Powering AI

This issue brief is part of the GeoTech Center’s “Atlantic Council Commission on Artificial Intelligence: US leadership in the age of AI” report, which offers an action-oriented roadmap for strengthening US domestic AI capacity, aligning with allies, and sustaining global leadership.

Electricity supply has shifted from being of peripheral interest to central for AI and technology companies. As AI companies release new, larger models and AI adoption accelerates, potential electricity shortfalls are looming larger. Accordingly, US AI leadership will depend, in part, on how the country meets surging AI-related demand. Three challenges are critical for policymakers and industry: building out the power system with new infrastructure and innovative technologies, making more efficient use of power systems, managing the gap between projected and actual AI load, and meaningfully addressing social license concerns.

A primary goal for US AI is to build an electricity system that can deliver sufficient electricity to power AI workloads. Electricity is already a bottleneck for US data centers, and insufficient or unreliable supply could slow or derail US AI ambitions. Although AI’s trajectory and ultimate impact remain uncertain, the technology will likely reshape US electricity infrastructure and markets. With the existing grid already due for replacement, the system is primed for a period of fundamental change.

The electricity sector has little historical experience managing rapidly growing demand and dynamic operating conditions, but it is now contending with immense uncertainty and transformation. From 2010 to 2021, US electricity demand was relatively flat, as demand rose only 1.4 percent to 3,805 terawatt hours, for a compound annual growth rate of only 0.13 percent. In contrast, US power demand has risen consistently in recent years, driven by electrification, industrial revitalization, and data center consumption for cloud computing needs and AI. While residential and industrial electricity demand are up 3 percent and 4 percent, respectively, from 2021 levels, commercial electricity demand rose 12.4 percent from 2021 to 2025. Data center hubs unsurprisingly outpaced average national growth rates: Virginia, which hosts the largest data center market in the world by operational capacity, accounted for 15 percent of the increase in total commercial demand growth, while other emerging data center hubs in states such as Texas and Georgia also outpaced the national average.

Under the right conditions, data centers can have a net positive impact on electricity affordability. The large and consistent power demand profile of data centers, along with hyperscalers’ ability to support investments in grid services and upgrades, means that new data center load can lower average retail electricity prices when costs are allocated according to the “beneficiary pays” principle. Furthermore, in US states with high demand growth over the past five years, average inflation-adjusted retail prices have generally declined or risen less than in states with flat or falling demand, where prices have generally increased. The net impact of data center load growth depends largely on the characteristics of the existing system and new load, as well as how utilities design rates, conduct system planning, and allocate costs.

Even as electricity demand is rising, supply and transmission are straining to keep pace. Much of the grid is at end of life: more than half of US coal plants were built in the Jimmy Carter presidency or earlier; coal plant maintenance has been deferred ahead of retirements; 55 percent of in-service distribution transformer units are more than thirty-three years old and are approaching the end of their service life; and much of the US electricity grid, including transmission, was built around the time of the moon landing.

The United States is struggling to bring new generation and transmission capacity onto the grid, especially in comparison to China. In 2025, the United States deployed 55 gigawatts (GW) of new generation to the grid. Conversely, China deployed 543 GW of new generation capacity that year, or more than the United States has cumulatively installed since 2008. According to Chinese sources, in 2023, China built about 40,445 kilometers (or 25,133 miles) of new transmission lines greater than 220 kilovolt (kV); the US built 888 line miles of high-voltage (more than 345 kV) transmission lines in 2024, the last full year for which data are available. The United States is not falling behind China in the electricity race; it is not even competing.

Taking all this together—growing demand, constrained supply, the need to replace aging infrastructure, and insurance risks from climate change—US electricity prices are rising and might begin soaring. While electricity prices changed little from 2010 to 2021, in real terms, they are now outpacing inflation and have risen nearly 21 percent from 2021 to 2024. In addition, if inflation and interest rates rise amid global supply chain outages or other crises, especially those surrounding oil, then capital costs will climb, which ultimately increases the total cost of the infrastructure buildout and can cause projects still under development to become unviable. With power demand projected to rise due to AI and other factors, incremental generation capacity deployment facing growing headwinds, and the grid nearing its end of life, most analysts believe prices will continue their march upward.

Rising electricity prices could spark backlash, constraining AI development. While electricity prices have traditionally not been as impactful for consumers as gasoline or groceries, nor as psychologically salient, public opinion polls suggest consumers are increasingly sensitive to electricity prices. Importantly, the effects of rising electricity will be uneven, falling most acutely on individuals earning lower incomes and living in rural areas. Electricity costs will figure increasingly prominently in political debates over AI.

US AI efforts will hinge, to a significant degree, on the nation’s ability to provide reliable and affordable electricity and to adopt pragmatic approaches. If the United States fails to pay sufficient attention to the energy landscape on which future AI growth rests, it runs the risk of higher prices, greater outages, more public backlash to AI, and insufficient energy to effectively compete with China.

