BIS paper warns AI data centres' power demand could blur central banks' inflation readings

Gambacorta and Polizzi place the noise in two readings at once, arguing that AI-driven productivity changes and energy constraints could muddy both capacity utilisation and inflationary pressure, according to Crypto Briefing's account of the paper ^[3]. Both errors lean the same way, toward reading an economy as hotter than it is. Output that rises because workers became more productive can look like strain on a utilisation gauge, and a power price pushed up by data centres competing for constrained grids ^[6] can look like demand running ahead of supply in the inflation figures.

The paper says AI systems, data centres above all, add significantly to electricity consumption and the emissions that come with it ^[4]. The International Energy Agency's figure for 2030 is more than 945 terawatt hours, more than double current use ^[5]. That puts today's global data-centre load below 472.5 terawatt hours ^[14]. Taking this year as the base, a doubling by 2030 needs at least 19% compound growth a year for four straight years ^[15]. The account of the paper does not put a figure on how much of any country's measured inflation comes from the electricity bill.

Whichever of the paper's two scenarios arrives ^[9], the money passes through the same channel twice. Spending on data centres and new generation shows up first as grid demand and a higher electricity price that a central bank has to interpret ^[6], and later as a return or a loss for whoever financed it. The authors warn that disappointing returns on that build-out could pose risks to financial stability ^[8].

I think the stability warning matters more to holders of capital than the signal noise does. A misread inflation print can be corrected at a later rate decision, while a loss on a data centre or a power plant stays on someone's balance sheet. The counter-case comes from the paper itself, which credits AI with gains in energy efficiency, forecasting and low-carbon technology ^[11]. If those gains arrive on the incremental path, they offset part of AI's own power demand and ease both problems at once. The noise thesis fails on a simpler test, if electricity prices in grid-constrained markets do not outrun other prices as data-centre demand climbs toward the IEA's 945 terawatt hours ^[5].