Stay informed with free updates

Many AI companies insist that their energy-hungry data centres aren’t leading to higher electricity bills for ordinary people, with Anthropic the latest to promise to pick up the entire tab and shield consumers.

However, over in the real world, the actual data looks like this:

The chart is from Goldman Sachs, which has published a report exploring the macroeconomic spillovers from the soaring AI demand for electricity, which is becoming one of the binding constraints on the “bonkers” data centre building spree.

The headline numbers are pretty stark. Electricity prices went up 6.9 per cent last year, more than twice the 2.9 per cent headline rate of the Federal Reserve’s preferred measure of inflation, the Personal Consumption Expenditures index (PCE, among friends).

Electricity inflation varies a lot from state to state, and it’s hard to pin down exactly how much of it is due to AI, but Goldman’s economists suggest that it’s a meaningful factor — with the share of US electricity gobbled up by data centres roughly doubling just since ChatGPT was rolled out in 2022.

And this is just the start. Goldman’s analysts predict that data centres will account for almost half of all US electricity demand growth over the next four years. They are sceptical that consumers can be shielded from the cost of building all the necessary extra power infrastructure to bring data centres online and keep humming:

While we expect these [regulatory] policies to shift a significant share of future utility capex costs to the data center companies, it will be difficult to fully insulate households and non-AI businesses from price increases because

1) the policies often don’t cover all aspects of data center grid costs;

2) AI companies may choose locations that are more lenient on passing through higher prices to other customers;

3) it is difficult to pin down how much of a given increase in costs can be fully attributed to data centers; and 4) power prices can spike before the policies are implemented.

In our forecasts, we assume that data centers bear about two thirds of the excess capex costs required to service data centers (which we proxy by capex in excess of nominal GDP growth), an while households and businesses finance the other third. Within that third, we assume that consumers bear two thirds of the capex costs and non-AI businesses the rest.

Overall, the investment bank’s economists forecast that the nationwide consumer electricity inflation rate will remain about 6 per cent over the coming two years. If non-AI customers have to bear half the cost of all the necessary capex — as opposed to the third that is Goldman’s primary scenario — it will rise to 8 per cent.

Even the more modest scenario inevitably entails a drag on spending and economic growth.

Higher prices will likely lower disposable income growth, which will exert a drag on consumer spending.

Combining the headline and core inflation impacts from Exhibit 11, we estimate that higher electricity prices will lower consumer spending growth by 0.2pp on average in 2026-2027 . . . We expect lower-income households to see the largest declines in both income and spending because electricity accounts for a greater share of their spending.

. . . Our estimates suggest that higher electricity demand will exert a 0.1pp drag on GDP growth in 2026-2027, as lower consumption is partially offset by higher utility capex.

Of course, these effects aren’t huge, and shouldn’t be seen in isolation. Hyperscalers spending squillions on building a massive series of gargantuan data centres is definitely helping US economic growth. Over time, perhaps the productivity miracle that the likes of Kevin Warsh are using as an argument for rate cuts could prove to be disinflationary. Crazier things have happened.

Goldman’s estimate the drag on growth is net of the investment splurge by utilities, but it stresses that the effect is dwarfed by its optimistic views of the growth that AI will stoke. But those gains probably need to materialise soon, otherwise the backlash from people with bigger electricity bills could get tasty.

Anyway, Goldman Sachs has made the entire report public so you can read it at your own leisure here.

Further reading:
— What’ll happen if we spend $3tn on data centres no one needs? (FTAV)