Stay up so far with free updates
Just register Artificial intelligence myFT Digest – delivered straight to your inbox.
Generative AI uses a whole lot of energy, but it surely's difficult to determine how much. There's no point in asking ChatGPT. It isn’t said.
Much of the educational research surrounding the subject was involved Counting the variety of AI servers shipped, but all varieties of things are missing here. For example, it’s unimaginable to know whether latest hardware will replace less efficient equipment or represent additional capability. Cooling, a giant variable, may very well be a contributing factor 50 percent or more to energy requirements. Latency may be necessary since we don't know if servers are used for model constructing, request processing, etc Idiot currency.
Goldman Sachs takes a special approach. The bank's economics team, led by Hongcen Wei, examined industrial power consumption in Virginia, the most important data center market on the earth. According to the Virginia Economic Development Partnership, greater than 35 percent of all known hyperscale data centers worldwide are situated in Virginia.
Here is the result:
According to Goldman, industrial electricity consumption in Virginia increased 37 percent from 2016 to 2023, while it remained relatively stable in most other states.
The first surge is probably going dedicated to cryptocurrency mining and was briefly interrupted by the pandemic, but while U.S. industrial electricity consumption fell through 2021, Virginia's energy consumption bounced back almost immediately.
The red line peaked shortly after ChatGPT launched in November 2022. The ongoing decline since then may be explained by Bottlenecks in energy transmissionwhich slowed construction rates and moved some capability out of state.
Goldman estimates that data centers increased Virginia's electricity consumption by 2.2 gigawatts in 2023. That's enough to power around 1.5 million homes and is reminiscent of the output of two nuclear reactors.
Data farms in Virginia account for just 0.5 percent of the present total electricity demand within the United States. Therefore, even after adding installations in other states, the statewide AI effect is less important to future demand than electric vehicles, industrial electrification, and air con.
Goldman expects total U.S. power demand to rise from 470 GW in 2023 to 567 GW in 2030, which incorporates a tripling of knowledge center power demand from 15 GW to 45 GW. The 15 GW figure is roughly the capability of all of them U.S. utility-scale solar farms used last 12 months.
What's notable, nevertheless, is that U.S. electricity demand growth is predicted to outpace real GDP growth. This last happened within the Eighties, 4 economic cycles ago:
Only about 20 percent of current U.S. electricity production comes from renewable energies, 60 percent of which come from fossil fuels. We hope Sam Altman is correct Nuclear fusionotherwise the longer term may very well be smoggy.
