, would coverage rise to about 40%. The five companies' capex rose from $239bn in 2024 to $412bn in 2025 (SEC), so for now the spending is recurring and rising, which is Cahn's case. The gap is closing in proportion because lab revenue has grown faster than capex: Anthropic's run-rate rose about 13-fold in a year. Whether that pace holds is the question this page cannot answer. Lab deals and the suppliers that depend on them are on AI compute commitments.
Questions people ask
- What is AI's $600 billion (now $1.5 trillion) question?
- Sequoia partner David Cahn's estimate of the revenue the AI build-out needs to pay for itself: take a year of AI data-centre spending and ask what end-customer revenue would earn it back at a normal software-like margin. He put it at $600bn in 2024 and $1.5tn in July 2026.
- How much revenue does AI need to justify hyperscaler capex?
- On Cahn's shortcut with our numbers: ~$825bn of 2026 capex guidance ÷ (1 − 50% margin) = $1.65tn a year. If 2026's capex were a one-off spread over a 5-year life, it would need $330bn a year for five years; if it recurs every year, the yearly need stays at the higher figure. Both ignore power and staff costs and assume all of the capex is for AI.
- How much revenue do AI labs make?
- About $131bn a year summed across OpenAI, Anthropic, xAI, Mistral, Cohere and four Chinese labs, on their latest reported run-rates (see /global/ai-lab-economics). That excludes AI sold by the hyperscalers themselves (e.g. Copilot, Gemini subscriptions) and AI's effect on advertising, which are not disclosed separately.
- Is there an AI bubble?
- The arithmetic shows a gap, not a verdict. Reported run-rates grew about 3.8× at OpenAI and 13× at Anthropic in a year; if lab revenue kept doubling every year, it would reach $1.65tn in about 3.7 years with capex held flat (try the growth setting). If growth slows, or capex keeps rising, it would take longer. We report the numbers and leave the forecast to the reader.