Data centre energy consumption: how much power AI is adding

Data center energy consumption statistics

The current statistics in one place, each covered in more detail below.

• Global use: About 485 TWh in 2025, up 17% in a year, and around 950 TWh by 2030 (IEA, 2026).
• AI: Electricity consumption in AI-focused data centres grew 50% in 2025 (IEA, 2026).
• United States: 176 TWh in 2023, 4.4% of US electricity (LBNL, 2025), and 11.8% by 2030 in the reference case (LBNL, 2026).
• Hyperscale data centres: 1,360 at the end of 2025, with 48% of worldwide data centre capacity (Synergy Research Group, 2026).

How much energy do data centres use today?

The IEA puts global data centre electricity use at around 485 TWh in 2025. It grew 17% that year, against about 3% growth in global electricity demand (IEA, April 2026). In its earlier Energy and AI report, the IEA estimated 415 TWh for 2024, about 1.5% of global electricity, and found consumption had grown around 12% a year since 2017, more than four times faster than total electricity demand.

That total is the IT equipment (servers, storage and networking) plus the facility overhead, mainly cooling and power distribution. Our glossary entry on data centre energy consumption explains the split.

Use is concentrated in a few places. The US accounted for 45% of data centre electricity in 2024, China 25% and Europe 15%. Within the US, nearly half of data centre capacity sits in five regional clusters (IEA, 2025). That is why the effect on a local grid can be much larger than the global share suggests.

Data centre energy consumption by country and region

• United States: Data centres used 176 TWh in 2023, 4.4% of US electricity, up from 58 TWh in 2014 (Lawrence Berkeley National Laboratory). The IEA expects data centres to account for nearly half of US electricity demand growth to 2030.
• China: The IEA projects an increase of about 175 TWh by 2030, a rise of around 170% on 2024 (IEA, 2025). Together, the US and China make up nearly 80% of global growth to 2030.
• European Union: The European Commission's Joint Research Centre estimated EU data centres used 45 to 65 TWh in 2022, 1.8% to 2.6% of EU electricity (JRC, 2024). The IEA expects European consumption to grow by more than 45 TWh, or around 70%, by 2030. Data centres now report energy performance under the Energy Efficiency Directive, and on 21 September 2026 the Commission proposed a rating scheme covering energy and water (European Commission).
• United Kingdom: NESO's Future Energy Scenarios 2025 put Great Britain's data centre demand at 7.5 to 7.7 TWh in 2024. Its ten-year forecast reaches about 33 TWh by 2035, and its pathways range from around 20 to 41 TWh in 2035 and 30 to 71 TWh by 2050 (NESO, 2025). NESO's newer 10 Year Outlook, published in September 2026, is lower, at 4.6 TWh in 2025 and 22 TWh in 2035, which shows how much UK forecasts are still moving (NESO, 2026).

How many data centres are there?

We could not find an authoritative count of every data centre in the world. Sites range from a server room in an office to campuses drawing tens of megawatts.

The best-tracked group is hyperscale data centres, the large sites run by cloud and internet companies. Synergy Research Group counted 1,360 at the end of 2025, with 48% of worldwide data centre capacity and almost 800 more in the known pipeline.

AI data centre energy demand: how much is AI adding?

AI is now the main driver of data centre energy growth. The IEA found electricity use in AI-focused data centres grew 50% in 2025 and expects it to triple between 2025 and 2030. In its 2025 base case, electricity used by accelerated servers (the servers built around GPUs and other AI chips) grows about 30% a year to 2030, against about 9% for conventional servers. Accelerated servers account for almost half of the net increase in global data centre consumption.

The hardware is getting denser too. According to the IEA, AI server power density rose 11 times between 2020 and 2025, and the IEA expects a further fourfold rise by 2027, when a single rack could draw as much at peak as 65 households.

