Electricity Needed to Support Data Centers

US data center power consumption has seen rapid growth driven to support Artificial Intelligence (“AI”) and cloud computing. Utilities’ decisions to build new electricity supply to meet future data center energy usage takes careful analysis and contracting. Without proper planning, the risk is real that utilities will overbuild based on early forecasts or incorrect assumptions.

In 2024, US data centers are expected to have a consumptive capacity requirements of 21 GW (excluding cryptocurrency needs), roughly 3% of US loads which is up 10.5% from 2023. Earlier this year, the Federal Energy and Regulatory Commission (“FERC”) forecasted data center capacity requirements to increase to 35 GW by 2030, a 166% increase.  The Electric Power Research Institute (“EPRI”) in May 2024 projected that data center consumptive loads could rise to be 4.6% under a low case and 9% under a high case of total US energy consumption. The number of US data centers has doubled from 2021 to 2024 based on EPRI numbers. In a April 2024 generation analysis, Goldman Sachs forecasted US data centers at 8% of US consumptive load with 47 GW of new capacity needed by 2030, a 224% increase. The state of Virginia, in particular, has a major data center load representing a reported 25.6% of 2023 electricity consumed statewide as calculated by EPRI.  EPRI also projected in May 2024 that this load could grow to 29% to 46% of Virginia’s total consumptive loads by 2030.

Fundamentally, the demand for electricity to support data centers has been affected by the type of computing, the quantity and type of equipment at the facility, computing demand, and the efficiency of the computing equipment. Some reports indicate that a typical Google 2020 type of internet search could take 0.3 Wh of computing electricity.  Today, it has been reported by EPRI and others that a typical AI search performed by a ChatGPT engine could require 10 times as much electricity (2.9 Wh). This data center power usage could create a massive need for new electricity supply depending on what actually happens.

Nvidia, a major leader in the AI space, recently published a blog post that indicated that newer AI chips may be 4 times as energy efficient as current AI chips. Intel also indicated that a new AI chip it has released may be two times as energy efficient as its older AI chips. Some academics report that neuromorphic chips that wait for threshold events before passing on AI data decisions can be 1,000 times more energy efficient than current AI Graphics Processing Units (“GPUs”). In July 2024, Nvidia stated on their blog that Nvidia GPUs have reduced their energy consumption by 45,000 times over the last 8 years from 2016 to 2024. While observers would expect that future improvements are likely to be dramatically more modest, it does show how forecasting in a straight line is probably not the best approach.

The growth of AI itself is difficult to forecast. While some put numbers to its growth based on server supply production, changes in the workplace are potentially transformative. As a result, the realized long-term growth will take some time to play out and be less server supply constrained than shorter-term forecasts.

Meanwhile, meeting the electricity supply of new data centers is important to utilities and others that must serve that load. While data centers can take 1-2 years to implement, new generation and transmission can take a lot longer to plan, contract and bring online. For example, new generation could take 4-6 years or more to procure, permit and bring online.  If the utility brings on too much supply, and the user does not contractually commit to the same life of supply or other keep whole provisions, the supply overhang can be detrimental to the utility and those that have to pay for the excess capacity. Moody’s rating service has put utilities on notice to be careful in making the right decisions. Utilities that do not heed Moody’s advice may be downgraded as well as risk non-recovery of rates based on imprudent decisions.

Proper planning, therefore, is key as well as understanding the possible drivers and components that lead to the forecasting conclusion. Staying ahead of the future load requirements and the right investment decisions will be critical to grid stability, the success of utilities and others. Establishing the right type of supply contracts will also be important to all parties. Ultimately, the need for future electricity supplies to support data center power usage provides many opportunities for developers and utilities in advancing new competitive supply projects.