AI’s buildout comes with $4.2 trillion in homework

Viddy the homework, droogies. The AI buildout has reached the part of the exercise where somebody must explain where the money comes back.

Reuters’ October 3 analysis examines whether revenue and productivity gains can arrive fast enough to support the infrastructure spending. Its reporting draws on several estimates with different horizons. Keep the columns separate.^[1]

The revenue column

Bain’s September 29 report says funding projected compute demand would require $6 trillion in annual AI revenue by 2031. Existing consumer and enterprise applications could provide $1.2 trillion to $1.8 trillion. That leaves at least $4.2 trillion a year to find elsewhere.^[2]

Bain points to search and advertising, autonomous machines, physical AI and products that do not yet exist. A substantial amount of the spreadsheet is waiting for someone to invent its customers.^[2]

The spending column

PwC’s September 2 central scenario projects $31.6 trillion in cumulative global data-center capital expenditure through 2050. Equipment refreshes every four to six years help keep the spending cycle going. This is a multi-decade projection, so putting it beside one year’s revenue does not produce a useful profit calculation.^[3]

The small print

These estimates describe what would have to happen under particular assumptions. They do not establish today’s losses or guarantee tomorrow’s sales. Reuters also notes that useful infrastructure can outlast an investment bust.^[1][2][3]

The next lovely slide deck can start with paying customers, renewal rates and the cost of replacing the machinery. Humanity’s glorious future should manage a receivables column.