Hyperscaler reports show the same curve. Google's data centres used 42.4 TWh of electricity in 2025, up from 30.6 TWh in 2024 (Google Environmental Report 2026). Microsoft used 37.0 TWh of electricity in fiscal year 2025, more than three times its fiscal 2020 figure of 10.8 TWh (Microsoft Environmental Data Fact Sheet 2026).

Energy per AI task is falling fast, by at least an order of magnitude a year in recent years according to the IEA. Use is growing faster still.

Data centre energy forecasts to 2030: why the range is so wide

Every major forecast expects data centre electricity use to keep rising. They differ on how far, because the answer depends on how many AI chips are shipped, how hard and how long they are run, how quickly efficiency improves, and whether grids and supply chains can keep up.

The LBNL 2025 update shows how sensitive the numbers are. Changing one assumption, how much power AI servers draw when idle and how heavily they are used, moves its 2030 estimate from 649 TWh to 782 TWh.

Grid limits matter as well. The IEA estimates around 20% of planned data centre projects could be delayed by grid constraints. Slow connections are pushing developers, mostly in the US, towards on-site gas, and 15 to 27 GW of it could power data centres by 2030 (IEA, 2026). Treat any single forecast as a midpoint and plan against the range.

Data centre capital spending and expansion

Global investment in data centres nearly doubled from 2022 to around half a trillion dollars in 2024 (IEA, 2025). The largest technology companies spent more than USD 400 billion on capital expenditure in 2025, and the IEA expects that to rise by another 75% in 2026 (IEA, 2026). According to the IEA, these companies are driving the current wave of data centre expansion, much of it for AI.

What drives data centre energy consumption?

A data centre's electricity goes to three places, according to the IEA:

• Servers, about 60% on average. AI accelerators draw far more power per server than conventional CPUs, which is why accelerated servers lead the growth.
• Storage and networking, about 5% each at most.
• Cooling and power distribution, from about 7% in efficient hyperscale data centres to over 30% in less efficient enterprise ones.

The overhead is measured as power usage effectiveness (PUE): total facility energy divided by the energy used by IT equipment. A PUE of 1.0 would mean no overhead at all. Google reported a fleet-wide average PUE of 1.09 for 2025 (Google, 2026). Providers publish PUE on different bases (by region, campus or wider grouping), so figures compare best within one provider.

Utilisation is the third lever. A server that sits idle still draws power, which is why assumptions about idle power and usage move forecasts so much. Cooling also uses water, which we cover in our guide to AI water usage.

Data centre efficiency: PUE, chips and cooling

Efficiency gains slow the growth in electricity consumption, but they have not stopped it. The Uptime Institute's annual survey shows the industry's average PUE falling from 2.50 in 2007 to 1.65 in 2014, then flattening. In 2025 it was 1.54, the sixth year in a row with almost no change. Sites of 20 megawatts and above averaged 1.44.

Hyperscale operators do better. Google reported a fleet-wide PUE of 1.09 for 2025, 83% less overhead energy than the industry average by its own measure. Its bigger efficiency improvements came from more efficient technologies inside the servers: more than three times as much compute per unit of energy as five years earlier, mainly from its own AI chips, called TPUs (Google, 2026).

Server racks, power density and liquid cooling

The most common rack density in Uptime's survey averaged almost 9 kW in 2025, up from 8.3 kW a year earlier, and more than 80% of operators had no server racks above 30 kW. The IEA says AI server power density rose 11 times between 2020 and 2025, and that an AI data centre has to remove heat equivalent to 30 natural gas boilers per server rack. Uptime expects liquid-cooled IT hardware and higher temperature set points to raise the share of facilities with a PUE of 1.3 or better, which was 15% of respondents in 2025.

Data centre emissions: why location matters

Data centre emissions depend on two things: how much electricity is used, and how much carbon each kWh carries. Carbon intensity varies by grid and by hour, so the same workload can have a very different footprint in a region running on wind and nuclear than in one running on gas or coal.

