Artificial intelligence feels like software, but its expansion is increasingly constrained by steel, turbines, transformers, transmission lines, and permits. Training and serving models requires dense clusters of accelerators operating around the clock. A single facility can resemble a large industrial load, yet data centers often arrive on planning timelines much shorter than those of power plants and high-voltage lines.

The global totals are significant but the local concentration is the harder grid problem. Data centers used an estimated 415 terawatt-hours of electricity in 2024, about 1.5% of the world's consumption, according to the International Energy Agency. Its base case reaches roughly 945 TWh by 2030, just under 3% of global demand. As IEA Executive Director Fatih Birol put it, “AI is one of the biggest stories in the energy world today.”

Growth is fast and geographically uneven

The IEA projects data-center electricity use to rise about 15% annually from 2024 to 2030, more than four times the growth rate of other electricity consumption. Accelerated servers, principally associated with AI, account for almost half of the net increase. The United States and China together account for nearly 80% of projected global growth through 2030.

Those numbers do not mean data centers will overwhelm the entire global system. They mean particular substations and regional grids may face unusually large additions. Nearly half of U.S. data-center capacity is concentrated in five regional clusters, the IEA reports. A project requesting hundreds of megawatts can materially change a utility's forecast, especially if several proposed projects seek service in the same county.

U.S. estimates reveal the uncertainty. Lawrence Berkeley National Laboratory calculated that data centers consumed 176 TWh in 2023, or 4.4% of U.S. electricity, up from 58 TWh in 2014. Its range for 2028 is 325 to 580 TWh, equivalent to 6.7% to 12% of national electricity use. The width of that Department of Energy forecast reflects uncertain AI adoption, chip efficiency, utilization, and the possibility that announced projects never get built.

The timelines do not match

A data center can be operational in two or three years. New generation and transmission often take much longer. Even when developers propose ample power, grid connection is a bottleneck. At the end of 2024, roughly 10,300 U.S. projects representing 1,400 gigawatts of generation and 890 GW of storage were seeking interconnection. The Berkeley Lab queue analysis warns that a queue is not a construction pipeline—many projects withdraw—but the scale shows how crowded access to the network has become.

Large loads create a related planning problem. Utilities must reserve capacity and may need new substations or lines before knowing whether every proposal is firm. If forecasts are too low, reliability suffers and projects wait. If they are too high, customers may pay for infrastructure that is underused. Regulators must also decide who bears upgrade costs: the new data center, the utility's broader customer base, or some negotiated combination.

Co-location adds another complication. Placing a data center beside a power plant can shorten the path to electricity, but a large load behind the same interconnection point can affect resource adequacy, transmission usage, and rates. In December 2025, the Federal Energy Regulatory Commission ordered PJM—the grid region serving more than 67 million people—to create transparent service rules for co-located large loads. In 2026, FERC broadened its review to all six regional grid operators, emphasizing speed, reliability, and ratepayer protection.

The Power Grid Challenge Behind the AI Data Center Boom — supporting image
The Power Grid Challenge Behind the AI Data Center Boom — supporting image

Supply is only one part of the answer

No single resource can solve the timing problem. The IEA expects renewables to meet nearly half of additional global data-center electricity demand through 2030, with natural gas next and nuclear contributing more later. Batteries can shift energy over hours, but they do not create annual energy. Firm resources can support continuous load, but new nuclear plants and transmission face long lead times. On-site generation can help, while also raising questions about emissions and fuel infrastructure.

Efficiency and flexibility matter because they can arrive faster. Better chips, cooling, scheduling, and software reduce electricity per unit of computation, although lower costs may stimulate more use. Data centers can also schedule some training workloads for periods of abundant power, provide realistic ramping forecasts, or accept limited curtailment in exchange for quicker connections. Latency-sensitive inference is less flexible, so each facility's workload matters.

A planning problem, not an energy slogan

The grid challenge is not captured by declaring AI either unsustainable or economically indispensable. The relevant questions are specific: where is the load, how firm is the request, what generation and network upgrades serve it, who pays, and what happens if the forecast is wrong?

AI's electricity demand is growing faster than the infrastructure institutions use to meet it. Aligning those clocks requires transparent load forecasts, faster interconnection studies, cost-allocation rules, efficiency standards, and contracts that reward flexibility. Without that planning, access to power—not algorithms—may determine where the next phase of AI is built.

Sources