Electrons as an AI bottleneck

Energy is a potential bottleneck for AI but not a major cost driver. Amortization of hardware and labor account for the bulk of costs for both AI training and inference workloads; energy comprises only 2–6 percent of costs for training. Nevertheless, impressive technical advances in AI, including longer model “thinking,” suggest that inference workloads will become more electricity intensive. Whether for training or inference workloads, data centers cannot run without electricity. If US electricity supply-demand balances are disrupted, then AI will suffer.

Can data centers curtail electricity consumption at key moments? Some level of data center curtailment might not be catastrophic, especially for inference workloads. A Duke University study indicated that data centers could curtail inference operations during peak-of-the-peak demand periods, while some companies at the AI-energy nexus are aiming to reposition data centers as “flexible, intelligent grid assets.” Some AI models already appear to be headed toward surge pricing, while some consumer-facing inference workloads can tolerate greater latency without severely degrading the product. When training AI models, however, power interruptions could hold severe consequences, especially for frontier models.

AI inference and training workload capacity factors might converge. Due to AI advances, models can solve more complex tasks. In turn, inference workloads can extend across multiple days or weeks. If inference workloads indeed prove more electricity intensive or compute constrained, AI model providers might employ tiered or surge pricing: some are already offering two times capacity outside of certain peak demand windows. Computational resources, memory chips, and EUV machines are key AI bottlenecks. But with total planned US compute capacity in 2030 thirty times larger than mid-2025 levels, electricity constraints could also throttle both model training and consumer inference usage, constraining US AI efforts.

Social license to operate: Electricity affordability and environmental impacts

Increasing demand, constrained supply, and grid replacement needs could lead to rising prices and reliability concerns. Political concerns around AI, data centers, and electricity prices might sharpen if affordability concerns intensify.

Because most social license concerns revolve around affordability, it’s worth contextualizing electricity prices alongside other consumer expenditures. Gasoline for vehicles cost the average consumer unit—as defined by the US Bureau of Labor Statistics—about $200 per month in 2024. Electricity costs totaled about the same amount. Food at home ran the average household about $500 per month during the same year. If electricity prices rise, the US AI buildout could face a political backlash. Electricity affordability is moving to the center stage in US politics: the Trump administration unveiled its Ratepayer Protection Pledge, while Pennsylvania Governor Josh Shapiro sued the Pennsylvania–New Jersey–Maryland Interconnection (PJM), the regional transmission organization that oversees transmission system reliability and coordinates electricity markets across thirteen states in the eastern United States—including Virginia, home to Data Center Alley, and Washington, DC—to speed up its interconnection queue. Regulatory and market reform, strong contracts, and effective rate design are essential to ensure that data centers pay their fair share of electricity costs while contributing positively to grid reliability and lowering per-customer system costs.

Electricity costs will also be impacted by natural gas dynamics, especially the interplay between exports and domestic use. US exports of liquefied natural gas are expected to more than double from 2025 levels by 2029, and they will constitute more than 20 percent of all US dry gas production by 2030. Pipeline exports to Mexico are also expected to rise, albeit more gradually. Rising export volumes will lift natural gas prices, all else equal, raising fuel prices for electricity. Ultimately, shifts in natural gas demand have limited effects on the data center buildout, as even a massive 4–6 gigawatt-sized data center will only consume about 1 billion cubic feet per day of natural gas, depending on turbine efficiency, or about 1 percent of total US natural gas consumption. Nonetheless, relying solely or even primarily on natural gas to power AI data centers amplifies exposure to fuel price volatility, especially if energy-inefficient turbines like single-cycle peakers are employed.

Attempting to meet AI power needs via coal might also prove contentious. Coal is politically unpopular: only 27 percent of Americans say the United States should support coal mining, versus 65 percent and 45 percent, respectively, favoring production of renewables and nuclear energy. Tech companies might have a difficult time navigating the political demands of their clients and user base, which largely oppose coal, and the Trump administration, which favors the technology.

Finally, using coal to power AI poses technical risks. The median age of a US coal plant is forty-five years, while many have delayed maintenance. Ramping up throughput at these facilities could strain hardware and potentially degrade reliability, especially because coal plants are unable to quickly adjust output, a feature that will likely become more important in matching AI inference workloads’ unpredictable demands.

