AI capex, measured against every boom before it

Capital spending on AI has become one of the largest investment cycles in economic history, and the cleanest way to see it is against the booms it is now passing:

Total US AI spending has scaled from $33bn in 2022 to an estimated $862bn in 2026. Six technology leaders have spent about $1.29tn of capex since 2020, versus $0.9tn by six global oil majors, and Microsoft, Meta, Alphabet and Amazon alone account for roughly 94% of it. One chipmaker’s revenue now equals 8% of all US information-processing equipment investment.

Two stress points are visible in the data. First, demand pull-forward: US business purchases of computers and peripherals totalled $585bn over 2.5 years: five to six years of pre-COVID-normal demand compressed into an already-saturated market. Buying pulled forward at that scale historically precedes mean reversion, because more demand now mechanically means less later. Second, input-cost inflation: GPU rental prices and semiconductor producer prices both turned sharply upward in 2026, meaning part of the rising capex figures now reflects paying more per unit of compute, not adding more compute.

The arithmetic that decides how this ends: every $1 of AI infrastructure needs the compute layer to charge ~$1.50, the model layer ~$2.70, and end users ~$4 for returns to clear across the stack. With global data-centre capex heading toward $1tn a year, that implies $2.5–4tn of annual AI revenue, against total global IT spending of roughly $6tn. The maths only closes if AI captures budgets beyond software, from the ~$45tn global wage bill.

What we’ll watch: hyperscaler capex guidance against their free cash flow; GPU rental prices as the truest demand signal; and world computer hardware exports as a share of GDP versus their 2000 peak (0.56%).