2024 United States Data Center Energy Usage Report

The results presented here indicate that the electricity consumption of U.S. data centers is currently growing at an accelerating rate. Figure ES-1 shows a compound annual growth rate of approximately 7% from 2014 to 2018, increasing to 18% between 2018 and 2023, and then ranging from 13% to 27% between 2023 and 2028. This surge in data center electricity demand, however, should be understood in the context of the much larger electricity demand that is expected to occur over the next few decades from a combination of electric vehicle adoption, onshoring of manufacturing, hydrogen utilization, and the electrification of industry and buildings.

This report utilizes a “bottom-up” model, with data and assumptions starting at the equipment level being scaled up and aggregated to generate total results. Figure 2.1 shows the overall structure of the model, including input data sources and the major units of analysis.

Figure 2.1. Flow chart for the data center electricity model used in this study.

As shown in Figure 3.5, we assume that conventional servers idle at 51% of their maximum power in 2014, dropping to 36% in 2023 and 27% in 2028.

As noted above, servers in the SPEC database are understood to be more efficient than the general market, so this is in line with those expectations. However, it is notable that the SPEC trend is relatively constant, if not increasing, with idle power fractions at 20-25% in recent years. This could imply a functional lower bound for idle power, relative to maximum power, that will remain constant into the future. Our current model assumption essentially states that the general server market will near this lower bound by 2028. However, this may underestimate idle power draw.

Figure 3.5. Idle power for conventional servers, as a percentage of maximum operating power.

Servers in internal and small data centers average 11% utilization in 2014, rising linearly to 20% in 2027. Colocation data centers average 21% utilization in 2014, rising to 35% in 2027. Hyperscale data centers average 45% in 2014, rising to 50% in 2027. AI accelerated and non-accelerated servers doing training are assumed to have a constant 80% operational time throughout the whole period, while the same servers doing inferencing have a constant 40% operational time (see Figure 3.6).

Figure 3.6. Operational time of servers given data center type.

Figure 3.9. Total server installed base for 2014–2028 with higher bound shipments (left). Adjusted installed base with lower bound GPU shipments (right).

Figure 3.10. Assumed average drive capacity (TB) of new storage devices shipped in each year.

Figure 4.1. Distribution of servers by data center type.

Figure 4.4 reveals a striking variation in PUE values across different types of data center spaces, reflecting the diverse efficiency practices implemented within each category. PUE values vary according to cooling system types, albeit not as significantly as the variance among different types of spaces. Utilization of economizers, adiabatic cooling, or dry coolers can contribute to lower PUE values, highlighting the impact of different cooling methodologies on overall energy efficiency.

Figure 4.4. Simulated PUE and WUE (site) ranges by data center cooling system and space type.

Figure 4.5. Aggregate PUE and WUE across space type categories considering the facility locations and mix of cooling systems present in 2023.

The total annual server energy use from 2014 to 2023 is presented in Figure 5.1, along with a future scenario range of server energy use through 2028. Server energy usage grew from about 30 terawatt-hours (TWh) in 2014 to nearly 100 TWh in 2023, more than tripling during that period. A large portion of this increase came from GPU-accelerated AI servers, which grew in energy usage from less than 2 TWh in 2017 to more than 40 TWh in 2023. Conventional servers, primarily dual processor servers, increased significantly during the same period as well, doubling from about 30 TWh to nearly 60 TWh.

Figure 5.1. Server annual electricity usage by type.

After 2023, server energy use is presented as a range to reflect various scenarios of future equipment shipments and operational practices. Specifically, the count of future GPU shipments varies within the higher and lower bounds previously described in the AI Accelerated Server Shipments section. Second, the average operational power of AI servers varies between 60% and 80% of the rated power as noted in the Server Operational Wattage section. Finally, the average operational time of AI servers varies between 75% to 85% of the year as noted in the Server Operational Time section. Together the variations of these inputs provide a range of operation energy, with the low and high end representing about 240 and 380 TWh in 2028, respectively, as shown in Figure 5.1.

Figure 5.2. Server annual electricity use by space type.

Figure 5.2 presents total annual server energy use allocated by data center space type. In 2014, over 60% of server energy consumption was in internal data centers. By 2023, this fell to nearly 10%, with hyperscale and colocation data centers accounting for almost 80%. After 2023, internal data centers’ share of server energy continued to fall, reaching below 2% by 2028. Hyperscale and colocation data centers continue to grow in proportion, and by 2028 it is expected that hyperscale and colocation will account for over 90% of server energy consumption, primarily driven by AI workloads.

Figure 5.5 presents total annual data center energy use from 2014 to 2023, along with a future scenario range of total data center energy use in 2024 and 2028. Data center energy use remained fairly stable between 2014–2016 at about 60 TWh. Energy use began to increase as the amount of accelerated AI servers in the server stock began to become significant in 2017, and by 2018 data centers consumed about 76 TWh, representing 1.9% of total U.S. electricity consumption. U.S. data center energy use continued to grow at an increasing rate, reaching 176 TWh by 2023, representing 4.4% of total U.S. electricity consumption.

Together the scenario variations provide a range of total data center energy, with the low and high end representing about 325 and 580 TWh in 2028, as shown in Figure 5.5, representing 6.7% to 12.0% of total U.S. electricity consumption.

Figure 5.5. Total data center electricity use from 2014 through 2028.

Figure 5.6. Total data center electricity use from 2014 through 2028 by equipment type.

Figure 5.10 illustrates the spatial variation of the annual average water consumption intensity and GHG emission intensity factors across U.S. counties, derived from the balancing authorities governing those areas. This visualization reveals significant, yet uneven, variability across different regions of the country.

Figure 5.10. Water consumption and GHG emission intensity factors of electricity use by county.

Figure 6.3. Projected U.S. Bitcoin mining energy consumption under a moderate price growth scenario (2024–2028).