Physical systems now set the pace of AI

The next constraint on artificial intelligence is the pace of the physical system around it.

The International Energy Agency's April 2026 review measured data-center electricity use at about 485 terawatt-hours in 2025, and that count includes ordinary internet services as well as artificial intelligence.

Software can be changed in hours. A power plant, a chip factory, or a high-voltage line takes years. The Agency's review says the capital expenditure of five large technology companies exceeded 400 billion dollars in 2025 and is expected to rise another 75 percent in 2026. The Agency presents that spending as the surge in data-center investment. It is company capital expenditure, not a separate count of AI chips, and the review says not every announced project will be built. Satellite tracking in the same review shows that data centers built for AI more than tripled in capacity over the prior 18 months.

Ordering more accelerators does not put them to work if the power connection is missing. One reading of this record is that usable scale is the throughput of the whole system, chips and memory and power and capital arriving together, rather than the size of any one order.

Energy is a small share of the world and a large share of some places. Data-center electricity use grew 17 percent in 2025. Use at AI-focused centers grew about 50 percent that year. The Agency's central projection puts all data centers at about 950 terawatt-hours in 2030, roughly double the 2025 level and about 3 percent of global electricity demand. AI-focused centers triple inside that projection. The 485 figure is a measurement for 2025. The 950 figure is a projection for 2030. A global share near 3 percent is not a world power crisis. A load that must be served at one site and one hour is a different problem.

A site drawing 100 megawatts without a break would use about 0.876 terawatt-hours in a year of 8,760 hours. That product is arithmetic, not an Agency statistic. Several such sites in one region can change the local grid while the global share stays small.

The wires can bind before new plants do. The Agency's Electricity 2026 report says more than 2,500 gigawatts of projects sit in connection queues. The queue covers renewables, storage, and large loads, including data centers. It is not an AI-only total. Meeting electricity demand through 2030 would require annual grid investment to rise by about 50 percent from about 400 billion dollars. Planning and building grid infrastructure can take 5 to 15 years. A data center can take 1 to 3 years. The same analysis says grid-enhancing technology and more flexible connection rules could free capacity for about 1,200 to 1,600 gigawatts of advanced-stage projects. About 750 to 900 of those gigawatts would come from non-firm connection agreements, and about 450 to 700 from tools such as dynamic line rating and from upgrades such as reconductoring. That potential covers many kinds of projects, not data centers alone. Prices for key grid components have nearly doubled over the prior five years.

Inside the machines, memory and the movement of data can bind before the count of chips does. The April review says a shortage of high-bandwidth memory developed over the prior six months and is expected to last at least through the end of 2027. Citing IDC, it says current production of that memory can support a maximum of around 25 gigawatts of AI-ready servers a year through 2027, in line with the Agency's base case. The power density of AI servers rose 11 times from 2020 to 2025, and the review says it is set to rise about fourfold further by 2027. A rack in an advanced center, about the size of a large refrigerator, could by 2027 have a peak power demand equal to about 65 households. Adding graphics processors does not raise useful output in proportion if memory, packaging, or the links between servers are the slow part.

Water and copper are real constraints and easy to overstate. Direct water use depends on climate and on the cooling design. A November 2025 paper in Nature Sustainability models an annual water footprint for AI servers in the United States of 731 million to 1,125 million cubic meters, depending on the expansion path across 2024 to 2030. In that paper's breakdown, about 71 percent of the footprint is indirect water from generating electricity. It is a model, not a meter reading. S&P Global's 2026 copper study forecasts copper tied to data centers, including a share of the grid assigned to that load, rising from 1.1 million metric tons in 2025 to 2.5 million in 2040. A scenario range in the same study runs from 1.7 million to 2.7 million tons in 2040. The study also projects a possible shortfall of 10 million metric tons in global copper supply in 2040 if supply does not expand. That shortfall is a whole-economy scenario, not a result caused by AI alone.

A cheaper task can raise total use. The Stanford Institute for Human-Centered Artificial Intelligence, in its 2025 AI Index, reports that the price of querying a model scoring the equivalent of GPT-3.5 on the MMLU benchmark fell from 20 dollars to 0.07 dollars per million tokens between November 2022 and October 2024, a drop of more than 280 times. That is a price, not a measurement of electricity. The Index also says hardware energy efficiency has improved by 40 percent a year. The IEA says energy use per AI task has fallen by at least an order of magnitude a year, and that a simple text query now typically uses less electricity than a television over the same span. Video generation, reasoning, and agent tasks can use hundreds or thousands of times more energy per query than simple text. Demand is the product of efficiency, of how many people use the tools, and of which tasks they run.

Capital can slow the build before physics does. The IEA puts the capital cost of a cutting-edge AI data center, including the information-technology equipment, at 40,000 to 50,000 dollars per kilowatt, five to ten times a nuclear plant on that per-kilowatt comparison. Its growth estimates imply about 3.9 trillion dollars of data-center investment from 2026 through 2030, more than company balance sheets can fund alone. In the second half of 2025, the Agency says investor sentiment moved from broad optimism toward caution about returns. A project sized for a demand path that does not arrive leaves idle equipment and a power contract. A project that is not built leaves real demand unmet.

The map of compute is concentrated, and the work does not all need the same building. The World Bank says high-income countries held 77 percent of global co-location data-center capacity in June 2025, measured in megawatts. That is co-location capacity, not the whole of AI infrastructure. Training a frontier model wants a dense cluster and fast links. Much inference can run in smaller sites, and some of it on the device. The IEA says three-dimensional chip stacking can cut the energy spent moving data off the chip, at the cost of more heat and harder manufacturing. Co-packaged optics, which replace some electrical signals with optical links, are entering initial production for switches between racks, with chip-to-chip use expected from around 2027 to 2028. The Agency says the effect of that step on the efficiency of the whole fleet is still uncertain. Putting models into vehicles, factories, and other devices can move some electricity use out of data centers, and the total of many small devices can still be large. The magnitude is uncertain.

The supplies that can arrive inside this decade are narrower than the research horizon. Data-center companies signed renewable power contracts covering more than 40 terawatt-hours in 2025. The Agency also records contracts for existing nuclear plants, including a Microsoft agreement for about 835 megawatts from a restart of Three Mile Island Unit 1, targeted for 2027, and an Amazon agreement for up to 1,920 megawatts from the Susquehanna plant. Announced offtake for small modular reactors grew from about 25 gigawatts at the end of 2024 to 45 gigawatts by the end of 2025, and the first of those projects is not expected before about 2030. Many of the larger next-generation nuclear arrangements are not yet backed by firm financing. Nuclear fusion is not a data-center supply in that account. A later plant, or a later optical network, does not connect a project that is in the queue now.

Three paths for 2030 start from the measured 485 terawatt-hours of data-center electricity in 2025. They describe all data centers, not AI alone. An 8 percent annual rise reaches about 713 terawatt-hours. A 20 percent annual rise reaches about 1,207. Those two paths are arithmetic on the 2025 measurement. They are not Agency forecasts, and they carry no probabilities. The Agency's central projection is about 950 terawatt-hours. The Agency says bottlenecks make the more aggressive near-term paths less likely, and that a higher path remains possible after 2030 if equipment bottlenecks ease and energy-heavy uses spread. Its chart, updated 16 April 2026, separates the historical series from the base case.

What would weaken this reading is a joint easing, in the same years, of power connections, high-bandwidth memory, and the cost of capital, while the mix of tasks stays as light as simple text. These sources do not show that break. The sector can add a great deal of compute and still be limited by whichever part of the system arrives last.