Data Centre Spending to 2035
Microsoft, Amazon, Alphabet and Meta spent $409 billion on capital projects in their last reported year — 72% more than the year before, and almost a third of everything the four of them have spent in the past decade combined. What the next ten years costs depends on three things: how fast that keeps growing, how much of each dollar buys chips rather than concrete, and how long a chip lasts before it is written off. The four companies do not agree on that last one. Set them yourself.
The model
Base case
Where the money goes
Sources and method
Where the numbers come from. The base is capital expenditure as reported in each company's own cash flow statement for its most recent completed fiscal year: Amazon $131.8bn and Alphabet $91.4bn (calendar 2025), Microsoft $115.9bn (fiscal year to June 2026), Meta $69.7bn (calendar 2025) — $408.9bn together, against $237.3bn the year before. The $1.40tn comparison is the same four companies' reported capex summed across their ten most recent fiscal years. The replacement cycles are the useful lives each company states for servers and network equipment: Amazon shortened its estimate from six years to five, citing the pace of AI hardware; Meta extended to five and a half; Microsoft and Alphabet both moved from four years to six. The chip-versus-buildings split starts at Alphabet's reported 2025 mix of roughly 60% servers.
What we have assumed. Everything past the base year. Growth, the drift in chip share, and the replacement cycle are yours to set — there is no forecast here, only arithmetic on inputs you choose. Two caveats worth stating plainly: reported capex is total capital expenditure, which for Amazon in particular includes warehouses and delivery infrastructure as well as data centres, so the base overstates pure data-centre spending by an amount none of them disclose; and servers are treated as retiring on the cycle you set while buildings and power plant are assumed to outlast the ten-year window. The Interstate comparison uses ~$500bn, the usual present-day restatement of the system's $114bn 1956–1992 cost. The widely repeated "45×" version of that comparison does not check out against any of these totals, which is why we compute it live rather than quoting it.