Energy infrastructure and a (potential) market correction

AI is already a powerful tool. Still, if its short-term benefits are overhyped, if AI workloads make dramatic gains in energy efficiency, or if interest rates rise due to energy shortages, a market correction or even a bubble could emerge, likely lowering projections of energy demand. Technological changes are uncertain but could upend forecasts in either direction: the reality is energy infrastructure could be overbuilt or underbuilt. In addition, some sort of AI model consolidation, such as a “winner take most” scenario, could lead to a single AI model dominating the global marketplace, with uncertain impacts for energy demand. The United States might continue to provide a large share of global data center capacity, in which case global AI adoption could increase overall US energy supply requirements. Given the inherent uncertainty surrounding AI, policymakers and industry actors must consistently adjust assumptions as developments unfold.

Data centers’ water impacts are overstated but socially salient

Larger data centers can each “drink” up to 5 million gallons of water per day, or about 1.8 billion gallons annually—the equivalent of a town with a population of about ten thousand to fifty thousand residents. While this might seem large, the United States uses about 118 billion gallons of water each day for irrigation. Put another way, the average hamburger requires more than five hundred gallons of water to produce. The average ChatGPT search requires somewhere between 0.003 and 0.007 gallons, or less than two tablespoons. In other words, the water needed to produce a single hamburger would power approximately 71,000 ChatGPT searches, conservatively.

Nevertheless, data center water consumption has emerged as a focal point of public interest, and water scarcity can be a real regional concern, especially in the developing world. In certain water-scarce areas of the country such as California or Arizona, AI might compete for limited water resources with other use cases. In Arizona, for example, agriculture and golf courses account for about 72 percent and 2 percent of the state’s daily water usage, respectively.

Finding: Too little attention is paid to less glamorous areas of the electricity system, such as transmission. The US electricity grid is aging, and the United States dramatically lags China in building out new transmission. If the United States cannot sustain its current grid, much less build the incremental generation and transmission needed for AI, then its ability to capture the global AI market will be hampered—potentially severely. Deliberate action is needed.

• Recommendation: Overhaul aging grid infrastructure and expand the power grid. Energy companies and public utilities must invest not only in replacing or retrofitting aging transmission lines, distribution lines, and supporting grid infrastructure, but also in building new generation, storage, transmission, and distribution projects. Upgrading and expanding the grid must consider both physical and digital infrastructure upgrades, including advanced transmission technologies, Internet of Things (IoT)-enabled system connectivity and monitoring, and AI-driven grid optimization and control. System efficiency improvements often generate positive returns in the long term and utilities can leverage hyperscalers’ access to capital to make these investments, thus improving reliability, mitigating cost increases, and accelerating US AI development.

Finding: An “all of the above” approach to powering AI’s electricity needs is critical, but policymakers must also recognize the reality of what is feasible over different time horizons. Speed is critical. Technologies that quickly increase the generation and transmission capacity of the grid will enable rapid development of more powerful AI models. Natural gas will remain integral to meeting growing demand while solar, onshore wind, and battery storage can be deployed cheaply and quickly.
Other fuel sources—nuclear energy, geothermal, hydro, wind, and even coal—will play an important role in maintaining reliability of the existing grid. Nonetheless, in terms of incremental generation, these fuel sources will not be constructed at the scale and speed sufficient to supply AI’s incremental power needs due to infrastructural and political constraints. Nuclear energy construction timelines often span more than a decade. Geothermal holds immense promise but has yet to reach commercialization.

Hydro is a valuable resource, but its potential has already largely been captured. Onshore wind has yet to be fully tapped; offshore wind, however, faces transmission constraints, market challenges, and political opposition. The United States hasn’t built a coal power plant since 2013 due to poor economic fundamentals, and the most recent facility now faces a nearly two-year outage through March 2027.

• Recommendation: Adopt a diversified approach to building electricity generation capacity, implementing both short- and long-term solutions. For the next five years, US near-term incremental electricity generation capacity growth will overwhelmingly consist of solar, natural gas, onshore wind, and batteries working in tandem with one another.

Finding: Non-energy costs matter: ensuring unconstrained US and allied access to AI accelerator chips holds major energy implications.

• Recommendation: Factor energy considerations into decision-making around export controls. In addition to the national security concerns of sharing advanced chips that strengthen AI capabilities of adversaries, US and allied data centers need rapid access to compute capacity. Epoch AI researchers in 2024 estimated that AI accelerator chips comprised roughly 44 percent of amortized hardware capital expenditures and energy costs. AI accelerator chips likely comprise a larger share of costs now. If domestic AI companies face higher semiconductor prices due to exports of advanced chips to adversary states, hyperscalers will be more capital constrained, thus limiting the buildout of data center capacity and the energy infrastructure that will serve it. Hyperscaler power purchasing agreements and procurement at scale hold promise in scaling new technologies, such as small modular reactors, geothermal, innovative battery chemistry, and transformers. If fewer data centers can be built domestically or in allied nations because of chip sales to adversary states, then it will be more difficult to achieve economies of scale for new energy solutions.

