How Much Power Does AI Require?

From training frontier models to real-time inference, AI's electricity demand is growing faster than almost anyone predicted.

Artificial intelligence has become the fastest-growing source of new electricity demand in the developed world. The International Energy Agency (IEA) estimates that AI-related electricity consumption could grow 10x between 2025 and 2030, potentially consuming as much power as the entire country of Sweden by the end of the decade. But the power requirements vary dramatically depending on the type of AI workload, the hardware used, and the scale of deployment.

This article breaks down the real numbers behind AI power consumption, from a single GPU to a 500,000-GPU training cluster.

Training vs. Inference: Two Different Power Profiles

AI workloads fall into two broad categories, each with distinct power characteristics:

Training

Training a large AI model involves feeding massive datasets through neural networks over days or weeks. During training, every GPU in the cluster runs at near-maximum utilization, drawing peak power continuously. A single training run for a frontier model can consume 50–100 GWh of electricity — equivalent to the annual consumption of 5,000–10,000 average US homes.

Training power is characterized by:

• Sustained, near-100% GPU utilization for days to weeks
• High power draw with minimal idle periods
• Timeline-driven — faster training means less energy overall, creating an incentive for more powerful (and more power-hungry) hardware

Inference

Inference — running a trained model to generate responses — has a more variable power profile. A single ChatGPT query consumes approximately 10–100x more energy than a Google search query (about 0.3 Wh vs. 0.0003 Wh for search). At scale, inference can be as energy-intensive as training. A popular AI service with 100 million daily users might consume 10–30 MW of inference compute power continuously.

Inference power is characterized by:

• Variable utilization depending on user demand
• Latency-sensitive — requires compute close to users
• Growing rapidly as AI becomes embedded in search, productivity tools, and enterprise software

GPU Power Consumption: The Building Block

Every AI cluster is built from graphics processing units (GPUs), and each generation of GPU draws more power than the last:

Note: TDP (Thermal Design Power) varies by configuration. Actual power draw at full load can exceed rated TDP.

100,000 GPU Cluster: 70–100+ MW

A 100,000-GPU cluster using Nvidia H100s (700W peak each) draws approximately 70 MW of GPU power alone. When you add supporting infrastructure — networking switches, storage servers, cooling systems, and power distribution losses — the total facility power draw reaches 85–110 MW.

For comparison:

• A 100,000-GPU H100 cluster consumes as much power as 60,000–80,000 US homes
• It requires approximately 30–40 MW of cooling capacity (typically liquid cooling)
• At a PUE of 1.15 (typical for a modern liquid-cooled facility), total facility draw = GPU power × 1.15
• Annual electricity cost at $0.08/kWh: approximately $50–70 million

Larger clusters are already being planned. Microsoft and OpenAI have discussed clusters of 500,000 to 1 million GPUs, which would require 400 MW to 1 GW of dedicated power infrastructure.

Real-World AI Clusters

xAI Colossus (Memphis, Tennessee)

Built in an unprecedented 122 days in 2024, the Colossus cluster initially deployed 100,000 H100 GPUs and has since expanded. The facility draws an estimated 50–100 MW and required a dedicated substation from the Tennessee Valley Authority (TVA). Its rapid construction and massive power draw made headlines worldwide and demonstrated that AI compute timelines could far outpace traditional data center construction schedules.

Microsoft's AI Infrastructure

Microsoft has announced plans to spend over $50 billion on AI infrastructure globally. Individual data centers in its AI fleet are designed for 100+ MW each, with clusters of multiple facilities planned in regions with available power and interconnection capacity. Microsoft has signed multiple power purchase agreements (PPAs) specifically to support its AI compute load.

Texas AI Corridor (800 MW planned)

Our tracking identifies the Texas AI Corridor as an 800 MW planned AI campus near Fort Worth, designed specifically for GPU training workloads. The project includes on-site natural gas generation to supplement grid power — a sign that traditional grid interconnection cannot keep pace with AI power demand.

Impact on Grid Planning

The power requirements of AI are creating unprecedented challenges for grid planners:

• Load forecasting uncertainty: Traditional load forecasting models do not account for the possibility of a single customer (a data center developer) requesting 500+ MW of new load in a region that has seen flat demand for decades.
• Capacity planning: Utilities that were planning for 1–2% annual load growth are now facing requests that would increase their peak load by 10–50% within 3–5 years.
• Generation adequacy: In regions like PJM and ERCOT, resource adequacy concerns are being amplified by the addition of hundreds of megawatts of AI compute load without corresponding generation construction.

These challenges are compounded by interconnection queue delays and transformer supply shortages, creating a multi-layered bottleneck that AI companies, utilities, and regulators are all struggling to address.

Key Takeaways

• AI training consumes 50–100 GWh per model — equivalent to thousands of homes' annual use
• Next-gen GPUs draw 700–1,000W each, driving per-cluster power to 100+ MW
• A 100,000-GPU cluster consumes 70–100+ MW of facility power
• Clusters of 500,000–1,000,000 GPUs are being discussed, requiring 400 MW–1 GW
• AI is the primary driver of the new wave of gigawatt-scale data center campuses
• Grid planning must adapt to a new reality of massive, unpredictable load additions