PwC's data centre outlook, published last week, puts annual capital expenditure at roughly $800bn this year and has it reaching about $1.8 trillion a year by 2050.
Twenty-four-year projections are not forecasts in any strong sense and should be read as arithmetic about assumptions. But the shape is the useful part, and the shape is not a boom.
A boom has a peak in it somewhere. This is a run rate that roughly doubles over a generation, which is the profile of electricity generation or motorway construction — a category of spending that recurs because the thing being built is consumed rather than completed.
What changes if it is infrastructure
Three things, and the industry is behaving as though none of them are true yet.
Constraints stop being bottlenecks. A bottleneck is a temporary obstruction you clear and move past. Every physical limit this buildout has hit has been described that way: the power connection, the transformer, the enclosure, the chip. At a $800bn annual cadence they are not obstructions on the way to a finished state. They are the permanent conditions of operating, and they want permanent supply chains rather than emergency procurement.
This paper has covered them one at a time — more generation waiting in interconnection queues than is currently operating on the grid, and $242m going into a plant that makes the electrical enclosures the power arrives in. Each was written as the next constraint. Read against a run rate, they are the same constraint recurring, which is what infrastructure feels like from the inside.
The financing changes. Boom capital is equity and it accepts a peak. Infrastructure capital is debt with a long tenor priced against a predictable revenue stream, and it requires exactly the thing this industry does not yet have: contracted, long-dated demand. A hyperscaler's balance sheet has been substituting for that contract.
And the regulation changes. Nobody writes a utility framework for a boom. They write them for things that will still be consuming land, water and power in thirty years.
The physical design is already conceding it
The hardware roadmap says infrastructure even where the language says boom.
Higher power density per rack, liquid cooling instead of air, optical networking between racks rather than copper — these are not incremental refresh decisions. They are the choices you make when you expect the facility to be power-constrained for its entire life and you are optimising for watts rather than for floor space.
Air cooling was adequate for a general-purpose data centre and is not adequate here. The switch is expensive, it is difficult to retrofit, and nobody undertakes it for a demand curve they expect to subside.
Where the number should be treated carefully
Cumulative figures to 2050 are close to meaningless and this desk has said so about the same $31.6 trillion projection: an industry whose accountants cannot agree on how long a GPU lasts is not in a position to compound anything over twenty-four years.
The near-term annual figure is different. Roughly $800bn in 2026 is a measurement of committed and in-flight projects rather than an extrapolation, and it is the number that has consequences now — for grid planning, for water allocation, and for the local tax bases being asked to host the buildings.
What to watch
Not the 2050 total.
Watch whether any large operator signs a multi-decade power purchase agreement at a fixed real price. That is the specific instrument that converts a spending forecast into an infrastructure commitment, because it is a promise to keep consuming after the current demand story has been tested. Until somebody signs one, the industry is still telling utilities it is a boom, and utilities are planning accordingly.
The projection of annual data centre capital expenditure at roughly $800bn in 2026 rising to about $1.8 trillion per year by 2050, and cumulative investment of $31.6 trillion through 2050, are from PwC's Global Data Centre Outlook published on 2 September 2026. The characterisation of enterprise data centre hardware being reshaped around higher power density, liquid cooling and optical networking is as described in Deloitte's technology outlook for 2026. Interconnection queue volumes and the Eaton enclosure investment are as previously reported by this publication. Long-range projections of this kind are estimates and are treated here as such. The distinction between boom and run-rate spending is our own.
Topics technologydata centresinfrastructurecapital spendingenergy
