The market still underestimates the power-hungry AI. On January 3, Barclays analyst William Thompson and his team released a report analyzing and summarizing the forecasts of about 70 pages of congressionally mandated research reports issued by the U.S. Department of Energy (DOE).
According to this forecast, U.S. data center power demand is expected to grow 2-3 times from 2023 to 2028.
Barclays said that this growth forecast is much higher than the 15% forecast by many institutions, and also higher than the 14-21% expected by Barclays in its previous report PoweringAI:CalibratingUSDataCenterEnergyDemand. However, Barclays pointed out in this report that there is uncertainty in the forecast due to rapid changes in data center deployment, equipment types, operating practices and cooling systems.
In addition, the U.S. Department of Energy also predicts that in the next few years, the growth of water consumption in U.S. data centers will exceed the growth of electricity demand. At present, direct water consumption has soared from 21 billion liters in 2014 to 66 billion liters in 2023, and is expected to increase by about 2-4 times by 2028.
Data center energy demand grows beyond market expectations
A U.S. Department of Energy report shows that U.S. data centers will consume 176TWh of electricity in 2023, accounting for 4.4% of the total U.S. electricity demand. It is expected that electricity demand will increase to 325-580TWh by 2028, accounting for 6.7%-12% of the total electricity demand in the United States.
According to research from the Department of Energy, AI servers are the main driver of electricity demand growth.
According to the report, the power demand for AI servers is expected to increase from approximately 40TWh in 2023 to approximately 165TWh (low estimate) to 325TWh (high estimate) in 2028, an increase of approximately 4-8 times. Research shows that by 2028, AI training consumption will exceed AI inference demand, accounting for 50-53% of the total energy consumption of AI servers, as more high-power GPUs will be used for AI training.
As for the power demand of traditional servers, the U.S. Department of Energy expects little change, from about 60TWh in 2023 to about 60TWh (low estimate) to 70TWh (high estimate) in 2028.
Data center water use growth will outpace power demand growth
The Department of Energy study also said that direct water consumption in U.S. data centers (mainly related to cooling) soared from 21 billion liters in 2014 to 66 billion liters in 2023, and is expected to increase approximately 2-4 times by 2028 to approximately 145-275 billion liters.
Interestingly, the report also proposes that direct water use intensity (water consumption per kilowatt-hour) is expected to increase from approximately 0.38L/kWh in 2023 to approximately 0.45-0.48L/kWh in 2028. Barclays hypothesizes this reflects a change in the proportion of water-cooled chillers to cope with the increasing energy density of AI data centers.
It is worth noting that the report's forecast does not include the nearly 80 billion liters of indirect water use, which is water used to produce electricity, such as steam for nuclear power and coal power plants. If this part of the water consumption is included, it will also increase the water consumption by about 4.5 liters per kilowatt hour.