AI Data Center Power Consumption: What GPU Density Does to the Grid

Planning a GPU deployment requires understanding how much power the racks consume and whether the local grid can support them. In 2024, data centers worldwide consumed approximately 415 TWh, or 1.5% of global electricity. However, this global figure does not show how much power an individual AI data center requires.

For AI data center planning, regional and rack-level figures are more useful. Data centers consumed 26% of Virginia's electricity in 2023, AI racks can require 50 to 150 kW each, and one measured GPU training system averaged 93% utilization over eight hours. These figures help determine rack capacity, site selection and power redundancy requirements.

AI Data Center Power Consumption: Global and US Figures

Power consumption figures vary depending on the year, geographic region and type of data center being measured. The following figures include facility-level data center power usage, accounting for power usage effectiveness (PUE).

These figures provide a baseline for understanding data center electricity demand. However, planning an AI data center also requires understanding the power consumption of individual GPUs and racks.

How GPU Density Increases AI Data Center Power Requirements

AI data centers require more power per rack because high-performance GPUs consume significantly more electricity than traditional CPUs.

According to the Congressional Research Service (CRS), advanced data center GPUs available in early 2025 had a maximum thermal design power (TDP) of 350 to 700W, compared with an average of 150 to 350W for contemporary CPU series. Hanwha Data Centers also cites Deloitte research estimating that next-generation GPUs from 2024 could consume up to 1,200W each.

Higher GPU power consumption increases the amount of electricity required per rack. The same Deloitte research cited by Hanwha provides the following AI rack density figures:

These figures show why facility-wide average rack power consumption can be misleading when planning an AI data center. GPU racks may require substantially more power than general-purpose server racks.

CRS also defines a hyperscale data center using three criteria:

1. Power capacity: More than 100 megawatts (MW) of electricity
2. Server count: At least 5,000 servers
3. Facility size: At least 10,000 square feet of floor space

For comparison, CRS states that 100 MW is approximately enough power to support 80,000 US households.

AI Training vs. Inference: How Power Consumption Differs

AI training and inference have different power consumption patterns. Inference power consumption varies by the type of request, while training can sustain high GPU utilization over extended periods.

For inference, published estimates compiled by Our World in Data show the following electricity consumption per request:

• Google Gemini: Approximately 0.24 Wh per median query
• OpenAI ChatGPT: Approximately 0.34 Wh per query, according to Sam Altman
• Epoch AI: Approximately 0.3 Wh per query
• Standard AI agent request: Approximately 1.1 Wh, based on IEA 2026 modeling
• Reasoning-agent request: Approximately 50 Wh, based on IEA 2026 modeling

These figures show how AI energy consumption varies by workload, particularly between standard queries and more complex agent requests.

Training workloads have a different power profile. CRS reports a December 2024 study of an eight-GPU training system that operated at 93% average utilization over an eight-hour period.

During this test:

• Median chip power draw was 7.92 kW
• Total electricity consumption was 62 kWh
• Average GPU utilization was 93% over eight hours

At a larger scale, CRS cites an April 2025 report estimating that training a single large AI model required 25.3 MW of power.

These examples illustrate why AI data center power planning needs to account for both sustained training loads and variable inference demand.

Projected AI Data Center Electricity Demand Through 2035

The International Energy Agency (IEA) projects that global data center electricity consumption will reach approximately 945 TWh by 2030, representing just under 3% of global electricity demand.

Most of this growth is expected to occur in the United States and China, which together account for approximately 80% of the projected global increase through 2030.

• United States: An additional 240 TWh, representing 130% growth
• China: An additional 175 TWh, representing 170% growth

The IEA also models three scenarios for global data center electricity consumption by 2035:

1. Lift-Off: More than 1,700 TWh (~4.4% of global electricity), assuming faster AI adoption and a more resilient supply chain than the Base Case
2. High Efficiency: Approximately 970 TWh (2.6%), assuming greater efficiency improvements in software, hardware and infrastructure than the Base Case
3. Headwinds: Approximately 700 TWh (under 2%), assuming slower AI adoption and supply-chain constraints

CRS also cites an April 2025 projection that the power required to train frontier AI models could double annually.

These projections indicate that AI data center electricity demand will continue to grow, although the rate depends on AI adoption, infrastructure availability and efficiency improvements.

Regional Power Demand and Electricity Costs

AI data center power consumption is concentrated in specific regions, where large deployments can account for a significant share of local electricity demand.

According to an Electric Power Research Institute (EPRI) estimate reported by Pew, data centers consumed approximately 26% of Virginia's electricity supply in 2023.

Other US states with high data center electricity consumption in 2023 included:

• North Dakota: 15%
• Nebraska: 12%
• Iowa: 11%
• Oregon: 11%

Outside the United States, Our World in Data reports that data centers account for more than 20% of Ireland's electricity consumption.

This concentration can also affect electricity costs.

In the PJM electricity market, data centers accounted for approximately $9.3 billion of the 2025–2026 capacity price increase, according to Pew. Estimated residential electricity bill increases included:

• Western Maryland: $18 per month
• Ohio: $16 per month

However, data center growth does not necessarily result in higher electricity prices in every region.

CRS reports an LBNL-led analysis covering 2019 to 2025 that identified grid infrastructure investment as the main driver of US electricity prices. States experiencing high data center growth generally saw electricity price decreases, although some experienced increases or little change.

The PJM figures apply specifically to one regional capacity market and one auction year.

What AI Data Center Power Consumption Means for GPU Deployment Planning

The power consumption figures above highlight three factors to consider when planning an AI data center.

1. Rack power capacity: AI racks currently require approximately 50 to 150 kW. Power distribution systems must be sized for the actual GPU rack requirements rather than the lower average consumption of traditional server racks.

1. Grid location: Regional electricity demand affects site selection. Virginia's 26% data center electricity share and PJM's capacity market prices illustrate why local grid conditions need to be evaluated before selecting a facility.

1. Power redundancy: GPU training workloads can sustain high utilization over extended periods, while inference demand can fluctuate. Each workload requires an appropriate power redundancy design.

Customers using zWARE typically achieve 30 to 40% higher effective GPU utilization.