AI in Energy Statistics 2026: How Much Power AI Uses and What It Saves

A single text prompt to Gemini uses about 0.24 watt-hours, less than a television consumes over the same few seconds. An agentic task that involves reasoning can use around 50 watt-hours. Both figures come from the IEA’s April 2026 report, and the gap between them helps explain the rapid growth in AI-related electricity demand.

AI and energy intersect in two ways. AI consumes electricity, making data centers one of the fastest-growing demand sources on many power grids. At the same time, AI helps operate the energy system, from detecting faults on power lines to supporting drilling decisions in oil fields.

We build energy software for utilities, renewable energy operators, and energy service companies, so our work covers both sides. This guide brings together verified AI and energy statistics for 2026. It starts with the electricity AI consumes, then examines how the technology is changing the energy sector. Every figure includes its reporting period and original source.

TL;DR

Energy for AI:

• A median Gemini text prompt uses 0.24 Wh, while an agentic reasoning task uses about 50 Wh.
• Data centers consumed about 485 TWh of electricity in 2025. Under the IEA’s base case, demand will reach around 950 TWh by 2030, or about 3% of global electricity use.
• US data centers consumed 192 TWh in 2024, equal to 4.7% of the country’s electricity use. Their share could reach 11.8% by 2030.
• ERCOT’s queue for large electricity loads grew from about 63 GW to more than 230 GW in 13 months. For comparison, the entire state’s peak demand was 85 GW.
• Data centers accounted for $29.4 billion, or 46%, of PJM capacity charges across its four most recent auctions.

AI for energy:

• 81% of North American utilities use AI, while 78% of US utilities use it to manage data center interconnection demand.
• AI-based fault detection can reduce power outage durations by 30% to 50%.
• With widespread adoption, AI-managed power lines could unlock up to 175 GW of transmission capacity without new construction.
• NESO reduced solar forecasting errors by 20% using an AI nowcasting model developed with Open Climate Fix.
• Saudi Aramco reported $5.3 billion in realized value from AI and other digital technologies in 2025.

How Much Energy Does AI Use per Query?

A simple text prompt consumes less than half a watt-hour. Energy use can rise by 3 orders of magnitude when a task involves reasoning, multiple agents, or video generation.

AI companies began publishing official per-query figures in 2025, and the IEA compiled them in its April 2026 report:

• Google reported that a median Gemini text prompt uses 0.24 Wh.
• OpenAI estimated that an average ChatGPT query uses 0.34 Wh.
• Microsoft Research measured a median of 0.34 Wh for frontier models with more than 200 billion parameters.
• Google reduced the energy used per prompt by a factor of 33 during the 12 months ending in May 2025.

At these levels, basic text chat isn’t the main driver of electricity demand. The IEA estimates that replacing every conventional internet search with a simple AI text query would consume less than 4 TWh per year. That is under 1% of the electricity data centers use today.

More complex workloads consume far more energy. The IEA estimates that a medium-sized language model uses about 0.05 Wh of GPU electricity to generate a text response, while a reasoning model uses around 1.14 Wh. An agentic task involving reasoning and four to six model calls uses about 50 Wh, roughly 1,000 times as much as the basic text response.

Video generation can consume hundreds or thousands of times more energy than text, depending on the video’s length and resolution.

This is the central tension in AI energy use in 2026. Individual tasks are becoming more efficient, but workloads are shifting toward the most energy-intensive applications.

How Much Electricity Do Data Centers Use?

Data centers consumed slightly more than 1.5% of the world’s electricity in 2025. Their electricity use grew more than 5 times faster than global demand.

Worldwide

According to the IEA’s April 2026 update, data center electricity consumption increased by 17% in 2025. Demand from AI-focused data centers grew by about 50%, while total global electricity demand rose by 3%.

In 2024, the United States accounted for 45% of global data center electricity consumption. China accounted for 25%, while Europe represented 15%.

Data centers are expected to drive slightly less than 10% of global electricity demand growth through 2030. Their local impact is much greater. In states and metropolitan areas with large data center clusters, these facilities can account for 20% to 30% of total electricity demand.

United States

The most detailed US forecast comes from Lawrence Berkeley National Laboratory’s 2025 Update, published for the Department of Energy in June 2026:

• US data centers consumed 192 TWh in 2024, equal to 4.7% of the country’s electricity use and 14% more than in 2023.
• Consumption is projected to rise 22% in 2025 and another 29% in 2026.
• Under the Reference Case, demand reaches 464 TWh in 2028 and 649 TWh in 2030. The 2030 figure represents 11.8% of US electricity use.
• The full range of scenarios for 2030 runs from 521 to 843 TWh, or 9.5% to 15.3% of US electricity.
• Data centers could account for 33% of all US electricity demand growth between 2024 and 2030.
• AI servers could consume 55% of all data center electricity by 2030.