Finding: Maximize the utility of existing generators and grid assets by installing advanced transmission technologies, improving energy management, and investing in energy efficiency measures. Maximizing the capacity of existing infrastructure offers a cost-effective and timely complement to building new transmission lines and energy generation assets. The following actions should be prioritized to optimize grid performance.

• Recommendation: Invest in advanced transmission technologies. Dynamic line ratings, advanced conductors, power flow controllers, topology optimization, and AI- and IoT-driven optimization software can increase the throughput of existing transmission lines, reducing grid congestion and the need for new infrastructure. In addition, advanced transmission technologies are strategically advantageous to the United States as they can be deployed relatively quickly along existing corridors and are not subject to the arduous permitting and study processes that often delay new infrastructure projects.
• Recommendation: Install high-voltage direct current (HVDC) transmission lines where suitable. HVDC lines have many advantages over alternating current lines and are particularly useful for long-distance power transmission.
• Recommendation: Pursue innovation and strengthen efficiency in energy management of AI workloads. Improving AI workload efficiency, more efficient query processing, storage management, cooling systems, and other innovative tools can support data center operations, lower energy inputs, and help hyperscalers manage their systems with greater precision. Grid managers should also explore options to harness AI software for grid infrastructure. Regulators and policymakers must ensure that the cost of these software solutions can be recovered by utilities through the rate base or through targeted funding or incentive opportunities.
• Recommendation: Advance efficient use of electricity for AI and other sectoral use cases to help taper energy demand growth. No common measurement or benchmark exists for AI electricity usage. Additional inquiries by hyperscalers and independent researchers yielding a shared framework for measuring efficiency, with metrics such as queries per watt or energy consumption per inference operation, could serve as a foundation for improving AI workload efficiency that could inform government and industry infrastructure planning and development or the implementation of performance-based incentives. Energy-efficient technologies in residential, commercial, and industrial sectors reduce energy consumption and can unlock power for AI applications without compromising the quality or usefulness of energy services for those users. Demand-response programs can help customers lower their power consumption during periods of peak electricity demand. In turn, reducing consumption at peak periods both lowers users’ bills and makes additional power available in the system.
• Recommendation: Incentivize the installation of energy-efficient technologies in utility rate design, implement supportive policies that enable the building of new infrastructure, and reward the modernization of inefficient legacy systems. Investments in energy efficiency can yield long-term cost savings, improve grid reliability, and lower power system costs. Utilities can create incentive programs for energy-efficient appliances for consumers and leverage efficiencies in grid infrastructure and operations. Utilities should also conduct system planning to provide clarity on long-term system needs to which infrastructure developers can respond. Regional transmission organizations and government agencies can complement utility planning frameworks with regional and national plans, respectively. Policymakers can also jump-start efficiency opportunities with targeted policies, such as agency-directed technical assistance, grant programs for critical infrastructure projects, and incentives for both consumer products and utility assets.

Finding: Expediting permitting and engaging communities effectively would accelerate the pace of the energy infrastructure buildout. Approval processes for energy infrastructure projects can last for more than a decade in some cases, and the current patchwork of permitting authorities is creating a system that is insufficient to meet the demands of AI and consumers.

• Recommendation: Pass comprehensive permitting reform legislation. This would set enforceable deadlines for permit decisions, increase transparency on decision timelines, centralize permit coordination in a single agency, enable concurrent review of multiple project permits, and set reasonable limits on judicial review.
• Recommendation: Engage stakeholders, including the grid operator and communities, early and often in development efforts. Effective engagement reduces the likelihood of community opposition that can halt or significantly delay project development. To avoid costly litigation and project delays later in development, community benefit agreements and transparent, consistent communication with local communities should become the standard.

Finding: Engaging on electricity costs and contextualizing water consumption relative to other use cases are critical for AI development. AI’s electricity challenges are real. Fears about rising energy costs exacerbated by AI is a legitimate consumer concern, and the US AI complex should help address this problem head on by strengthening the grid, expanding transmission, easing the natural gas turbine shortage, and investing in generation sources with zero fuel cost. At the same time, water consumption issues are largely overhyped.

• Recommendation: Pursue closed-loop water recycling systems that starkly lower consumption. Treating and reusing wastewater on-site drastically reduces the need for freshwater intake. This shift also significantly reduces the volume of discharge, cutting both environmental impact and operational utility costs.
• Recommendation: Engage in more vigorous public education campaigns. Contextualize water demands relative to other use cases, including agriculture, cryptocurrency, and golf courses.

Joseph Webster is a senior fellow at the Atlantic Council’s Global Energy Center.

Frank Willey is an assistant director at the Atlantic Council’s Global Energy Center.

The GeoTech Center champions positive paths forward that societies can pursue to ensure new technologies and data empower people, prosperity, and peace.

Learn more

Image: Data centers filled with rows and rows of servers as blinking light indicate constant processing, high detail, cinematic angle composition