AI Data Center Power Consumption: The Real Numbers for 2026

How Data Center Power Consumption Is Measured

Data center power consumption is measured in two ways: total electricity draw in kilowatts (kW) or megawatts (MW) for a single facility, and terawatt-hours (TWh) for industry-wide annual consumption. One terawatt-hour is one trillion watt-hours — the amount of electricity generated by a large power plant running continuously for roughly five months.

The standard efficiency metric is Power Usage Effectiveness (PUE): total facility power divided by IT equipment power. A PUE of 1.0 means every watt entering the building goes directly to servers. A PUE of 1.5 means that for every 100W powering servers, another 50W goes to cooling, lighting, and power distribution overhead. The global average PUE for data centers is approximately 1.58, according to the Uptime Institute 2024 Global Data Center Survey. Hyperscalers do better: Google reports a fleet-wide average PUE of 1.10; Meta has achieved 1.09 at some facilities.

Why PUE Matters Less for AI Facilities

PUE was designed when cooling was the primary overhead cost. In AI data centers, the dynamic shifts. GPUs draw so much power that they generate heat faster than conventional air cooling can remove it. This forces facilities to install liquid cooling systems — direct liquid cooling, immersion cooling, or rear-door heat exchangers — that themselves consume substantial power. A facility with excellent PUE but insufficient cooling capacity simply cannot run AI racks at full load.

The more relevant metric for AI facilities is power density per rack, measured in kW/rack. Standard colocation facilities support 5-10 kW/rack. AI-optimized facilities support 40-80+ kW/rack. This difference drives every infrastructure decision: floor reinforcement, power distribution, cooling capacity, and utility interconnect size. For context on how this affects physical design, the colocation data center explainer covers how facility operators are retrofitting for AI workloads.

Sources: IEA Energy and AI Report (January 2025), Socomec data center power analysis (2025).

The Global Power Consumption Numbers

The most authoritative source on global data center electricity is the International Energy Agency. Their January 2025 Energy and AI Report, updated April 2025, provides the clearest baseline:

"AI is the most important driver of growth in data center electricity demand and one of the key new energy consumers on a global scale." (International Energy Agency, Energy and AI Report, 2025)

Key figures from the IEA's 2025 Energy and AI Report:

The US numbers, tracked separately by Lawrence Berkeley National Laboratory, show the domestic picture:

• 2018: 76 TWh (1.9% of US electricity)
• 2023: 176 TWh (4.4% of US electricity)
• 2024: 183 TWh (4% of US electricity)
• 2028: 325-580 TWh projected (6.7-12% of US electricity)

That 2028 range is wide because it reflects genuine uncertainty. The lower bound assumes efficiency gains and slower-than-expected AI buildout. The upper bound assumes hyperscaler capacity plans proceed at the announced pace with no offsetting efficiency improvements.

The US capacity picture from Bloom Energy's January 2026 infrastructure report:

• 2025: ~80 GW total US data center demand
• 2028: ~150 GW projected (nearly double in three years)
• 68 hyperscale facilities of 50 MW or larger operational or under construction in the US as of early 2026

Why AI Uses So Much More Power Than Previous Computing

The power difference between AI computing and conventional computing comes down to one thing: parallelism at scale. Training a large language model requires performing billions of matrix multiplications simultaneously across thousands of GPUs. Each GPU is doing useful work every millisecond, drawing full power continuously. Conventional server workloads are highly variable — a web server sits idle most of the time and spikes briefly when requests arrive. An AI training cluster runs at near-100% utilization for weeks.

GPU Power Draw vs Standard Server

A standard 1U server with dual Intel Xeon processors draws approximately 300-500W at peak load. An NVIDIA H100 SXM5 GPU alone draws 700W at maximum thermal design power. A full H100 server (DGX H100) with eight GPUs draws approximately 10,000W (10 kW) from a single server. A rack of 10 DGX H100 servers draws 100 kW — 10-20x a standard server rack.

The newest generation is higher still. NVIDIA GB200 NVL72 rack systems, which package 72 Blackwell GPUs together, draw up to 120 kW per rack unit. Data centers building for GB200 are designing for densities that did not exist in any commercial facility two years ago.

Inference vs Training Power

Training runs are power-intensive but time-bounded. A GPT-4 class training run lasts weeks to months and then ends. Inference — serving model responses to users — runs continuously as long as the product exists. As AI products scale to hundreds of millions of users, inference power demand accumulates indefinitely.

Industry estimates suggest that inference already accounts for 60-70% of AI compute power draw at large AI companies, with training accounting for 30-40%. As model deployment scales faster than new training runs begin, inference will dominate further. The power trajectory is not a bump from training — it is a permanent baseline increase from inference.

