The defining financial question of the decade so far is whether the enormous capital being committed to artificial intelligence infrastructure will earn a return. The spending is real, verifiable and unusually concentrated. The revenue is growing quickly but from a much smaller base.
Three things are worth separating. The first is chip demand, which is measurable and has been consistently underestimated. The second is the physical constraint, which is increasingly electrical rather than silicon: grid connection queues, turbine lead times and cooling water now set the pace of construction in several markets. The third is enterprise adoption, which is where the evidence is thinnest, because pilot projects convert to durable contracts at rates that vary enormously by sector.
Historical analogies are used freely and rarely carefully. The railway and fibre-optic build-outs both destroyed a great deal of shareholder capital while leaving behind infrastructure that proved genuinely transformative. Whether that pattern repeats depends less on the technology than on how much of the current spending is debt-financed and how quickly depreciation schedules catch up with hardware that is replaced every few years.
For readers outside the sector, the practical consequence is electricity prices and industrial policy. Governments that expected data centres to be a straightforward source of investment are now weighing them against household energy bills.
