AI feels weightless when it arrives as text in a browser. The machinery behind it is anything but.

Training and running large models happens in buildings filled with accelerators, networking equipment, cooling systems and electrical hardware. These data centers are often discussed through one enormous global number: how much electricity they use in a year.

That number matters. It can also hide the problem a power engineer actually faces.

Electricity is not a single pool that can be drawn from equally anywhere. A country may have enough generation in total while one county lacks the substation, transformers and transmission capacity to connect another large campus. The AI data-center problem is therefore not only how much power exists. It is where that power must arrive, at what hour and through which wires.

A small global share can be a very large local load

The International Energy Agency’s 2025 Energy and AI report estimated that data centers consumed about 415 terawatt-hours of electricity in 2024. That was roughly 1.5 percent of global electricity use.

Seen only as a percentage, 1.5 percent can sound manageable. The demand is not evenly spread across the globe.

The United States accounted for 45 percent of data-center electricity use, China for 25 percent and Europe for 15 percent. Within the United States, nearly half of existing capacity was concentrated in five regional clusters. The IEA also found that half of the data centers under development there were being planned in existing large clusters.

A typical AI-focused data center, in the agency’s comparison, can use as much electricity as 100,000 households. The largest facilities under construction were expected to use about 20 times that amount. These are comparisons of electricity consumption, not claims that a data center is socially or economically equivalent to a city.

What they reveal is scale at the connection point. Adding one very large load to one local grid is different from distributing the same annual electricity use across millions of homes and thousands of substations.

The range of plausible demand is unusually wide

The IEA projected global data-center electricity consumption to reach about 945 terawatt-hours in 2030, slightly more than Japan uses today. AI is the largest driver of the increase, alongside continued growth in other digital services.

That is a modeled outlook, not a meter reading from the future. The agency’s scenarios for 2035 ranged from 700 to 1,700 terawatt-hours depending on AI uptake, hardware and model efficiency, and the speed at which energy bottlenecks are resolved.

A separate 2024 report from Lawrence Berkeley National Laboratory illustrates the uncertainty at the national level. It estimated that US data centers used 176 terawatt-hours in 2023, or 4.4 percent of the country’s electricity. For 2028, the projected range was 325 to 580 terawatt-hours, equal to 6.7 to 12 percent of US electricity use.

The high end is almost twice the low end.

I think that gap should make us cautious in both directions. It is too large to treat the highest forecast as destiny. It is also too large for grid planners to assume that recent efficiency gains will quietly absorb the growth.

The computers and the grid run on different clocks

The IEA estimated that around 20 percent of planned data-center projects could face delays unless connection risks are addressed. In advanced economies, a new transmission line can take four to eight years. Wait times for transformers and cables have doubled in three years.

This is where the conversation becomes less dramatic and more concrete. A developer can order servers and construct buildings. Supplying them may require a new substation, upgraded transformers, additional generation, transmission studies, permits and equipment with its own manufacturing queue.

The national electricity total does not solve any of those steps.

The same distinction appeared when I recently wrote about solar supplying more than 8 percent of global electricity. The next constraint was not whether panels could generate power. It was whether grids and storage could move a midday surplus to the place and hour it was needed.

AI data centers expose the other side of that problem. A facility may have a contract for enough renewable electricity over a year while still drawing from a constrained local grid during a particular hour. Annual matching and physical delivery answer different questions.

Some computing can move in time, but not all of it

One proposed response is flexibility. Some computing jobs can wait for an hour with more available power, or move to another data center on a less constrained grid. Onsite batteries and backup generation can also reduce the facility’s draw during short periods of grid stress.

This is not merely theoretical. In a paper in IEEE Transactions on Power Systems, Google researchers described a production system that forecast electricity carbon intensity and delayed flexible computing work to lower-carbon hours. Hourly capacity limits shifted eligible jobs while preserving the total daily computing capacity assigned to them.

The system demonstrates that some large-scale computing can be scheduled around grid conditions. It was developed and evaluated within Google’s own infrastructure, and it did not make every workload flexible. Search, live model responses and other latency-sensitive services cannot simply pause whenever electricity is scarce. Long training runs can also be costly to interrupt, especially when thousands of accelerators must remain synchronized.

The IEA notes another awkward tradeoff: an AI-focused data center can be ten times more capital-intensive than an aluminum smelter. Owners have invested heavily in keeping expensive chips busy, so asking them to reduce output may carry a high opportunity cost.

Flexibility needs a contract, not just clever software

When I looked at electric cars sending power back to the grid, the technical battery capacity was only the beginning. Chargers, standards, owner consent and compensation determined whether that capacity was actually available.

Data centers face a similar divide between potential and usable flexibility.

A grid operator needs to know how much demand can be reduced, how quickly, for how long and with what notice. A data-center operator needs payment or faster connection in return, plus limits that protect customer service and equipment. Without those rules, spare server capacity is not a dependable grid resource.

Siting may be even more important. Placing new campuses where transmission and generation are available can avoid some bottlenecks before they exist. Yet the reasons companies cluster are real: fiber connections, skilled workers, land, tax policy and proximity to existing infrastructure.

The honest answer is local

It is tempting to ask whether AI will “use too much electricity” and look for one global percentage in reply. I do not think the evidence supports such a neat verdict.

Data-center demand is growing quickly, but forecasts remain wide. Efficiency will improve, but cheaper and more capable AI may also increase use. New generation can be built, but a power plant hundreds of kilometers away does not help until the network can deliver its output.

The better questions are less universal: Which grid will serve the facility? What must be upgraded? Who pays? How much of the workload can move? What happens during the system’s hardest hour?

AI arrives on a screen in seconds. The infrastructure beneath it still moves at the speed of substations, transformers and public decisions.