AI data centres and the power problem how much energy does AI really use

[QuantumToken65]

TL;DR

• The International Energy Agency estimated that data centres used about 415 TWh of electricity in 2024, roughly 1.5 percent of global electricity, and projected this could more than double to around 945 TWh by 2030
• In its 2026 update, the IEA reported that data centre electricity use rose about 17 percent in 2025, with AI focused data centres growing around 50 percent
• In the US, Lawrence Berkeley National Laboratory estimated data centres used 4.4 percent of electricity in 2023, potentially rising to between 6.7 and 12 percent by 2028
• Per query energy is smaller than many assume: Google reported in 2025 that a median Gemini text prompt used about 0.24 watt hours, but total demand is growing because usage is exploding
• The real pressure points are local: grid connections, generation capacity, water and community impacts in places where data centres cluster

Why the question matters

Few technology debates generate as much heat, in every sense, as AI's energy use. Headlines warn that AI will overwhelm power grids, while others insist the concerns are exaggerated. Both sides can find numbers to support their case. Understanding what is known, what is estimated and what is uncertain helps cut through the noise

This thread pulls together the most widely cited data from the International Energy Agency, US national laboratories, companies and researchers. It explains how data centres use electricity, why AI changes the picture and how individual queries compare with total demand. It also covers efficiency improvements and the local challenges that matter most. The aim is a clear, sourced picture rather than alarm or dismissal

A word of caution about figures. Estimates of AI energy use vary widely because companies disclose limited data, methods differ and the field changes fast. Where possible, I cite the source and the year. Treat every number as an estimate with uncertainty, and expect revisions

How data centres use electricity

A data centre is a building full of servers, storage and networking equipment, plus the cooling, power conversion and backup systems that keep them running. Servers consume most of the electricity, and cooling is the next largest share. The industry measures overhead using power usage effectiveness, or PUE, the ratio of total facility energy to IT equipment energy. A PUE of 1.2 means 20 percent extra energy goes to cooling and other overheads

AI changes the composition of data centre load. Training and running large AI models relies on accelerators such as GPUs, which draw far more power per server than traditional computing and are packed densely into racks. Some AI racks now draw over 100 kilowatts, compared with a few kilowatts for traditional servers. That density pushes operators towards liquid cooling and new electrical designs

Global numbers

The International Energy Agency's Energy and AI report, published in April 2025, estimated that data centres consumed about 415 terawatt hours of electricity in 2024, around 1.5 percent of global electricity use. The United States accounted for the largest share, followed by China and Europe. In its base case, the IEA projected data centre demand rising to about 945 terawatt hours by 2030, slightly more than Japan's total electricity consumption today. AI was identified as the most important driver of that growth

In 2026, the IEA reported that data centre electricity use grew about 17 percent in 2025, far faster than overall electricity demand growth of about 3 percent, and that electricity use by AI focused data centres rose around 50 percent. It reiterated that total data centre demand was set to roughly double by 2030, with AI focused facilities growing even faster. The IEA also noted tightening bottlenecks in grids, equipment and supply chains. These findings make AI a significant, though still modest, share of global electricity

Context helps. Even at 945 terawatt hours, data centres would account for a single digit share of global electricity demand, alongside much larger uses such as industry, buildings and transport, including electric vehicles. Globally, AI is one of several drivers of rising electricity demand rather than the dominant one. Locally, the picture can be very different

The United States and local hotspots

In December 2024, Lawrence Berkeley National Laboratory published a report for the US Department of Energy estimating that data centres used about 4.4 percent of US electricity in 2023 and could reach between 6.7 and 12 percent by 2028, depending on AI growth and efficiency. That range reflects significant uncertainty. Either end would represent substantial growth in demand within a few years. Utilities in several states have revised their forecasts sharply upwards

Data centres cluster in specific regions with good connectivity, land, power and tax incentives. Northern Virginia hosts one of the world's largest concentrations, and states including Texas, Georgia and Ohio have seen rapid growth. In Ireland, official statistics showed data centres used about a fifth of the country's metered electricity in 2023. In such places, data centre demand strains grid connections and generation capacity even if global shares remain small

How much energy does a single AI query use

Estimates of per query energy have varied widely, partly because they depend on the model, hardware and workload. In August 2025, Google published a technical paper estimating that the median text prompt in its Gemini apps used about 0.24 watt hours of energy, emitted about 0.03 grams of carbon dioxide equivalent and consumed about 0.26 millilitres of water. Google said energy per median prompt had fallen by a factor of 33 over the preceding twelve months through efficiency improvements. Comparable figures from other companies remain limited

