Science
AI Data Centres and Electricity: What the IEA Evidence Shows—and What It Does Not
IEA scenarios point to sharply higher data-centre electricity use, but the outcome depends on adoption, efficiency, infrastructure and grids—not AI alone.
Editorial note: This source-backed explainer was prepared with AI assistance, reviewed under the newsroom workflow, and approved for publication. It is not individualized professional advice or a report of a new event.
The number is large, but it is not an AI-only meter
The International Energy Agency estimates that data centres consumed about 415 terawatt-hours (TWh) of electricity in 2024, or roughly 1.5% of global electricity consumption. The IEA also says data-centre demand has grown by about 12% a year in recent years. Those figures describe the electricity used by data centres as a whole: servers, storage, networking, cooling and other facility systems. They are not a direct measurement of electricity attributable to AI alone. [1]
That distinction matters. AI workloads are an important driver of new server demand, but a facility may run many kinds of computing at the same time. A headline that converts all projected data-centre growth into “AI electricity” overstates what the evidence establishes. The IEA presents an energy-system analysis with assumptions and scenarios, not a live meter that can isolate every model-training run, inference request or conventional workload.
What the IEA’s 2030 scenario says
In its 2025 base case, the IEA projects data-centre electricity use to more than double, reaching around 945 TWh by 2030. Servers average about 60% of modern data-centre electricity use in the IEA’s equipment breakdown. Cooling is a smaller share in efficient hyperscale facilities—about 7%—but can exceed 30% in less-efficient enterprise facilities. The comparison shows why the question is not simply how many accelerator chips are installed. Power conversion, networking, storage, cooling design and utilization all shape the facility total. [1]
The IEA projects accelerated servers, largely associated with AI, to grow by about 30% per year and to account for almost half of the net increase in data-centre electricity demand in its base case. “Largely associated” is deliberately more careful than “every new unit is AI.” The estimate identifies a major contribution within a broader system whose demand also reflects non-AI services, facility construction, deployment choices and regional constraints. [1]
| IEA evidence | Careful interpretation |
|---|---|
| About 415 TWh in 2024 | Total data-centre electricity, not AI-only consumption |
| Around 945 TWh in 2030 base case | A scenario projection, not a guaranteed outcome |
| Servers around 60% on average | A facility-level average; actual mixes vary |
| Cooling about 7% to over 30% | Efficiency depends strongly on facility design and conditions |
Efficiency helps, but it does not settle the total
More efficient hardware and software can reduce the electricity needed for a particular computation. Better cooling and facility engineering can also reduce overhead. The IEA’s High Efficiency Case assumes stronger gains across hardware, software and infrastructure and produces more than 15% energy savings relative to the comparable outlook by 2035. That is evidence that efficiency can materially change the trajectory; it is not evidence that demand must fall. [1]
The reason is a familiar system-level trade-off. If computation becomes cheaper per task, organizations may run more tasks, use larger services, offer more responsive features or apply AI in settings where the cost previously discouraged use. Efficiency can therefore lower energy intensity while total electricity still rises. Whether that happens depends on demand, workload mix, utilization, location and grid conditions. The source record does not justify a universal claim that a more efficient chip, model or cooling system automatically reduces total electricity use.
The IEA’s scenario range makes the uncertainty explicit: its outlook spans roughly 700 to 1,700 TWh depending on AI adoption, efficiency progress and energy-system bottlenecks. This is not a prediction interval for one inevitable future. It is a way to show how different assumptions about technology and infrastructure produce very different system requirements. [1]
Why local grids and communities may notice first
Global percentages can conceal local pressure. A data centre’s electricity demand arrives at a particular connection point, in a particular region, with a particular generation mix and transmission capacity. A fast-growing cluster may therefore raise questions about connection queues, reliability, land and water use, wholesale prices or the timing of new generation, even while the global share remains modest. The IEA’s emphasis on grid and infrastructure bottlenecks is a reminder that electricity availability is not only a global accounting problem. [1]
The practical implication is not that AI should be treated as uniformly harmful or that efficiency should be dismissed. It is that decisions should compare the full system: computing demand, cooling, power supply, siting, utilization and the emissions characteristics of the electricity serving the facility. A more efficient server in a poorly constrained location is not the same intervention as a more efficient server paired with flexible workloads and additional low-emissions supply.
What the U.S. evidence adds—and its limits
The 2024 Lawrence Berkeley National Laboratory report offers a separate U.S.-focused historical estimate and scenario range through 2028. It is based on prior studies and historical server-shipment data rather than a single direct measurement system. That makes it useful context for understanding the U.S. trajectory, but it should not be blended with the IEA’s global scenario as if the two were identical measurements. [2]
Taken together, the sources support a measured conclusion: data-centre electricity demand is becoming a significant energy-system issue, and AI-related servers are an important contributor to the expected increase. They do not establish a precise global tally for AI alone, a fixed 2030 result or a guaranteed effect from any one efficiency improvement. The responsible reading is conditional: demand, efficiency and infrastructure will interact, and the observed outcome will depend on choices made across all three.
Editorial disclosure: This is an explanatory, source-backed article, not individualized professional advice and not a report of a new event.
Frequently asked questions
1. Does 415 TWh mean AI used 415 TWh in 2024?
No. It is the IEA’s estimate for electricity used by data centres overall. AI is a major driver of new demand, but the figure also includes non-AI computing and the energy used by supporting systems such as cooling and networking.
2. If chips become more efficient, will total electricity use fall?
Not necessarily. Efficiency can reduce energy per computation, while lower costs and wider adoption increase the number or size of computations. The total result depends on demand, workload mix, facility design and grid conditions.
3. Is 945 TWh a firm forecast?
No. It is the IEA’s 2030 base-case projection. The IEA also presents a much wider scenario range, reflecting uncertainty around adoption, efficiency and infrastructure bottlenecks.