U.S. data center energy demand moved from a specialist infrastructure issue to a material business planning variable between 2014 and 2023. The DOE and Lawrence Berkeley National Laboratory estimates cited in recent federal reporting show a sector that more than doubled electricity use over that period, with faster growth after 2018 and a wide range of plausible outcomes by 2028 and 2030.
In 2023, U.S. data centers consumed about 176 terawatt-hours of electricity, equal to roughly 4.4% of total U.S. electricity use. That was up from about 70 terawatt-hours in 2014, according to the DOE report. The same federal analysis projected 2028 consumption at 325 to 580 terawatt-hours per year, or 6.7% to 12.0% of total U.S. electricity use.
Those figures are not a single forecast of the future. They are scenario-based estimates tied to assumptions about server deployment, AI accelerator use, idle power, storage, networking, and the mix of hyperscale, colocation, and internal facilities. For business readers, the central point is that load growth has become large enough to affect site selection, energy procurement, facility design, and utility planning.
Data Center Energy Projections Through 2030
Recent Growth Rates Changed The Baseline
The growth rate is as significant as the absolute volume. DOE/LBNL research cited in the notes shows data center electricity use growing at about a 7% compound annual rate from 2014 through 2018. From 2018 through 2023, that rate accelerated to about 18% annually. From 2023 through 2028, the projection range widened to 13% to 27% annual growth.
The 2025 Berkeley Lab update put a further marker on 2030. In its reference case, total U.S. data center electricity use reached about 649 terawatt-hours in 2030. The estimate range was about 578 to 782 terawatt-hours, depending on assumptions about installed equipment, AI chip utilization, and server idle power, according to the 2025 Berkeley Lab update.
Why The Range Matters
A range from 578 to 782 terawatt-hours is not statistical noise. It reflects uncertainty in how quickly AI-oriented computing equipment is installed, how intensively accelerators are used, and how much power servers draw when not running at full utilization. These variables are operational, not abstract. A facility with high utilization and dense accelerator clusters can create a different load profile than a conventional enterprise facility with lower average server activity.
For corporate buyers and utilities, the range also limits overconfident planning. If demand lands near the lower bound, infrastructure stress and procurement needs are still material. If it moves toward the higher bound, utilities, grid planners, and large technology buyers may face more pressure to coordinate interconnection queues, long-term power contracts, and backup power decisions. A related analysis of DOE findings discusses those forecast limits in more detail.
Technical Sources Of Load Growth
Data Center Energy Growth Drivers
Data center energy growth is not only a story about more buildings. The equipment mix inside those buildings is changing. The DOE/LBNL research notes that internal, on-premises data centers accounted for more than 60% of server energy use in 2014. By 2023, that share had fallen to about 10%. By 2028, internal facilities were expected to fall below 2%, while hyperscale and colocation facilities were expected to account for more than 90% of server energy usage.
This shift matters because hyperscale and colocation sites concentrate load in fewer, larger facilities. That can improve management of cooling, operations, and procurement in some cases, but it also concentrates interconnection requests and local power demand. The federal estimates do not imply that every region faces the same exposure. They do show that business risk is shifting from distributed server rooms toward high-capacity campuses and third-party facilities.
Storage And Networking Are No Longer Secondary
Servers remain the central load source, but storage and networking are gaining importance. Storage equipment electricity use rose from about 7 terawatt-hours in 2014 to about 16 terawatt-hours in 2023. By 2028, the projection reached about 22 terawatt-hours, with flash storage accounting for roughly half of that use.
Networking equipment is also expected to grow. The research notes point to about 23 terawatt-hours per year of anticipated network equipment growth by 2028. InfiniBand switches were projected to represent about 45% of network equipment electricity use by then, up from under 10% in 2023. Port speeds below 50 Gbps were projected to fall below 2% of usage by 2028.
Those details are relevant for AI clusters because accelerator-heavy systems depend on high-throughput interconnects, not just processors. Faster networking can raise the power budget for the facility even when the headline procurement decision centers on servers or AI chips. Storage growth also suggests that data retention, training data movement, checkpointing, and inference support systems can add measurable load outside the accelerator rack.
Business Exposure Across Facility Types
Hyperscale And Colocation Operators Carry More Load Risk
The movement from internal data centers toward hyperscale and colocation facilities changes which companies carry operational exposure. Enterprise technology departments once managed much of the energy footprint inside corporate facilities. By 2023, according to the research notes, hyperscale and colocation sites accounted for nearly 80% of server energy use. By 2028, they were expected to exceed 90%.
That change gives cloud providers, colocation landlords, and large AI infrastructure operators a larger role in power procurement and capacity planning. It also gives their customers indirect exposure. A software firm buying cloud compute may not manage the electrical system, but its workload design, model selection, data movement, and utilization patterns still influence demand. Inefficient software can become a cost issue even when the physical facility is outsourced.
Procurement Decisions Depend On Utilization
The 2030 scenario spread shows why utilization deserves more attention. AI chips that sit idle still draw some power, and the LBNL range explicitly depends in part on assumptions about AI chip utilization and server idle power. That means procurement strategy cannot be separated from workload scheduling, model efficiency, and capacity reservation practices.
Operators that buy capacity far ahead of actual workload demand may face different economics from firms that align equipment deployment with verified utilization. At the same time, waiting too long can create service constraints if interconnection, power delivery, or facility capacity is delayed. The available research supports caution rather than a simple rule. Power planning depends on configuration, facility type, and actual workload behavior.
- Facilities teams need to track server, storage, networking, cooling, and support loads separately.
- Cloud and colocation buyers should ask how reserved capacity, idle systems, and burst demand are billed or allocated.
- AI teams should treat model design, training schedules, and inference utilization as energy variables, not only performance choices.
Public communication will also matter as these figures move into policy and education settings. For individuals curious about how to discuss infrastructure topics in educational settings, the online resource stampsinclass.com can be consulted for assistance, although the technical claims here rely on federal energy research rather than classroom material.
What Data Center Energy Means For Operators
Planning Should Treat Forecasts As Bounds
Data center energy planning should treat the DOE and LBNL numbers as boundary conditions, not precise certainties. The direction of travel is clear in the cited estimates: electricity use rose from about 70 terawatt-hours in 2014 to about 176 terawatt-hours in 2023, and the projected 2028 and 2030 ranges are substantially higher. The exact outcome depends on equipment deployment and use.
For operators, the practical response is measurement first. Power usage should be tied to workload categories, not only facility totals. Storage and networking should be included in the energy model. AI accelerator utilization should be tracked alongside procurement plans. Colocation contracts should clarify how energy costs, reserved capacity, and expansion rights are handled.
The DOE data does not prove that every data center project will face the same constraints. It does indicate that U.S. demand growth is large enough to shape business negotiations between cloud providers, colocation firms, utilities, equipment vendors, and enterprise buyers. The firms most exposed are those that treat power as a background utility rather than a design parameter in computing architecture.