The Cooling Problem

Standard air cooling works up to about 30 kW/rack before it becomes impractical. At 60+ kW, air cooling requires so much airflow that it consumes significant power itself and creates hot spots regardless. This is why AI data centers are deploying:

• Direct liquid cooling (DLC): coolant pipes routed directly to GPU heatsinks, removing heat at the source
• Rear-door heat exchangers: liquid-cooled doors that capture exhaust heat before it enters the hot aisle
• Immersion cooling: servers submerged in dielectric fluid

Each cooling method adds infrastructure cost and, in some cases, additional power draw for pumps and chillers.

Hyperscaler Power Consumption and Announced Plans

The five major AI infrastructure spenders — Microsoft, Google, Meta, Amazon, and Oracle — collectively announced over $300 billion in data center capital expenditure for 2025-2026. Power consumption follows capex at roughly proportional scale.

The Number Most Guides Don't Show

Meta's 5 GW Hyperion requirement is often cited as a large number without context. Here is what 5 GW means in practical terms. The US has approximately 900 GW of total installed generating capacity. Meta's single campus requires 0.56% of total US generation capacity running continuously — for one company's one campus.

At $10-12 million per MW for AI-optimized data center construction, 5 GW of capacity represents $50-60 billion in construction cost. That is before equipment. The Hoover Dam generates approximately 2 GW at peak output. Meta's single campus would require the equivalent of 2.5 Hoover Dams running continuously to stay powered.

The grid implications have prompted multiple US states to fast-track new generation approvals specifically for data center demand. Virginia, which hosts the largest concentration of data centers in the world in Northern Virginia, has seen utility applications for new gas peaker plants and solar farms driven primarily by data center contracts.

The Grid Impact of AI Data Center Growth

US electricity demand had been flat or declining for two decades before 2023, as efficiency improvements in lighting, appliances, and manufacturing offset economic growth. AI data center construction has reversed that trend. The US grid was not designed for the scale of new demand now arriving.

The Belfer Center at Harvard Kennedy School published a detailed analysis in 2025 identifying AI data centers as a watershed moment for the US electric grid. The core challenge: building new generation capacity takes 5-10 years. Data center construction takes 18-36 months. The gap creates a period where demand grows faster than supply can be built to meet it.

Consequences already appearing in 2025-2026:

• Power purchase agreements (PPAs) for renewable energy are being signed years in advance, as hyperscalers lock up grid capacity before it is built
• Nuclear power is attracting significant interest from AI companies. Microsoft's deal to restart Three Mile Island's Unit 1 reactor (renamed Constellation Energy's Crane Clean Energy Center) was directly tied to data center power needs, announced September 2023 and operational from late 2024
• Some data center developers are building behind-the-meter generation — natural gas turbines or fuel cells on-site — to avoid grid interconnection queues that now stretch 4-7 years
• States including Georgia, Texas, and Virginia have proposed data center-specific utility rate structures to recover grid upgrade costs from large users

For the broader infrastructure context, the what are AI data centers explainer covers facility design and the AI water usage explainer covers the cooling water dimension of the same infrastructure challenge.

Efficiency Improvements and the 2030 Outlook

The IEA's 945 TWh projection for 2030 is a base case, not a certainty. It assumes current trends in AI adoption, model size, and inference growth continue without major efficiency breakthroughs. Two factors could narrow the gap significantly:

Model Efficiency Improvements

The compute required per unit of AI capability has been falling. DeepSeek R1 demonstrated in January 2025 that comparable reasoning performance to much larger models could be achieved with substantially less compute. If efficiency gains continue at historical rates, the effective AI compute per TWh increases, meaning less energy is needed for the same volume of AI capability.

Hardware Efficiency Gains

NVIDIA's Blackwell generation GPUs deliver approximately 4x better energy efficiency per FLOP compared to the H100 generation for inference workloads. As older H100 infrastructure is replaced by Blackwell and future architectures, the energy per inference request falls. However, the Jevons paradox applies: cheaper, more efficient compute historically leads to more total compute being used, not less.

The IEA Projection Range

The IEA publishes a range, not a single number. Their high scenario reaches approximately 1,100 TWh by 2030; their low scenario reaches around 700 TWh. The 945 TWh base case splits the difference. The range reflects uncertainty about:

• How quickly AI inference scales with model deployment
• Whether efficiency gains in hardware offset demand growth
• Whether regulatory pressure on data center power use materializes in major markets
• The pace of renewable energy and nuclear power deployment specifically for data center supply

What is not in dispute is the direction. US data center electricity consumption grew from 76 TWh in 2018 to 183 TWh in 2024 — a 141% increase in six years. Whatever the exact 2030 number, the trajectory points substantially higher than today.

Frequently Asked Questions

Related Articles