AI data centers are consuming electricity at unprecedented scales, with a single large facility drawing enough power to supply a small city. Global data center electricity consumption reached 415 terawatts per hour (TWh) in 2024, representing 1.5% of total global electricity use, according to the International Energy Agency's Energy and AI report. The IEA projects that figure will more than than double to 945 TWh by 2030, roughly equivalent to Japan's entire annual electricity consumption today . This explosive growth is reshaping how companies plan AI infrastructure, because power availability has quietly become the defining constraint on AI expansion, surpassing even chip availability.

Why Are AI Data Centers So Power-Hungry?

The answer lies in the hardware itself. A single NVIDIA H100 graphics processing unit (GPU) draws 700 watts of power. A rack containing eight H100s draws 5.6 kilowatts just for the processors, before accounting for cooling, networking, or storage. Scale that to a 50,000-GPU training cluster and the power draw approaches 35 megawatts, equivalent to powering a small city . This is fundamentally different from traditional data centers, which were designed for occasional, bursty workloads like web hosting and storage. AI data centers run constant, intensive computation 24 hours a day, 7 days a week.

The density difference is staggering. Traditional data centers operate at 5 to 10 kilowatts per rack. Cloud AI clusters reach 20 to 40 kilowatts per rack. AI hyperscale facilities hit 40 to 120 kilowatts per rack. And planned AI gigafactories exceed 100 kilowatts per rack . This 5 to 10 times increase in power density per rack is what drives the cooling complexity, grid demand, and infrastructure costs that now define AI deployment planning.

How Does Power Consumption Change Across AI's Lifecycle?

Most discussions treat AI power usage as a single number, but that's misleading. Electricity consumption happens across distinct phases, and confusing training with inference leads to fundamentally wrong infrastructure decisions . Training a large AI model requires 25 to 50 megawatts of power during the training run itself, sustained over weeks or months. But training is a temporary spike. Inference, the phase where the trained model answers user queries, is the real power baseline.

Inference now accounts for approximately 80 to 90% of total AI computing today, and is expected to represent 75% of total AI energy demand by 2030 . One of the fastest-growing inference loads in enterprises is AI-powered search and retrieval-augmented generation (RAG) systems, which query live knowledge bases continuously. Unlike traditional search, these systems generate a new large language model (LLM) response for every single query, meaning GPU time is consumed on every search at scale. This persistent demand fundamentally changes how infrastructure must be designed and sized.

Where Does All That Power Actually Go?

Understanding power distribution helps explain why efficiency matters so much. In a typical AI data center, GPUs and compute systems consume 55 to 65% of total power, the dominant load that scales directly with GPU density. Cooling infrastructure accounts for 30 to 40% of power consumption, with older or air-cooled facilities consuming even more. Storage and networking make up roughly 5 to 10%, a growing share as high-speed fabric and NVMe storage become standard . Cooling alone consumes as much electricity as a small city in large AI facilities, which is why the industry is rapidly shifting cooling technologies.

How to Plan Your Data Center Cooling Strategy for AI Workloads

• Air Cooling Limits: Traditional air cooling supports up to 20 kilowatts per rack, which is insufficient for modern AI deployments and represents the baseline many enterprises still operate within.
• Liquid Cooling Systems: Liquid cooling pipes chilled water directly to server components and supports up to 100 kilowatts per rack, making it the standard for most hyperscale AI facilities today.
• Immersion Cooling Technology: Submerging servers in dielectric fluid supports 200 or more kilowatts per rack and reduces Power Usage Effectiveness (PUE) to near 1.0, representing the frontier of cooling efficiency.
• Direct-to-Chip Cooling: This approach targets heat removal at the GPU die itself and is being deployed by Google, Microsoft, and AWS in new builds for maximum efficiency.

Power Usage Effectiveness, or PUE, remains the critical efficiency metric for comparing data center performance. Best-in-class hyperscale facilities achieve 1.1 to 1.2 PUE, meaning they use only 1.1 to 1.2 watts of total facility power for every watt of compute power. The global average sits around 1.5 to 1.6, indicating significant room for improvement .

The scale of this challenge is becoming visible in regional power grids. In Virginia, data centers already consume nearly 25% of the state's total electricity supply, according to Virginia Business . A single large AI campus can reach 1 gigawatt of power consumption, roughly equal to the power output of one nuclear reactor. In 2025 alone, more than 10 gigawatts of new AI data center capacity were announced globally, creating a power bottleneck that utilities and governments are struggling to solve .

What makes this challenge even more complex is how enterprises are scheduling AI workloads. AI scheduling agents for remote teams and asynchronous AI tools for global operations distribute compute jobs across time zones. While this sounds efficient, it actually means AI data centers must maintain near-full GPU utilization across all 24 hours rather than peaking during business hours. This flattens the demand curve but eliminates opportunities for load shifting or grid-friendly scheduling .

The bottom line is clear: power availability, not chip availability, has become the primary constraint on AI infrastructure expansion. Companies planning AI deployments in 2026 and beyond must prioritize grid access and power infrastructure before purchasing GPUs. The most advanced chips in the world are worthless without the electricity to run them.