Energy Costs From Frontier AI Models

COLUMBUS, OHIO – JULY 24: The COL4 AI-ready data center is located on a seven-acre campus at the convergence point of long-haul fiber and regional carrier fiber networks on July 24, 2025 in Columbus, Ohio. COL4 spans 256,000 square feet with 50 MW of power across three data halls. There are concerns that new AI data centers are hiking up electricity rates due to their massive energy consumption for training and operating AI models, coupled with the need for advanced cooling systems. The Ohio Tax Credit Authority has offered sales tax exemptions and incentives for new AI data centers in Ohio for development costs. (Photo by Eli Hiller/For The Washington Post via Getty Images)

The Washington Post via Getty Images

Hyperscalers are companies that operate massive, globally distributed cloud and AI infrastructure capable of scaling computing resources elastically and automatically across millions of servers (such as the big Tech companies). Hyperscale data centers are required to meet the need for the immense computing power for training and inference of large AI models. To put these numbers in perspectives, Derek Thompson estimates the infrastructure demands of the AI boom require companies to collectively fund the equivalent of an Apollo program every 10 months. The growth of hyperscalers are creating huge, discontinuous, and strategic increases in electricity capacity and demand – outpacing traditional utility contracting structures. Tech companies are in a race to keep upgrading their semiconductor infrastructure. This has economic consequences that may not be readily apparent in current discourse about AI.

Hyperscalers’ Impact on Regional Economies

The International Energy Agency projects global data center electricity demand to more than double between 2022 and 2026, reaching over 1,000 terawatt-hours (TWh) in 2026 mainly caused by AI energy demand increasing at least 1000% between 2023 and 2026. Typically utility IRPs (integrated resource plans) assume 1–2% annual load growth while AI can create 15–35% growth in a single region over 2–3 years. A single hyperscale AI data center can consume as much electricity as a large city. To meet this skyrocketing demand, utilities must invest billions in new power generation capacity, transmission lines, and grid upgrades. Such capital costs may be passed onto utilities’ entire ratepayer base, leading to higher power bills in areas where large data centers are being built. Ordinary consumers and small businesses may be ultimately subsidizing the infrastructure buildout for tech giants.

Such increases in electricity prices are uneven in their geographical impacts given where data centers are located. It is estimated that only 15 states host about 80% of our nation’s data centers. In Northern Virginia’s “Data Center Alley” data centers already consume 25% of the region’s electricity, a total that could estimated to rise to 50% of the state’s total in some projections, leading to local utilities proposing new fossil fuel infrastructures.

The time it takes to build a new transmission line or power plant (often many years) is far longer than the time it takes to build a new data center (often 18-24 months). This mismatch creates long wait times for grid interconnection, which can delay the deployment of essential AI and other manufacturing projects. This can end up stressing regional grids and put a strain on local power capacity.

The rapid addition of massive, high-density, and variable AI loads (which can fluctuate in power use quickly) stresses the grid’s stability, leading to compromised reliability. If not managed, this can lead to lower system stability and an increased risk of power outages, which halts economic activity across all sectors.

A related question is what is the impact of such energy gentrification on other sectors of the economy. AI’s demand competes directly with the energy needs for other high-priority national goals, such as the electrification of transportation (electric vehicles) and heating, and the resurgence of domestic manufacturing. We need to acknowledge that hyperscalers have become stakeholders in shaping regional energy futures.

Policy Solutions and Frameworks to Address Hyperscalers’ Capacity Challenges

At a societal level, we may need newer style capacity markets that enable hyperscalers to procure long-term, firm, tradable commitments of grid capacity (MW, flexible load, storage, transmission rights) instead of relying on ad hoc utility contracts. Given the discontinuous demand from AI model scaling, and the challenge for utilities to build capital-intensive infrastructure on short notice as well as the differences in incentives between hyperscalers, utilities, regulators, and communities, we need newer regulatory and market structures that enable forward commitments and tradable rights. Essentially, we need regulatory and market mechanisms wherein unpredictable load surges from AI models become manageable and financially backed commitments from hyperscalars.

In a capacity market, prices should reveal whether new capacity is cheap or prohibitively expensive, where siting pressures are high, the relative scarcity of water, cooling, or peak capacity etc. Capacity markets can embed carbon intensity requirements, water-use pricing, environmental constraints, heat recovery rules and cooling efficiency standards, providing an institutional bridge between tech companies and utility companies. They could also close the chasm due to the mismatch in incentive alignment between hyperscaler growth and utility planning by using price signals and long-term contracting to align incentives, accelerate renewables, and protect households. An AI capacity market may enable hyperscalers offer “negative load” (load reductions or deferrals) as a capacity product, and allow the utility to call upon AI workloads to pause/shift during peak or contingency periods.

Utilities may also need to adapt tariff structures to distinguish large load customers from residential customers. Creating special categories of customer classes capacity reservation charges for dedicated grid upgrades and locational price signals might help.
The higher reliability needs of hyperscalers needs to be factored in as well.

Hyperscalers and Sustainability

The federal government has begun data collection efforts to assess energy usage by data centers. Reporting requirements such as mandating power usage effectiveness (PUE) ratios for AI data centers, requiring regular efficiency audits setting minimum efficiency benchmarks for AI accelerator chips might be potential solutions for utilities and governments to incorporate energy efficiency requirements into hyperscalers’ energy demands.

Regulators can mandate that developers of general-purpose AI models (GPAI), i.e., hyperscalers, measure and publicly report the energy consumption of their models during training and deployment. There are currently no binding PUE standards for private-sector data centers. However, the federal government’s data center optimization initiative (DCOI) requires existing federal data centers to achieve a PUE of 1.5 or better and new federal data centers to aim for 1.2. Some regions are considering minimum PUE standards for new data center developments. Proposed legislation, such as the Clean Cloud Act of 2025, aims to grant agencies like the EPA the authority to collect data on data center energy consumption. Clean energy mandates are another solution. Governments can incentivize hyperscalers to directly source power from renewable energy projects (such as those managed through a power purchase agreement (PPA)) and ensure that the new energy demand is met with new, clean energy generation rather than existing fossil fuels. There are also newer technical innovations that can reduce the energy demands from training AI models. With the AI arms race continuing unabated, we need new solutions to meet the energy demands from hyperscalers.