The IEA puts emissions from data centre electricity at around 180 million tonnes of CO2 today. In its 2025 base case they reach around 300 million tonnes by 2035, or up to 500 million tonnes in its high-growth Lift-Off case. Its April 2026 update projects emissions roughly doubling to around 350 million tonnes in 2035, about 2% of global electricity sector emissions. The IEA expects renewables to meet about half of data centre demand growth to 2035, with gas and nuclear each adding around 175 TWh.

Location-based vs market-based: the gap in hyperscaler reports

The GHG Protocol Scope 2 Guidance sets out two ways to report electricity emissions. The location-based method uses the average carbon intensity of the grid. The market-based method reflects contracts such as renewable energy certificates and power purchase agreements. Our glossary covers location-based vs market-based emissions in more detail.

For hyperscalers, the gap between the two is large:

• Google (Environmental Report 2026) reported 15.1 million tCO2e of Scope 2 emissions on a location-based basis for 2025, and 2.8 million tCO2e market-based.
• Microsoft (2026 Data Fact Sheet) reported 12.0 million tCO2e location-based for fiscal 2025, and 2.7 million tCO2e market-based. Its market-based figure was 0.26 million tCO2e the year before. In February 2025 it stopped counting short-term unbundled certificates (spot EACs) towards its renewable electricity matching.

Both figures are valid. They answer different questions. The location-based figure follows the physical grid where the electricity is used, so it is the one that moves when usage, region or timing changes.

Renewable energy and clean power procurement

Large operators cover their electricity demand with clean energy contracts, mostly long-term power purchase agreements. Google signed agreements for more than 12 GW of new clean energy in 2025, its largest annual total, and nearly 35 GW since 2010 (Google, 2026). The IEA expects renewables to meet about half of the growth in data centre demand to 2035.

Renewable energy procurement over a year is not the same as clean power every hour. Google's data centres and offices ran on about 65% carbon-free energy on an hourly basis in 2025, against 66% in 2024, while its electricity consumption grew 37%. Contracts like these lower a market-based figure. The location-based figure still follows the grid where the computing runs.

Energy security, electricity prices and the electric grid

The IEA calls data centres “a highly visible flashpoint for concerns around energy prices and the environment”. Whether they push up electricity prices depends on the local system. Where supply is tight, new load can trigger investment that raises prices. Where there is spare capacity, it can spread fixed costs and lower them. Timing is a problem too: new transmission lines take four to eight years in advanced economies (IEA, 2025).

The data centre energy challenge in Ireland

Ireland shows what happens when data centres cluster on a small grid. They used 23% of the country's metered electricity in 2025, up from 22% in 2024 and 5% in 2015, according to the Central Statistics Office. That is more than all urban homes, at 18%.

What data centre energy means for companies buying cloud and AI

If you run workloads on AWS, Azure or Google Cloud, or call AI models through an API, the electricity behind them is your provider's Scope 2. In your own inventory it sits in Scope 3, usually under Category 1, purchased goods and services, together with your share of the emissions from making the servers.

That has three practical consequences:

• Your share is not visible in provider totals: A hyperscaler's annual report covers its whole fleet. It does not tell you which part of that energy and carbon came from your workloads.
• Method changes your number: A market-based figure can fall when certificates are bought, while the grid behind your workloads stays the same. Know which basis any figure you report uses.
• Region and timing are levers: Moving flexible work to a cleaner region or hour cuts location-based emissions with no change to the code. Model choice and batching do the same for AI.

For how cloud emissions map to CSRD and the GHG Protocol, see our article on Scope 3 cloud emissions and CSRD.

How to see your share of data centre energy and carbon

Provider carbon tools report totals on their own methods and do not show the energy underneath. To manage your share of data centre energy, you need it at the same level you already manage cost: by service, account, region and hour.

Greenpixie is a data company. GPX Cloud adds energy, carbon and water data to every line of AWS, Azure and Google Cloud billing data:

• IT electricity is estimated for each billing line from the hardware behind it
• data centre overhead is added using the region's published PUE
• carbon comes from hourly, location-based grid carbon intensity for 269 cloud regions
• embodied emissions from making the servers are included, as part of Scope 3

GPX AI applies the same approach to AI usage by model and token, across cloud AI services, model provider APIs and AI tools, so AI and wider cloud usage can be reported together.

