AI Infrastructure Spending Is Outpacing Its Proven Returns

AI infrastructure investment is enormous by historical standards, but the scale of the buildout does not prove that it will earn an adequate return. The case depends on two things arriving at sufficient scale: revenue from AI services and productivity gains for the businesses using them. So far, the evidence is mixed—and the largest figures are projections, not money already spent.

How large is the AI buildout?

J.P. Morgan Asset Management says major cloud companies spent more than $400 billion on capital investment in 2025 and are on track to approach $800 billion in 2026. That is a rapid jump in annual spending, though it is not directly comparable to a cumulative, multi-decade projection. 9

PwC’s baseline forecast is for $31.6 trillion in global data-centre capital expenditure through 2050. It is a projection of future spending, not a tally of completed investment. Reuters reports that PwC considers the projected buildout larger than spending during the railway and dotcom booms, even after adjusting for inflation. 111

The U.S. investment pipeline is also striking. Economist Stijn Van Nieuwerburgh estimates that AI infrastructure investment could total $10.3 trillion from 2025 to 2032, averaging about 3.6% of U.S. GDP per year over that period. This is an estimate of a pipeline, not a settled spending total.

What revenue would justify the investment?

Bain’s 2025 analysis estimates that meeting anticipated demand for AI computing by 2030 would require about $2 trillion in annual revenue to fund the computing capacity. Even after accounting for AI-related savings, Bain estimates an annual gap of roughly $800 billion. These figures describe a scenario, not a confirmed shortfall that has already occurred.

Bain’s analysis also points to a challenge for the investment case: productivity improvements in existing markets may not be enough on their own. Additional revenue could depend on new uses and markets, including AI-guided robotics and discovering materials for batteries and semiconductors. Those are potential sources of future value, not guaranteed outcomes. 1

It is also important to distinguish revenue earned during the buildout from returns earned by its eventual users. Spending on chips, cloud services and data-centre construction creates revenue for suppliers. But that alone does not establish that companies adopting AI will gain enough to keep paying for the infrastructure at its current pace.

Are productivity gains visible yet?

There are signs of monetization alongside reasons for caution. J.P. Morgan reports fast growth in cloud revenue, but says U.S. labour-productivity growth averaged about 1.3% over the last three quarters—a pace it describes as consistent with the relatively low-productivity backdrop of the 2010s, rather than a clear AI-driven acceleration. 915

A Reuters report on company experience found that many businesses were still struggling to achieve meaningful returns on their AI investments. In surveys cited in that report, 15% of respondents in one executive survey said AI had improved profit margins over the previous year, while 5% in another said they had seen widespread value. These are survey results from specific periods, not a census of all businesses, but they underline the gap between broad expectations and demonstrated returns.

Could the payoff take decades?

A delayed payoff is possible, but it is not a guarantee that today’s investments will eventually prove profitable. Diane Coyle has noted that the productivity effects of past transformative technologies could take roughly 10 to 50 years to emerge. That historical comparison is a reason not to judge every technology only by its earliest returns; it is not a timetable for AI’s impact. 8

The near-term financing question still matters. Van Nieuwerburgh’s estimate describes a substantial investment requirement before any eventual economy-wide benefit is assured. Meanwhile, Bain’s revenue gap is conditional on its assumptions about future compute demand, costs and savings. Both figures are useful gauges of the scale of the bet, not definitive forecasts of success or failure.

The bottom line

AI infrastructure spending is projected at a scale that outstrips earlier railway and dotcom buildouts in PwC’s comparison, while near-term investment by major cloud companies is already climbing quickly. To justify that commitment, AI must generate durable revenue and benefits beyond the companies selling infrastructure—potentially through productivity improvements and new markets.

The evidence does not yet settle whether those returns will be large enough, or how quickly they will arrive. Current cloud revenue growth is encouraging, but recent productivity data and business surveys do not yet demonstrate an economy-wide payoff on the scale implied by the investment projections. 915