For comparison, LBNL estimated that US data centers consumed about 76 TWh in 2018, equal to 1.9% of the country’s electricity use.

LBNL’s 2030 Reference Case is higher than the IEA’s US projection of about 430 TWh. The difference mainly comes from the two organizations’ assumptions about how many AI chips will be shipped and deployed.

LBNL also translates electricity consumption into grid capacity requirements. At an average utilization rate of 50%, US data centers would need about 148 GW of interconnection capacity by 2030. Meeting that requirement would mean adding an average of 17.4 GW of capacity each year.

Where AI Load Hits the Grid

National averages hide the main pressure point. Data centers develop in clusters, where electricity demand can grow faster than utilities can add generation, transmission, and grid capacity.

Connection Queues

Texas shows how quickly demand can increase. In December 2024, ERCOT’s queue for large loads stood at about 63 GW, with data centers accounting for roughly 3-quarters of the requests. The queue reached 130 GW by April 2025 and more than 230 GW by January 2026. For comparison, the state’s record peak electricity demand is 85 GW.

Other regions are reaching similar limits. In Alberta, data center connection requests totaled 16 GW, exceeding the province’s peak demand. The grid operator responded by imposing a temporary 1.2 GW cap.

Expanding the grid takes time. The IEA reports that new grid connections can take five to ten years in many regions. Transformers can take two to three years to procure, while new gas turbines may take around five years. Most projects in connection queues won’t be built, but the uncertainty still makes grid planning more difficult.

Electricity Prices

The clearest evidence of data centers affecting electricity prices comes from PJM, the grid operator serving 13 US states and Washington, D.C. Its independent market monitor attributed $6.3 billion, or 38%, of the latest capacity auction’s $16.4 billion in charges to data center demand. Across the 4 most recent auctions, data centers accounted for $29.4 billion, or 46%, of total capacity charges.

Capacity prices increased from $28.92 per megawatt-day in 2024/2025 to $333.44 in 2027/2028. In utility surveys, 83% of US utility innovation leaders said the cost of infrastructure needed to support data centers was being passed on to households.

PJM’s experience isn’t universal. The IEA found that in regions with spare capacity, steady new demand can spread fixed system costs across more electricity users and reduce average prices. Prices are more likely to rise when supply is already tight, and demand grows faster than utilities can add generation and grid infrastructure.

What Powers AI

Technology companies have become some of the world’s largest buyers of new energy capacity. They sign a large share of corporate renewable power agreements while also securing natural gas generation and nuclear energy.

The scale of investment helps explain this shift. According to the IEA, capital spending by the largest technology companies exceeded $400 billion in 2025 and is expected to rise by another 75% in 2026. 5 companies alone now spend more than the entire world invests in oil and gas production.

The IEA’s April 2026 report shows how technology companies plan to meet their growing energy needs:

• Technology companies signed about 40% of all corporate renewable power purchase agreements worldwide in 2025.
• Data center agreements to buy power from small modular reactors grew from about 25 GW at the end of 2024 to 45 GW at the end of 2025. However, the first units aren’t expected to begin operating until around 2030.
• Companies are also securing power from existing nuclear plants. Microsoft’s agreement covers about 835 MW from the planned restart of Three Mile Island Unit 1 in 2027. AWS has an agreement with Talen for up to 1,920 MW from the Susquehanna nuclear plant.
• Between 15 and 27 GW of on-site natural gas generation could supply data centers by 2030, mostly in the United States.
• Orders for gas turbines increased by 70% in 2025.

Why AI Data Centers Are Different

AI hardware is changing data center electrical design. Server power density increased elevenfold between 2020 and 2025 and is expected to quadruple again by 2027.

A rack based on Nvidia’s 2020 Ampere architecture consumed about 13 kW. A Blackwell rack uses around 130 kW, while the announced Rubin architecture could reach 600 kW. At its peak, a single cabinet roughly the size of a refrigerator could draw as much power as about 65 households.

AI workloads also create rapid swings in electricity demand. Training and inference clusters have produced fluctuations of tens of megawatts in less than a second, about half their rated capacity.

The IEA expects data centers to install between 20 and 25 GW of battery storage by 2030. With the right incentives, that capacity could help data centers support the grid rather than simply draw power from it. Managing these rapid changes requires energy storage management software that responds in milliseconds.

Can AI Pay Back Its Own Energy Bill?

