You spend thousands on GPU instances each month, but the electricity powering those chips may cost more than the hardware itself. AI data center power consumption has surged 300% since 2022, with a single training cluster now drawing 150 to 250 MW of continuous power. Here is what the real numbers look like and where that energy goes.
What AI Data Center Power Consumption Means in Practice
AI data center power consumption is the total electrical load for GPU compute, networking, cooling, and facility overhead serving AI workloads. A single NVIDIA H100 draws 700W at peak. The B200 pulls 1,000W. Scale to 100,000 GPUs and compute alone demands 87.5 MW before cooling adds 30 to 50% on top.
The IEA reported global data centers consumed approximately 460 TWh in 2024. AI accounted for roughly 15% by early 2026, up from 2 to 3% in 2023. Our breakdown of how much energy AI uses covers the full scope across providers and models.
Where the Power Goes Inside an AI Facility
GPU Compute: 60 to 70% of Total Power
A DGX H100 system with eight GPUs draws 10.2 kW at full load. A rack of four DGX systems pulls 40 to 42 kW. At facility level, a 50,000-GPU cluster consumes 43 to 50 MW for compute alone.
Cooling: 15 to 25% of Total Power
Air-cooled facilities running dense GPU racks operate with data center power usage effectiveness (PUE) ratios of 1.4 to 1.6. Direct liquid cooling pushes PUE to 1.10 to 1.15. Google reports fleet-wide PUE of 1.10. Microsoft reports 1.12. Every 0.1 PUE reduction at a 200 MW campus saves 20 MW of continuous power, roughly $8.8 million annually.
Networking and Auxiliary: 10 to 15%
InfiniBand switches, storage arrays, and power distribution account for the rest. A large training cluster requires thousands of NDR switches drawing 1,200 to 1,500W each, collectively adding 1.5 to 2.5 MW.
Real Power Numbers by Provider and Model
Training GPT-4 consumed approximately 50 GWh on 25,000 A100 GPUs. Meta’s Llama 3 405B used 30 GWh over 54 days on 16,384 H100s. Google’s Gemini Ultra required an estimated 45 to 55 GWh. On the inference side, OpenAI serving over 1 billion weekly queries consumes roughly 350 GWh per year, exceeding the annual electricity use of 32,000 UK households.
If you are evaluating providers, facility PUE directly affects your cost per computation. A provider at PUE 1.5 wastes 36% more electricity than one at 1.1 for identical workloads. Providers investing in renewable energy for data centers address both cost and carbon dimensions.
Why AI Power Consumption Keeps Growing
Three factors compound the growth. Frontier models grow 2 to 4x larger each generation. Inference volumes expand as companies embed AI into search and enterprise software serving billions of users. New modalities like video generation demand 5 to 10x more compute per request than text.
The IEA projects AI electricity demand could reach 300 TWh by 2030, exceeding the United Kingdom’s annual consumption. Quantization and distillation reduce per-query energy by 50 to 80%, but total demand still rises because deployment scale grows faster.
Frequently Asked Questions
How much power does a single AI data center use?
A modern AI data center draws 50 to 500 MW continuously. The largest planned facilities target 5 GW. Conventional enterprise data centers draw 10 to 30 MW, making AI facilities 5 to 15x more power-hungry per square foot.
What is a good PUE ratio for an AI data center?
A PUE of 1.10 to 1.15 represents best-in-class efficiency with direct liquid cooling. The industry average is approximately 1.58. Google, Microsoft, and Amazon consistently achieve PUE below 1.20.
Will AI data center power consumption decrease as hardware improves?
Each GPU generation delivers 2 to 3x better performance per watt. The NVIDIA B200 provides 2.5x the inference throughput of an H100 at 1.4x the power draw. However, model sizes and deployment volumes grow faster than efficiency improves, so total consumption is projected to increase 3 to 5x by 2030.