PwC's 2050 model has data centres spending 93 cents of every capex dollar on equipment

In a fixed asset register there is a column for useful life, and that column is where a $31.6 trillion projection turns into something a controller has to sign. Shells and land get decades. A rack of AI accelerators can be obsolete within a few years ^[6]. PwC handles this by calling data centres hybrid assets ^[8], a phrase that names the accounting problem without settling it.

The arithmetic is more useful in dollars than in percentages. At a 70% equipment share, roughly $560bn of this year's projected capex is kit and about $240bn is everything else ^[1]. At 93%, the 2050 figure splits into about $1.67tn of kit against about $126bn of everything else ^[2]. Annual spending in the model rises 2.25 times, while the non-equipment portion, the part that behaves like a building, falls by nearly half in nominal terms ^[3]. The share of spend that is not equipment drops from 30% to 7%, under a quarter of its current weight ^[4].

Scepticism about the headline total is cheap and largely deserved. TNW notes that a 24-year capex model rests on assumptions about demand, chip prices and the continuation of the current investment cycle that nobody can hold with much confidence ^[19]. McKinsey has separately put data centre investment at nearly $7tn by 2030, which TNW reads as broadly consistent given the different horizons ^[15]. Spread over five years that is about $1.4tn annually ^[5], well above the $800bn PwC assigns to this year, though the source does not say which year McKinsey starts counting from ^[15].

The share number survives on fewer assumptions. It needs only chip generations to keep getting shorter rather than longer, which is what TNW reports operators finding, and it may outlast the $31.6tn total for exactly that reason ^[16]. What it does not do is move the queue. Transformers, grid connections, cooling equipment and planning approvals all move more slowly than money ^[17], and the specifics named are grid connection queues in Texas and Denmark alongside transformer lead times measured in years ^[18]. Europe's EUR30bn gigafactory programme, the continent's largest coordinated response, is already running late ^[12].

Here is what the deck says: infrastructure spending, with the permanence that word carries. Here is what the money buys: hardware on a replacement clock set by vendors, sitting inside a shed. The usable test for anyone signing an infrastructure commitment this quarter is to sort every line by replacement cycle, under five years against over fifteen, then check the tenor of the money paying for it. A colo lease and a GPU order do not land in the same box, and the distance between those boxes is where a financing assumption is doing work the asset cannot support.