The gold rush in artificial intelligence is creating a different kind of resource squeeze, electricity.
According to JP Morgan, the world’s data centres are on track to consume nearly a tenth of all new global power capacity by the end of the decade.
The bank has lifted its forecast for total data centre power use by as much as 40% between now and 2028, pointing to bigger chips, heavier workloads and slower efficiency gains than once hoped.
It now expects global data centre capacity to hit 242 gigawatts by 2028, up from 97 GW last year, enough, in round terms, to power more than 200 million homes.
AI servers are the culprit and the opportunity. Power demand from them alone is growing at roughly 60% a year, JP Morgan says, with each new generation of chips drawing ever more current to crunch complex models.
Keeping all that silicon cool is becoming its own industry. The bank reckons the market for liquid-cooling equipment, the pipes, pumps and micro-channel lids that replace traditional fans, will expand by about 50% annually through 2028, reaching some $30 billion.
It’s not just a matter of preventing chips from melting. The energy used to chill these vast server farms has become one of their biggest cost and environmental headaches.
Data centre efficiency, measured by something called “power usage effectiveness” (the ratio of total facility energy to computing energy), has barely improved in recent years, stuck around 1.1 to 1.2.
Liquid-to-liquid cooling, which circulates fluid through chips directly rather than blowing air, could bring that figure down further, though practical hurdles mean it may take a few years to scale.
On the supply side, the story looks like an industrial-scale refit. JP Morgan sees a 64% compound growth rate for power-supply units over the next three years as data centres switch to higher-voltage systems and more integrated designs.
That could lift demand for specialist electrical equipment from Asian suppliers such as Delta, Vertiv, LS Electric and Hyundai Electric.
The question, as ever, is whether grids can keep up. The analysts warn that transmission bottlenecks and the long lead times for new substations could become the main constraint on AI growth.
One answer, they suggest, could be nuclear, which may yet find a surprising ally in the data economy.