The artificial intelligence race is expanding from chips and data centers into the power system. AI model training and inference require large numbers of servers to operate around the clock, along with cooling, storage, networking, and backup equipment. As a result, the electricity consumption of a single large-scale campus is gradually approaching that of a major industrial facility. The U.S. Energy Information Administration projects that total U.S. electricity consumption will increase from 4.195 trillion kilowatt-hours, or 4,195 TWh, in 2025 to 4,269 TWh in 2026, before rising further to 4,399 TWh in 2027. Electricity demand growth is expected to be more pronounced in the commercial and industrial sectors, with data center expansion serving as an important driver. Lawrence Berkeley National Laboratory estimates that data centers could account for approximately 11.8% of total U.S. electricity consumption by 2030, with different scenarios ranging from 9.5% to 15.3%.
As electricity demand increases rapidly, market attention is no longer focused solely on whether sufficient power is available. The question of who should pay for new power plants, transmission lines, and substations has also become a policy dispute. If utilities incorporate these investments directly into general electricity rates, households, small and medium-sized businesses, and traditional manufacturers may end up sharing the cost of technology companies’ AI infrastructure expansion.
Data Center-Intensive Regions Are Already Facing Higher Electricity Costs
Data centers can affect electricity prices through three main channels: generation, reserve capacity, and transmission and distribution investment. When new demand exceeds existing supply, grid operators must activate more expensive generating units and pay higher fees to ensure sufficient capacity remains available during peak periods. If transmission systems cannot accommodate the electricity demand of large campuses, utilities must also build substations, transmission lines, and interconnection facilities.
These pressures are particularly evident in PJM, the largest regional power grid in the United States. Due to rising data center demand, power plant retirements, and the slow pace of new supply additions, PJM capacity prices have increased more than tenfold over two years. Ohio-based Belden Brick saw its monthly capacity charges rise from approximately US$1,600 to US$12,000, while its total electricity costs increased by as much as 90%. As of December 2025, industrial electricity prices in Pennsylvania and Ohio had risen by 31% and 26% year over year, respectively, compared with a national average increase of 7%.
Data centers are not the only cause of rising electricity prices. Natural gas prices, extreme weather, transmission congestion, power plant retirements, and aging equipment can also affect electricity bills. The price impact of AI-related electricity demand also varies significantly by region. Areas with high data center density, limited supply margins, and lagging grid construction generally face greater pressure.
The White House Is Encouraging Technology Companies to Pay for New Power Infrastructure
In March 2026, the White House introduced the Ratepayer Protection Pledge, which was signed by Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI. The participating companies pledged to build, procure, or bring online the new generation resources needed to meet data center demand and to pay for the grid upgrades required to serve those facilities. They must also negotiate dedicated rate structures with utilities and state governments. Even if actual electricity use falls below initial projections, the companies would still be responsible for the power and infrastructure built on their behalf, reducing the risk that the cost of unused capacity is shifted to existing customers.
The pledge remains voluntary and has not directly established a nationwide, legally enforceable cost-allocation framework. Implementation still depends on state regulation, utility rate structures, and corporate contracts. Reuters reported on July 13 that the White House was planning to invite utilities, data center developers, and state governments to participate in a follow-up initiative.
The U.S. Federal Energy Regulatory Commission also directed six regional grid operators on June 18 to explain within 60 days whether their existing rate structures were just and reasonable or to propose revisions. The review covers accelerating interconnection for large loads, preventing transmission construction costs from being shifted to other customers, establishing flexible load services, and addressing rules for data centers co-located with dedicated power generation.
Direct Corporate Cost Assumption Cannot Eliminate Broader Supply-Demand Pressure
Requiring technology companies to pay for dedicated interconnection facilities and new generation can reduce the risk that ordinary customers directly subsidize data centers, but it cannot fully isolate the power market from broader price changes. When multiple companies simultaneously purchase transformers, gas turbines, switchgear, and engineering services, they may still drive up equipment and construction costs, affecting the expansion plans of other utilities.
Supply-chain constraints have become a major limitation on data center development. In the first quarter of 2026, lead times for generator step-up transformers in the United States exceeded 160 weeks. Some equipment must be ordered three to five years in advance, while transformer prices are expected to rise by approximately 4% to 10% over the next year. Because large transformers are generally customized, expanding factory capacity, completing engineering certification, and training specialized workers all require significant time. Even when data center buildings and servers are ready, operations may still be delayed because essential power equipment has not yet been delivered.
Technology companies’ capital expenditures therefore no longer cover only GPUs, servers, and cooling systems. They must also secure generation capacity, transformers, transmission and distribution equipment, and engineering contractors well in advance. The price and delivery schedule of power infrastructure are beginning to determine when AI computing resources can enter service.
Behind-the-Meter Generation and Long-Term Energy Contracts Are Emerging as Alternatives
As public grid expansion fails to keep pace with data center development, some developers are adopting behind-the-meter generation. Under this model, natural gas generators, fuel cells, renewable energy systems, or energy storage facilities are installed behind the customer’s meter to supply power directly to the campus. Between 2024 and 2025, announced behind-the-meter generation projects for data centers in Texas exceeded 20 GW, while planned capacity nationwide reached approximately 56 GW over the same period.
On-site generation can shorten the wait for grid interconnection, but it must still address fuel supply, equipment redundancy, emissions, and local environmental concerns. Most large campuses also cannot fully disconnect from the public grid and still depend on it for backup power, frequency regulation, and emergency dispatch. Whether data centers pay rates sufficient to cover these system services therefore remains an important element of cost allocation.
Technology giants are also expanding their use of long-term energy contracts. In January 2026, Meta signed agreements with Vistra, TerraPower, and Oklo to support up to 6.6 GW of existing and new nuclear capacity by 2035. Long-term corporate contracts can provide revenue support for nuclear plant life extensions and new reactor technologies, but new nuclear projects still face uncertainty over approvals, construction timelines, and costs. They therefore cannot independently fill all data center power needs in the near term.
AI Competition Is Beginning to Encompass Energy Access and Grid Management Capabilities
Expanding generation and transmission remains central to meeting long-term electricity demand, but smart grids, energy storage, and demand flexibility can also reduce peak-load pressure. Dynamic line ratings, power flow control, and topology optimization can improve utilization of existing transmission assets, while virtual power plants and on-site storage can adjust loads when the grid is under stress, reducing the need to build large amounts of reserve capacity for a limited number of peak periods. Some AI training workloads can also be shifted to off-peak hours or transferred to data centers in other regions. Real-time inference, cloud services, and critical systems, however, still require highly reliable, round-the-clock power. Future data center competitiveness will therefore depend simultaneously on computing efficiency, energy access, grid interconnection speed, and load management capabilities.
Technology giants’ commitment to paying for new generation and grid upgrades can reduce the risk that AI construction costs are directly shifted to households and small and medium-sized businesses. However, equipment shortages and tight overall supply-demand conditions may still affect regional electricity prices. The ultimate effectiveness of these policies will depend on whether new supply can come online on schedule, how local rates are designed, and whether corporate commitments can be converted into specific and enforceable contracts. As the constraints on AI infrastructure expand from chip supply to the power system, grids, transformers, energy storage, and energy management equipment will also become critical infrastructure shaping the pace of the next phase of computing expansion.