In theory, AI could offset part of its own energy use. The IEA estimates the technology’s potential savings are roughly on the same scale as its consumption. The difference is timing: data center demand is growing now, while most of the savings depend on widespread adoption that hasn’t happened yet.

The left column reflects demand that is already moving through construction and grid connection pipelines. The right column shows what existing AI applications could deliver if the energy sector overcomes the barriers to wider deployment. The IEA calls this its Widespread Adoption scenario.

The rest of this article examines how close the energy sector has come to that scenario.

The IEA’s 2025 report highlighted one comparison directly. AI could unlock up to 175 GW of capacity on existing transmission lines, more than the projected increase in data center load by 2030.

AI Adoption in Utilities

Most North American utilities now use AI in at least one part of their operations. In just one year, the main reason has shifted from meeting climate targets to managing growing electricity demand from data centers.

Two large surveys show the change.

• Itron surveyed 500 electric utility executives in the United States and Canada. It found that 81% already use AI, while 41% have fully integrated AI-related technologies. A year earlier, only 27% expected to reach that level of integration within five years. Grid optimization was the leading application, cited by 57% of respondents, while 51% ranked demand forecasting among their top three uses.
• National Grid Partners surveyed 134 US utility innovation leaders between May and July 2026. It found that 78% were deploying at least one AI application to manage data center-related interconnection demand. Another 74% said data center load growth was affecting grid reliability.

The shift in priorities is significant. Reliability now ranks among the top 3 concerns for 73% of utility leaders, up from 43%. Over the same period, the share prioritizing net zero fell from 54% to 16%. In addition, 87% said rate-case frameworks created before the AI demand boom limit the returns utilities can earn from innovation.

These adoption figures show how widely utilities use AI, but not how deeply they have integrated it. A utility running a single forecasting model counts the same as one using AI throughout its control room. The IEA’s global assessment is more cautious, as the barriers section below explains.

AI on the Grid

The grid offers one of the clearest economic cases for AI in energy, partly because building new transmission lines can take four to eight years.

The IEA’s 2025 analysis identified three potential benefits:

• AI-based fault detection can reduce outage durations by 30% to 50% by finding and locating faults more quickly.
• Remote sensors and AI-based line management could unlock up to 175 GW of transmission capacity without building new lines.
• Widespread use of AI in power plant operations and maintenance could save up to $110 billion a year by 2035.

The IEA’s 2026 update adds that AI and other grid-enhancing technologies can help utilities use existing infrastructure better while they wait for grid expansions. AI can also monitor transformers and other equipment for early signs of failure, reducing unplanned outages.

However, greater AI use creates new security risks. The technology can make cyberattacks easier to carry out, while a more digital grid gives attackers more potential entry points.

These figures represent potential benefits, not savings already achieved. No public source currently tracks how much of that value utilities have captured.

Realizing these benefits depends on integrating sensor feeds, asset records, and outage logs that utilities often store in separate systems. This data integration is central to most smart grid software development and energy data engineering projects.

AI in Renewable Energy Forecasting

Forecasting is one of AI’s clearest, proven uses in renewable energy. Every megawatt of forecast error must be covered by reserve capacity, and that reserve costs money.

Solar in Great Britain

The strongest recent evidence comes from Britain’s National Energy System Operator (NESO) and the nonprofit Open Climate Fix. Their PV Nowcasting project received £500,000 through NESO’s innovation allowance. It combines machine learning with satellite imagery, weather data, and live solar output.

The project delivered several measurable results:

• Mean absolute error fell by 20%, meeting NESO’s target. Forecasts for the next zero to eight hours were also about 40% more accurate than those produced by the previous models.
• After the system entered operational use, forecast error fell by about 50 MW, according to NESO’s innovation annex.
• NESO has used the model in its control room since 2025. Open Climate Fix reports that its Quartz Solar tool is 2.8 times more accurate than earlier tools and helps avoid around £30 million in annual imbalance costs.

The source of each figure matters. NESO, the grid operator, provides the performance results. The £30 million in avoided costs and the 2.8-times accuracy claim come from Open Climate Fix itself.

Wind

The most frequently cited wind forecasting example is older. DeepMind and Google applied machine learning to 700 MW of wind capacity in the central United States. The system forecast electricity output 36 hours in advance.

The companies reported that these forecasts increased the value of the wind energy by about 20% compared with making no advance commitments to the grid. However, Google didn’t publish the method used to calculate that increase.

For teams planning renewable energy software, the NESO case shows where the commercial value comes from. Better forecasts matter because trusted predictions let operators buy less reserve capacity.

