Datacenter Power Consumption

Abstract

The proliferation of digital technologies has driven rapid growth of data centers across the United States. These facilities, housing the servers that power the internet, cloud computing, and big data applications, have become critical infrastructure. However, their energy consumption is also expanding rapidly. This white paper explores the rapid expansion of data centers in the US, the associated surge in electricity demand, and the implications for the power industry.

Growth of Datacenters

Datacenters have experienced exponential growth in recent years, fueled by factors such as the increasing reliance on cloud computing, the rise of streaming services, and the proliferation of internet-connected devices. The demand for data storage, processing, and distribution has necessitated the construction of larger and more numerous data center facilities.

The United States, home to many of the world’s largest tech companies, holds nearly half (5,381) of the world’s datacenters¹. These datacenters are concentrated in a few states, largely driven by access to network infrastructure. Far from catching up to demand, the amount of data moving across America’s networks and churning in datacenters is only increasing. Additionally, the rise of computationally intensive technologies like Large Language Models (LLMs) increase the power demand of existing and future datacenters. The International Energy Agency (IEA) predicts data center power consumption will double by 2026 and continue to grow chasing consumer and industry demand².

How Much Power Do Datacenters Use?

In 2023, datacenters in the United States consumed 150 TWh of electricity, a figure equal to the entire nation of Poland¹. Only 5 states consume more electrical power than the combined datacenters of the United States³. For a human-scale perspective, this is enough power to drive an electric car from Earth to the Sun nearly 6 thousand times, or to power the average American home for over 14 million years.

This power consumption is split into three major categories: computing equipment, environmental equipment, and other facilities equipment. Computing equipment includes the servers, storage, and networking equipment for which the datacenter is built. Environmental equipment for a datacenter keeps the computing equipment operating in the correct temperature range, as overheating is a major concern. Other facilities equipment is everything else required to run the building, including security, fire suppression, and lighting. The industry metric Power Usage Efficiency (PUE) is used to measure how much of the consumed power is used by the computing equipment.

Predicted power consumption by datacenters. Image from Nature Magazine4. — Figure 1. Predicted power consumption by datacenters. Image from Nature Magazine⁴.

PUE = $\frac{Total Facility Energy} {IT Equipment Energy}$ 5

A PUE of 1 indicates all power is used for computing with no overhead. The National Renewable Energy Laboratory reports that the industry average PUE is 1.8⁶. This means over 83 TWh of datacenter power consumption in the United States was used by environmental and facilities equipment. PUE across the industry has improved greatly from the early 2000s, but efficiency gains do not offset the magnitude of increased demand¹.

An example of PUE showing how much power is consumed by various equipment. Image from Energy Star7. — Figure 2. An example of PUE showing how much power is consumed by various equipment. Image from Energy Star⁷.

The recent rise in LLM-based technology contributes a meaningful percentage of this consumption, despite being a newer technology. Training the GPT-3 model alone consumed an estimated 1.3 GWh of power, and the arms race to develop larger and more capable models will continue to grow this power consumption⁸.

Challenges

Unlike many traditional industries, where loads are cyclical throughout the workday, datacenters run at high consumption around the clock. A datacenter uses the same power as hundreds of thousands of homes, but can be built in a fraction of the time, placing strain on the grid operator to build infrastructure to cope with the increased demand. There is also the issue of impact; while a single factory experiencing a power interruption event could disrupt that one business, in today’s connected world, an outage at a single datacenter could disrupt thousands of businesses across the world and millions of people.

The demand incurred by datacenters is not well distributed by geography or by proximity to power generation. Datacenters are clustered in regions with high quality network access and industry to support them. While datacenters consumed almost 4% of all electrical power in the United States, datacenters accounted for over 25% of Virginia’s power consumption, largely residing in the state’s “Data Center Alley” in the suburban counties outside the nation’s capital^1,9.

Dramatic expansions are planned for datacenter operations and subsequent power demand. Image from S&P Global 10. — Figure 3. Dramatic expansions are planned for datacenter operations and subsequent power demand. Image from S&P Global ¹⁰.

Conclusion

The datacenter industry consumes a tremendous amount of power, and the rate of growth is accelerating, driven by ever-rising computing needs of consumers, industry, and government. Electric utilities responsible for supporting datacenters should prepare for dramatic increases in demand.

Endnotes

Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption. EPRI, 28 May 2024, www.epri.com/research/products/000000003002028905
IEA, “Electricity 2024”, IEA, Paris, 2024. Available: https://www.iea.org/reports/electricity-2024, Licence: CC BY 4.0
“US Electricity Profile,” U.S. Energy Information Administration, Nov. 02, 2023. eia.gov/electricity/state
N. Jones, “The Information Factories,” Nature, vol. 561, pp. 561, 163–166, Sep. 13, 2018. Available: https://datacenters.lbl.gov/sites/default/files/nature.pdf
V. Avelar, D. Azevedo, and A. French, “PUE: A Comprehensive Examination of the Metric,” The Green Grid, 2012. Available: https://datacenters.lbl.gov/sites/default/files/WP49-PUE%20A%20Comprehensive%20Examination%20of%20the%20Metric_v6.pdf
“High-Performance Computing Data Center Power Usage Effectiveness,” NREL. https://www.nrel.gov/computational-science/measuring-efficiency-pue.html
“Reduce Energy Losses from Power Distribution Units (PDUs),” ENERGY STAR, 2023. https://www.energystar.gov/products/data_center_equipment/16-more-ways-cut-energy-waste-data-center/reduce-energy-losses-power (accessed Aug. 20, 2024).
S. Mehta, “How Much Energy Do LLMs Consume? Unveiling the Power Behind AI,” Association of Data Scientists, Jul. 03, 2024. https://adasci.org/how-much-energy-do-llms-consume-unveiling-the-power-behind-ai/
“Data Centers | Virginia Economic Development Partnership,” Virginia Economic Development Partnership. https://www.vedp.org/industry/data-centers
S&P Global, Estimated Operating and Planned Data Center Space and Power Consumption. [Online Image]. Available: https://www.spglobal.com/marketintelligence/en/news-insights/research/datacenter-companies-continue-renewable-buying-spree-surpassing-40-gw-in-us