Global electricity demand from data centres rose by 17% in 2025, compared with 3% growth in electricity use overall, according to the International Energy Agency. Yet the main obstacle for many proposed AI facilities is not the world’s total supply of power. It is obtaining enough electricity at the right location.
Artificial intelligence may be delivered through software, but it runs on servers, cooling equipment, transformers, substations and transmission lines.
Technology companies can order computing equipment and revise expansion plans within months. The electricity networks serving those machines often take years to plan, approve and build.
That difference in speed is turning grid access into a commercial constraint on AI investment.
AI demand is growing faster than grid infrastructure
The IEA’s 2026 update on energy and AI projects that data-centre electricity consumption will rise from about 485 terawatt-hours in 2025 to 950 terawatt-hours in 2030. That would be close to 3% of global electricity demand.
A terawatt-hour, or TWh, is one billion kilowatt-hours. It measures electricity consumed over time, rather than the maximum power a facility may draw at a particular moment.
AI-focused facilities are expected to expand faster than the wider data-centre market. The IEA expects data-centre consumption to roughly double by 2030, while electricity use at sites focused on AI could triple.
There is an important counterweight. The agency says the electricity required for each AI task is falling rapidly as chips, models and software become more efficient. However, more people are using AI and more demanding applications, including AI agents, are being introduced. The increase in use is expected to outweigh the savings per task.
The United States shows how large the range of possible outcomes remains. A June 2026 report from Lawrence Berkeley National Laboratory estimated that data centres could account for 11.8% of US electricity consumption in 2030 in its reference case. Its sensitivity scenarios ranged from 9.5% to 15.3%.
These are projections, not settled demand. They depend on equipment sales, chip efficiency, cooling systems, server use and the number of projects that are actually built.
Why does a grid connection take so long?
A large AI data centre is an industrial-scale electricity user. It cannot be connected in the same way as an office or warehouse.
Before approving the connection, a utility or grid operator must study whether the local network can supply the proposed load without weakening reliability. The answer may require a new substation, larger transformers, upgraded local lines, additional generating capacity or new high-voltage transmission.
Transmission carries large amounts of electricity over long distances. Distribution networks then deliver it to individual customers. A region may have enough generating capacity in total but still lack the wires and substations needed to move that power to a particular data-centre site.
The latest IEA analysis of electricity grids estimates that a data centre can take one to three years to build, while planning, permitting and completing new grid infrastructure can take five to 15 years. Prices for important grid components have nearly doubled over the past five years.
Location also matters. The IEA previously found that half of US data-centre capacity under development was in existing large clusters. Clustering brings fibre connections, skilled contractors and established suppliers, but it can concentrate new electricity demand where the grid is already busy.
The 2,500 GW queue is not a data-centre total
The IEA says more than 2,500 gigawatts of projects are stalled in grid-connection queues worldwide. This figure is sometimes presented as evidence of data-centre demand alone, but it includes renewable power plants, electricity storage and large users such as data centres.
A gigawatt, or GW, measures power capacity at a given time. It is different from a terawatt-hour, which measures energy used over a period. The two figures cannot be compared directly.
Connection queues are also more complicated than an ordinary waiting list. Each project must be studied to determine how it would affect the network and who should pay for any upgrades. Applications can be withdrawn, revised or moved to another location, forcing planners to repeat part of the work.
Some developers apply at several sites while searching for the fastest and least expensive connection. In remarks in June, US Federal Energy Regulatory Commission member David Rosner said speculative applications could clog queues, duplicate demand forecasts and lead to infrastructure being planned for projects that never appear.
On 18 June, FERC ordered the six regional grid operators under its jurisdiction to justify their existing rules or propose reforms for connecting data centres and other large users. They were given 60 days to respond.
The orders focused on readiness requirements, flexible connections and agreements intended to keep existing households and businesses from paying for infrastructure built for a large project that does not materialise. A tariff in this setting is the rulebook and pricing structure for transmission service, not a tax on imports.
Faster connections need construction and flexibility
More power generation will be required if demand follows the central forecasts, but new power plants will not solve a local transmission shortage by themselves.
The IEA estimates that annual grid investment would need to rise by roughly 50% from today’s level of about $400 billion by 2030 to meet forecast electricity demand. Building new lines, substations and equipment remains the durable solution in areas where demand has outgrown the network.
There may also be more room in existing grids than static limits suggest. Dynamic line rating uses real-time weather and equipment data to calculate how much electricity a transmission line can safely carry. For example, cooler or windier conditions can allow a line to carry more current without overheating.
The IEA estimates that grid technologies and regulatory reforms could together release enough capacity to connect between 1,200 GW and 1,600 GW of advanced-stage projects now in queues. Of that total, it estimates 750 GW to 900 GW could be enabled through non-firm connection agreements.
A non-firm agreement gives a project earlier access to the grid on the condition that its consumption or output can be limited at certain times. The remaining 450 GW to 700 GW could come from technologies and upgrades such as dynamic line rating, power-flow controls and replacing existing conductors.
Those are modelled global estimates, not guaranteed capacity at every site. Voltage, substation limits, local demand and other engineering constraints still require project-specific studies.
Data centres could contribute by shifting non-urgent computing to less congested hours or regions, using batteries and making spare server capacity more flexible. The financial incentive is mixed. AI facilities contain extremely expensive equipment, so owners generally want their processors working rather than waiting for electricity.
Electricity availability could redraw the data-centre map
The US Department of Energy reached a similar conclusion in its draft 2026 National Transmission Needs Study, released on 9 July. It identified additional transmission needs associated with growing demand from data centres, manufacturing and other large industrial users. The study remains a draft under consultation, rather than a final investment plan.
For regions trying to attract data centres, access to electricity may now matter more than a generous tax break or inexpensive land. A site with excellent fibre links but a five-year wait for power can lose to a less obvious location with spare grid capacity.
The same calculation applies to developers. Existing clusters still offer commercial advantages, but continued concentration can make each additional connection slower and more expensive.
AI companies may move at software speed. Their physical infrastructure does not. In some markets, transformers, transmission capacity and a place in the connection queue may determine how quickly new computing capacity can come online.