AI in Oil and Gas

Oil and gas companies adopted high-performance computing earlier than almost any other part of the energy sector. Saudi Aramco now goes a step further by reporting the financial value it attributes to AI and other digital technologies each year.

The IEA tracks the sector’s computing capacity. In 2000, oil and gas companies operated 11 of the world’s 500 fastest supercomputers. By 2024, that number had increased to 24. Their combined computing capacity grew by nearly 70% a year, faster than the overall supercomputing market.

The main applications include subsurface data processing, reservoir simulation, remote operations, predictive maintenance, leak detection, and regulatory compliance.

Saudi Aramco provides the clearest company-reported figures. It uses the term “technology realized value” for higher revenue, lower capital spending, or reduced operating costs attributed to AI and other digital tools:

• Aramco reported $5.3 billion in realized value for 2025, bringing the total since 2023 to $11.3 billion, according to its full-year results.
• The company attributed $1.8 billion directly to AI in 2024.
• By the end of 2025, Aramco had identified 442 AI use cases, deployed more than 200, and had over 100 in development.

Aramco identifies reservoir mapping, drilling, well productivity, and maintenance as important sources of value. Better asset data has also helped the company reduce spending on pipeline anti-corrosion materials.

The $5.3 billion figure covers AI and other digital technologies, and Aramco doesn’t publish the method used to calculate it. Therefore, treat it as a company-reported estimate rather than an independently audited return on investment.

Aramco’s results also depend on an asset few companies can match: nearly a century of geological and operational data. Operators with shorter data histories may gain more by starting with oil and gas software that brings existing sensor and well data into one system before applying advanced AI.

What Holds AI Back in the Energy Sector

Energy companies have more data than their level of AI adoption suggests. According to the IEA, the main barrier isn’t the technology itself. It’s the shortage of people with the skills to use it.

The IEA’s 2026 survey of energy companies ranked limited digital skills as the biggest obstacle to wider AI adoption. Fragmented data came next, followed by concerns about data protection, privacy, and cybersecurity.

Other findings show how much of the supporting infrastructure is still missing:

• Open electricity data policies cover only 10% of global electricity consumption.
• Policy frameworks that support AI in the energy sector cover less than half of global energy demand.
• Around 1% of energy-related patents reference AI.
• Only 2.3% of energy startups have an AI-related value proposition, compared with 7% in life sciences.

Utility surveys show the same constraints. In National Grid Partners’ 2025 survey, 61% of utility leaders said a shortage of AI talent could slow their plans. In 2026, 87% said existing rate-case frameworks limit the returns utilities can earn from innovation.

This means a utility can prove that an AI project works and still struggle to recover its investment through regulated rates.

AI Regulation in Energy

In the EU, AI used to protect the power grid can be classified as high-risk. AI used only to optimize its performance generally is not.

Point 2 of Annex III of the EU AI Act classifies AI as high-risk when it serves as a safety component in the management or operation of critical digital infrastructure, road traffic, or the supply of water, gas, heating, or electricity.

The Digital Omnibus on AI, which entered into force on 27 July 2026, narrowed this definition. AI used solely for user assistance, performance optimization, service efficiency, or process automation isn’t considered a safety component.

The amendment also delayed the compliance deadline. High-risk requirements for Annex III systems will now apply from 2 December 2027 rather than 2 August 2026.

Grid operators also follow a different registration rule. Unlike most high-risk systems, those covered by point 2 are registered in non-public national registers instead of the public EU database. In Germany, the Bundesnetzagentur maintains this register.

For energy software teams, the practical question is whether a model’s failure could physically endanger the grid or people. A load forecasting model that informs a planner is likely outside the high-risk category. A model that automatically activates protection equipment is likely within it.

What the Numbers Say Together

AI’s energy impact falls into two categories: measured demand and modeled savings. Consumption is already visible. Data center electricity demand grew by 17% in 2025, US facilities consume 4.7% of the country’s electricity, and capacity markets such as PJM are already pricing in the additional load.

The savings are less certain. The IEA estimates that AI could unlock 175 GW of transmission capacity and save up to $110 billion a year in power plant operations. Both figures depend on a level of adoption the energy sector hasn’t reached.

Where results have been measured, they are meaningful but specific. NESO reduced solar forecasting errors by 20%, while Saudi Aramco reported $5.3 billion in realized value from AI and other digital tools.

For utilities, renewable operators, and energy companies, the challenge over the next few years is to turn projected benefits into verified operational results. Our AI and ML development for energy software supports that process, from building reliable data pipelines to developing models that operators can trust in the control room.

We’ll update this page as the IEA, LBNL, and grid operators release new data. Every figure includes its reporting period and original source.

FAQs