For perspective, 0.24 watt hours is roughly the energy a typical microwave uses in about a second, or a modern LED bulb in a few minutes. Longer tasks, image and video generation, and reasoning models that think through many steps can use substantially more per request. Agents that run for minutes or hours consume more still. Per query figures therefore vary by orders of magnitude depending on the task

The key insight is that small per query energy can still add up to large totals when multiplied by billions of requests, plus training. Total demand is rising because usage is growing explosively and because new capabilities such as reasoning and video are more compute intensive. Efficiency gains are real but are being outpaced by growth in demand, at least for now. This pattern, where efficiency enables more use, is often called the rebound effect

Training versus inference

Training a frontier model requires enormous computation over weeks or months on thousands of accelerators. Public estimates for the energy used to train the largest models run into many gigawatt hours, though companies rarely disclose exact figures. Training happens relatively infrequently for each model. Inference, the work of answering user requests, runs continuously and at growing scale

Over time, inference is expected to dominate AI energy use. As more people and businesses use AI daily, and as agents perform longer tasks, inference demand compounds. This shifts attention towards efficient serving: smaller models, better hardware, caching and smarter routing that uses large models only when needed. Choosing the right sized model for a task is one of the simplest ways to save energy

Efficiency improvements

Hardware efficiency keeps improving. Each generation of AI accelerators delivers more computation per watt, and specialised inference chips push efficiency further. Data centre design is also improving, with liquid cooling, higher voltage power distribution and better airflow reducing overhead. Leading hyperscale facilities report PUE values well below industry averages

Software and model efficiency matter as much as hardware. Techniques such as quantisation, distillation, mixture of experts architectures and speculative decoding reduce the computation per answer. Smaller models now match the performance of much larger models from a year or two earlier on many tasks. These gains help explain Google's reported reduction in energy per prompt

Operational choices make a difference. Scheduling flexible workloads such as training for times and places with abundant clean electricity, and designing data centres to reduce demand during grid peaks, can ease strain. Some operators reuse waste heat for district heating, particularly in northern Europe. These approaches turn data centres from pure loads into more flexible grid participants

Grids, generation and communities

The most immediate constraint is often the grid connection. Utilities in many regions report long queues for large new connections, and building transmission lines and substations takes years. Some data centre developers have turned to on site generation, including gas turbines, to bypass delays, which raises emissions and air quality concerns. Others are signing deals for new renewable, storage and nuclear capacity

Water use is another local issue. Many data centres use water for evaporative cooling, and facilities in water stressed regions face scrutiny. Operators increasingly use closed loop or air cooling designs and report water usage effectiveness. Communities near large campuses also raise concerns about noise, land use and electricity prices

Clean energy commitments are under pressure. Several large technology companies have reported rising emissions despite climate goals, citing data centre growth. They are investing in renewable power purchase agreements, storage and nuclear energy, including restarting retired reactors and backing small modular reactors, discussed in another thread here. Whether clean supply can keep pace with demand is one of the central energy questions of the decade

AI as part of the solution

AI can also help reduce energy use elsewhere. The IEA's Energy and AI report discussed opportunities for AI to optimise power grids, improve building efficiency, accelerate materials discovery for batteries and solar, and enhance industrial processes. Some estimates suggest these savings could be substantial if widely adopted, though the IEA has cautioned that adoption of such applications is lagging behind growth in AI energy demand. Realising the benefits requires deliberate effort

Balanced accounting is essential. Claims that AI will solve climate change and claims that it will doom grids both oversimplify. The honest picture is a significant and growing new load, concentrated in specific places, alongside real but uneven opportunities for efficiency and optimisation. Policy, investment and transparency will shape which effects dominate

What to watch

• Updated IEA and national laboratory estimates of data centre demand
• Disclosure of per query and per model energy figures by more AI companies
• Grid connection timelines and new transmission investment in data centre hotspots
• Growth of on site gas generation versus clean energy procurement
• Hardware and model efficiency improvements, especially for inference
• Local policy responses on water, land use and electricity tariffs

Closing thoughts

AI's energy use is real, growing and increasingly important for grids and climate goals, but it is not yet a dominant share of global electricity. Per query energy has fallen sharply, while total demand rises with explosive usage and more compute intensive tasks. The biggest challenges are local, where data centres cluster and grids strain. Efficiency, clean energy investment, flexibility and transparency are the main levers

If you work in energy, data centres or AI infrastructure, please share your perspective in the replies. Local stories about grid connections and community impacts would be especially valuable. Please include sources for any figures you share


Key references for citation

• International Energy Agency, Energy and AI, April 2025, and 2026 update on data centre electricity use
• Shehabi and colleagues, Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report, December 2024
• Google, Measuring the environmental impact of delivering AI at Google scale, August 2025
• Central Statistics Office Ireland, data centres metered electricity consumption, 2023
• Uptime Institute, global data center survey