The internet is often described as a cloud, which makes it sound weightless. In physical terms it is a very large collection of machines that run on electricity: warehouse-scale computers, radio towers, undersea cables, and roughly ten billion screens and phones drawing power wherever their owners happen to sit. Global electricity demand is around 30,000 terawatt-hours a year. Estimates of what the internet and its supporting infrastructure consume sit somewhere between 3 and 5 percent of that, a share that grows a little every year and a lot when a new class of workload arrives.
Breaking the total into three parts makes the picture clearer.
Data centres
A data centre is a building full of computers whose electricity use has two components: the servers, storage, and networking equipment inside, and the cooling and power systems that keep the machines alive. Industry analysts at the International Energy Agency put global data centre consumption at roughly 415 terawatt-hours in 2024, about 1.4 percent of world electricity and comparable to the entire consumption of France. Wikipedia’s account of data center energy use collects the published estimates and explains why they differ by a factor of two or more depending on what gets counted.
The numbers at the scale of a single building are worth pausing on. A large cloud campus now draws 300 megawatts or more. At 300 megawatts and an 80 percent average load, one campus consumes about 2.1 terawatt-hours a year, which is roughly the electricity demand of a city of 500,000 people. Efficiency has improved at the same time. The ratio of total facility power to computing power, a measure the industry calls PUE, averaged around 1.55 a decade ago and now sits near 1.2 at well-run sites. At the very best facilities the ratio approaches 1.1, which means almost every purchased watt reaches a chip.
Networks carry the traffic
The second component is the transport layer: fixed broadband, mobile towers, core routing, and the long-haul links between cities and continents. Published estimates for network electricity land between 250 and 350 terawatt-hours a year, with mobile networks accounting for the larger share. Radio equipment is sensitive to load and distance. A rural cell site that serves fifty customers has roughly the same power budget as an urban site serving thousands, which is why coverage obligations are expensive in energy terms.
Per bit of traffic, network efficiency has improved far faster than traffic has grown. Streaming video is a good case study. An hour of high-definition video, delivered over a network and decoded on a television, was estimated at around 0.15 kilowatt-hours in 2015. The same hour on the same connection today runs closer to 0.08, largely because encoding standards halved the bandwidth and screens became more efficient.
The devices at the end
The third component is the one readers touch. A smartphone costs 3 to 5 kilowatt-hours a year to charge. A modern laptop runs 30 to 60. A 55-inch television watching four hours a day lands near 120 kilowatt-hours a year, and a gaming console playing three hours of demanding games adds another 150 or so. Household routers, printers on standby, chargers left in walls: the small stuff sums to something. A typical European home with a dozen connected devices spends 200 to 300 kilowatt-hours a year keeping them running.
A rough total
Add the three columns and the internet arrives at something like 1,200 to 1,500 terawatt-hours annually, with data centres growing fastest. Cryptocurrency mining sits outside most of these figures and adds roughly 100 terawatt-hours of its own, dominated by a single proof-of-work network. The growth story going forward is concentrated in one place. Training and serving large AI models happens in the densest, newest data centres, and the IEA expects global data centre demand to approach 945 terawatt-hours by 2030 in its central scenario, more than double the 2024 figure.
Where the trend line leads
The encouraging part of the history is that computing demand has grown by orders of magnitude while electricity use has grown much more slowly. Data centres used roughly the same electricity in 2020 as in 2015, despite processing several times more data. Chips got more efficient, virtualization packed servers more tightly, and the worst facilities were retired. That decoupling is now being tested by workloads that are unusually compute-intensive, and whether it holds is an engineering question as much as an environmental one.
None of this requires anyone to stop streaming video. It does suggest that the electricity behind a download is real, measurable, and worth knowing about.