Stay informed with free updates
Simply sign up to the Artificial intelligence myFT Digest — delivered directly to your inbox.
Generative AI uses a lot of energy, though figuring out how much is hard. There’s no point in asking ChatGPT. It won’t say.
Much of the academic research around the theme has involved counting the number of AI servers shipped, but this misses all sorts of stuff. It’s impossible to know, for example, whether new hardware is replacing less efficient rigs or is additional capacity. Cooling, a big variable, could be adding 50 per cent or more to the energy requirement. Latency might also matter, since we don’t know whether servers are being used for model building, request handling, or dork currency.
Goldman Sachs takes a different approach. The bank’s economics team, led by Hongcen Wei, has looked at commercial power consumption in Virginia, the largest data centre market in the world. More than 35 per cent of all known hyperscale data centres worldwide are in Virginia, according to the Virginia Economic Development Partnership.
Here’s the result:
Commercial power consumption in Virginia rose 37 per cent from 2016 to 2023 while remaining fairly flat in most other states, says Goldman.
The first ramp is probably for cryptocurrency mining, interrupted briefly by the pandemic, but while commercial power consumption fell across the US through 2021, Virginia’s energy draw rebounded almost immediately.
The peak of the red line comes shortly after ChatGPT’s launch in November 2022. The flatlining since can be explained by power transmission bottlenecks, which slowed build rates and moved some capacity out of state.
Goldman estimates that data centres boosted Virginia power consumption by 2.2 gigawatts in 2023. That’s enough to power 1.5mn homes, approximately, and is equivalent to the output of two nuclear reactors.
Virginia data farms only account for 0.5 per cent of current total US power demand so, even after adding installations in other states, the AI effect nationwide is less significant for future demand than electric vehicles, industrial electrification and air conditioning.
Goldman expects US total power needs to grow from 470GW in 2023 to 567GW in 2030, which includes a tripling of data centre power demand from 15GW to 45GW. The 15GW figure is about equal to the capacity of all US utility-scale solar farms deployed last year.
Notable, however, is that US power demand growth is expected to outpace real GDP growth. That last happened in the 1980s, four business cycles ago:
Only about 20 per cent of current US electricity generation is from renewables, with 60 per cent from fossil fuels. Here’s hoping Sam Altman’s right about nuclear fusion, otherwise the future may be smoggy.
