AI power demand became a location decision
Rising electricity demand from data centres put local grid capacity at the centre of investment decisions. Connection schedules and workload flexibility determined how much planned computing capacity could operate.
The IEA's Energy and AI report gave infrastructure investors a clearer view of the electricity demand behind computing growth. Its base case projected that global electricity consumption by data centres would reach around 945 terawatt-hours in 2030, against an estimated 415 TWh in 2024. The increase in total demand did not ensure that each proposed facility could secure a grid connection on schedule.
Founders building computing infrastructure, cooling equipment and energy software needed to assess those connections. Their addressable market depended on where facilities could operate, when power would arrive and which constraints a customer could pay to remove. A global demand forecast could guide those decisions, but each site required its own assessment.
The forecast concerned more than AI
Data centres also support storage, conventional cloud services and other digital activity. The IEA estimated that their total electricity use represented around 1.5% of global consumption in 2024. Its base case used industry projections for server shipments and assumptions about adoption, equipment efficiency and infrastructure. The projected total therefore included the whole facility load, with AI accounting for only part of it.
| Period or category | TWh per year |
|---|---|
| 2024 estimate | |
| 2030 IEA base-case projection |
IEA Energy and AI, executive summary and Chapter 2. Rounded global totals for all data centres. The 2030 figure is a scenario projection and includes services beyond AI.
Source1
A separate national study helped show the scale of uncertainty. Berkeley Lab estimated US electricity use by data centres at 176 TWh in 2023 and projected a range of 325–580 TWh for 2028. Those were US figures with a different baseline year and forecast horizon. They offered a range of outcomes for US planning and needed to be read separately from the IEA's global total. A supplier forecasting sales needed to translate either study into equipment orders within its own customer segment.
A global total concealed local pressure
The report found that nearly half of US data centre capacity was concentrated in five regional clusters. Existing clusters could offer connectivity, expertise and commercial relationships, while also competing for the same substations and transmission capacity. A country with sufficient annual generation could still have a local connection bottleneck. Energy available over a year and electrical capacity available at the required place and time answered different questions.
| Period or category | TWh per year · US data centres |
|---|---|
| 2014 estimate | |
| 2023 estimate |
Berkeley Lab's historical consumption estimates cover all US data centres, including uses beyond AI. They are not direct measurements of every facility. The figures have a different geographic scope from the IEA's global totals.
Source2
The IEA estimated that around 20% of planned data centre capacity through 2030 in its base case could face delays if grid risks were not addressed. This estimate measured exposure to unresolved grid constraints rather than the number of projects that had failed. The report noted that new transmission lines could take four to eight years in advanced economies. For a startup selling into a planned facility, the utility connection date could therefore be more consequential than a strong equipment demand forecast. Delayed energisation could postpone installation, acceptance and revenue together.
Measure efficiency against useful work
The IEA examined how improvements in software, hardware and infrastructure could meet the same demand for digital services with less electricity than in its base case. An efficiency gain means using less energy for the same amount of useful work. Falling consumption caused by fewer customers or delayed facilities would reflect weaker activity instead.
A cooling supplier could measure the electricity needed to remove heat under a specified load, while an inference supplier could measure energy per successfully completed task. Neither result alone established a customer's total bill. Utilisation, surrounding equipment and workload changes could alter the outcome. A credible pilot needed a baseline, a defined operating range and evidence that savings survived normal production conditions. The business case also had to include installation effort and maintenance, particularly when an operator could not interrupt a working facility to fit new equipment.
Useful capacity could come from different interventions
The report proposed locating new facilities where power and grid capacity were available and operating servers, generation at the facility or storage more flexibly. These options created different product markets. Software for planning sites could connect technical constraints to development choices. Controls could help an operator coordinate electrical and thermal loads. Storage and power equipment could support a defined operating requirement. Each needed to identify a constraint the customer could actually change.
Flexibility was particularly dependent on the workload. Some computing jobs could tolerate a change in timing; a service promising a rapid customer response had less room to move. The IEA also noted that curtailment could be expensive for AI facilities with high capital costs, while spare capacity might provide options. A product claiming to help the grid had to account for the value of computing work delayed or displaced, as well as demonstrate that a server could be switched off.
For investors, the route from demand to revenue depended on the customer and the constraint being addressed. A large projected power requirement could support demand for equipment and software, but the purchasing decision still sat with a named operator, developer or utility. Suppliers whose products made a particular connection usable, brought forward an operating date or reduced a verified running cost had a clearer commercial proposition. The report made location and system integration central to computing growth.
Investment implications
- The IEA's headline totals covered all data centres, with AI identified as the leading driver of growth.
- Connection constraints could delay projects even while global electricity supply and computing demand increased.
- Efficiency and flexibility products needed to demonstrate savings or usable capacity under the customer's actual workload.
Sources
- Energy and AI (April 2025) (opens in a new tab)
International Energy Agency
- Berkeley Lab Report Evaluates Increase in Electricity Demand from Data Centers (opens in a new tab)
Lawrence Berkeley National Laboratory
Published
