A projection published this week puts global data centre investment at about $31.6 trillion between now and 2050. It was picked up quickly, as very large numbers with distant dates tend to be.
It is worth asking what a reader is supposed to do with it, because the answer turns out to be nothing, and the reason is instructive.
Twenty-four years is beyond the reach of the inputs
A forecast is only as good as the shortest-lived assumption inside it. This one requires a view, held to 2050, on the cost of compute per unit of useful work, the price and availability of electricity, the rate at which models improve, what people will pay for the output, the regulatory position in a dozen jurisdictions, and the physical service life of hardware.
Take the last of those, because it is the one with a live controversy attached. This desk reported on Wednesday that the depreciation bill for the AI build-out arrives on a schedule set by useful-life estimates the companies make themselves — and that at least one of them is extending that estimate from fifteen years to twenty-five, a change that moves billions of reported earnings between periods without a single physical fact changing.
If the industry cannot settle how long a server lasts for the purpose of next year's income statement, a twenty-four-year capital forecast is not resting on a foundation. It is resting on the same estimate, compounded.
What the number is actually doing
Not predicting. Framing.
Set beside a separate figure circulating the same day — roughly $1.4 trillion of AI infrastructure spending by 2030 — the 2050 number is more than twenty times larger over roughly twice the horizon. Both are quoted approvingly and in the same news cycle, by people who do not appear to notice that they imply wildly different growth paths.
That is the function. A very large number attached to a distant date makes the present number look like an early instalment rather than a peak. Three-quarters of a trillion dollars a year is alarming if it is the top. It is prudent if it is the opening move in a thirty-one-trillion-dollar programme.
Nobody has to argue that case explicitly. The forecast does it by existing, and it does it in a period when the people committing the capital are being asked hard questions about returns.
The accountability problem
The defining property of a 2050 forecast is that nobody who acts on it will be assessed against it.
The executives approving capital today will have retired or moved on. The analysts citing it will have changed coverage. The firm publishing it will not be asked in 2050 how it did, because by then the number will have been superseded twenty times by successors that were also never checked.
A forecast that cannot be falsified within the tenure of anyone relying on it is not doing forecasting work. It is doing rhetorical work, and it is doing it for free, because being wrong carries no cost.
Compare the numbers this desk finds worth following: quoted lead times for the switchgear a data centre cannot be energised without, the interval between ordering racks and securing the grid connection, the useful-life assumption in a specific filing. Each of those is checkable within a year, by anybody, against a published source. Each has been wrong at least once in ways that mattered.
What would make a long forecast useful
A stated set of conditionals and a stated way to be wrong. If the projection came with the four or five assumptions that dominate it, the value each takes, and the sensitivity of the total to each — so that a reader could watch those inputs and update — it would be a model rather than a headline.
Most such projections do contain that apparatus somewhere in the underlying document. Almost none of it survives the trip into circulation, and the single number that does survive is the one carrying the least information.
How to read one
Ask what would have to be true. Thirty-one and a half trillion dollars over twenty-four years is about $1.3 trillion a year, every year, sustained through at least two recessions, several hardware generations and whatever replaces the current model architecture.
State it that way and it becomes a claim a reader can weigh, which is presumably why it is rarely stated that way.
The honest version of this forecast is short: capital spending on data centres is very large now, is growing, and is being justified by returns that have not yet arrived. Everything after that sentence is arithmetic performed on a guess.
The projection of approximately $31.6 trillion in data centre investment through 2050 is attributed to PwC and was reported on 2 September 2026 by Quartz and Fox Business. The separately reported figure of $1.4 trillion in AI infrastructure spending by 2030 appeared the same day in syndicated market coverage. The practice of companies setting and extending their own useful-life estimates for data centre equipment, and its effect on reported earnings, is as described in this publication's earlier reporting and in the companies' own disclosures. The analysis is our own.
Topics aidata centrescapital spending
