The chain of beliefs
Each link has to hold for the capex to earn that return. The figures move with the assumptions.
Follow the dollar
One year of spending in 2030, from end users down to the hyperscalers’ profit.
Depreciation: years into rates
Your useful life as an annual depreciation rate, against the lives the hyperscalers use for their assets.
Straight-line rate: 1 ÷ useful life. Shaded: CGAM’s 5–10 year range, 10–20% a year.
Blend servers and buildings
One life has to cover servers and buildings together, so it sits between the two. Split the capex to see the combined rate.
Sources: company filings on useful lives (Microsoft FY2023 10-K; Alphabet 2023 10-K; Oracle FY2025 10-K; Meta 2024 10-K; Amazon 2024 and 2025 10-Ks; CoreWeave; Nebius 2025 20-F); capex mix from Epoch AI’s one-gigawatt data centre cost model (May 2026).
Cloud margins today
Operating margins at the big cloud businesses on their latest reported figures.
Sources: Amazon Q2 2026 results (30 July 2026), 10-Ks for 2024 and 2025 and 10-Q for Q2 2026 (segment depreciation); Alphabet Q2 2026 results (22 July 2026) and 2025 quarterly results; Microsoft FY26 Q4 results (29 July 2026).
Operating margin vs end-user revenue
The hyperscalers’ operating margin on AI compute at each level of end-user spending in 2030.
Sensitivity
End-user spending for each combination, with its share of 2030 US GDP beneath.
How the model works
Formulas, simplifications and sources.
The calculation
- Depreciation is cumulative capex divided by the average useful life. As an annual rate that is 1 ÷ life, so 10 years is 10% a year.
- Required operating profit is the required return times the cumulative capex.
- EBITDA must cover depreciation and that profit. Compute revenue = (depreciation + required profit) ÷ EBITDA margin.
- End-user spending is compute revenue divided by the share of each end-user dollar that reaches the hyperscalers, which is 1 − AI gross margin because all the AI companies’ cost of sales is compute.
- Share of GDP is end-user spending, times the US share, divided by 2030 US GDP (today’s GDP compounded for four years).
- The operating margin is the required profit divided by compute revenue. It is a result, shown against the cloud businesses.
Simplifications
- 2030 is a steady-state test year with all the capex in use, as CGAM assumes.
- One useful life covers servers and buildings together; the blend calculator shows how the two combine.
- Land is not depreciated, so a blend that ignores it slightly overstates the rate.
- The return is on gross capital invested, which is stricter than a return on depreciated book value.
- All hyperscaler AI compute revenue, training and inference alike, is treated as AI companies’ cost of sales, as CGAM’s bridge implies.
- As a simplification, all calculations are before tax.
- Cloud benchmarks cover whole segments, not just AI, and reflect asset bases built before the AI boom.
Sources
The AI Capex Boom (Chris Clothier, September 2026). Cost structures, cloud margins, benchmarks and useful lives are from company filings of Alphabet, Amazon, Meta, Microsoft, Oracle, CoreWeave and Nebius, to 6 October 2026.