Location matters as much as size. In Greenpixie's data (November 2026), the same AWS m5.12xlarge instance running all year produces about 620 kg CO2e in Hong Kong (ap-east-1) and 284 kg in Spain (eu-south-2), 54% less. Its electricity use differs far less, 1,258 kWh against 1,067 kWh, so most of the gap comes from the local grid.

“While data centres are a visible target, in reality it is the businesses that run services within them that need to change their practices to reduce waste, improve efficiencies and ensure data centre usage grows sustainably.”

All figures are estimates, modelled from usage data, not metered. Greenpixie's cloud methodology has been independently verified in accordance with ISO 14064-3, prepared with reference to ISO 14064-1, and designed to be in line with the GHG Protocol Corporate Standard. The AI per-token layer builds on these verified factors and is validated against independent research and metered energy data.

See your own data centre energy and carbon in the Greenpixie app.

Frequently asked questions

How much energy do data centers use?

Around 485 TWh worldwide in 2025, according to the IEA, about 1.5% of global electricity on 2024 figures. The IEA expects that to roughly double to 950 TWh by 2030. In the US, data centres used 176 TWh in 2023, 4.4% of the country's electricity, according to Lawrence Berkeley National Laboratory.

How much of data center energy consumption comes from AI?

There is no single published share, because AI and conventional workloads often run in the same facilities. The direction is clear, though. Electricity use in AI-focused data centres grew 50% in 2025 and the IEA expects it to triple by 2030. Accelerated servers account for almost half of the projected net increase in data centre consumption to 2030.

Are data centers bad for the environment?

They have a real footprint: electricity that is still partly generated from fossil fuels, water for cooling and power generation, and the emissions from building the facilities and making the servers. The IEA puts emissions from data centre electricity at around 180 million tonnes of CO2 today. How large the impact is depends heavily on location, grid mix and efficiency.

Where are most data centers located?

The US accounted for 45% of global data centre electricity use in 2024, China 25% and Europe 15%, according to the IEA. Within the US, nearly half of data centre capacity sits in five regional clusters.

Do data center emissions count in my company's Scope 3?

If you use public cloud or AI services, yes. The provider's data centre electricity is its Scope 2, and your share appears in your Scope 3, usually Category 1, purchased goods and services. If you run your own data centre, its purchased electricity is your own Scope 2.

Sources

1. IEA (16 April 2026), Key Questions on Energy and AI, executive summary
2. IEA (16 April 2026), Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions
3. IEA (2025), Energy and AI, executive summary
4. IEA (2025), Energy and AI: Energy demand from AI
5. Lawrence Berkeley National Laboratory (December 2024), 2024 United States Data Center Energy Usage Report, and Berkeley Lab news (17 January 2025), Berkeley Lab report evaluates increase in electricity demand from data centers
6. Lawrence Berkeley National Laboratory (June 2026), United States Data Center Energy Usage Report: 2025 Update
7. European Commission Joint Research Centre (16 February 2024), Energy Consumption in Data Centres and Broadband Communication Networks in the EU
8. European Commission, Energy performance of data centres (accessed 3 November 2026)
9. National Energy System Operator (July 2025), FES: Pathways to Net Zero, Electricity Demand Summary data table (ED1), Future Energy Scenarios 2025
10. Google (2026), Environmental Report 2026, covering 2025
11. Microsoft (2026), 2026 Environmental Data Fact Sheet, covering fiscal year 2025
12. GHG Protocol (2015), Scope 2 Guidance
13. Uptime Institute (July 2025), Uptime Institute Global Data Center Survey 2025
14. Synergy Research Group (7 April 2026), Hyperscale Operators to Account for 67% of all Data Center Capacity by 2031
15. Central Statistics Office, Ireland (7 July 2026), Data Centres Metered Electricity Consumption 2025
16. Greenpixie, Cloud methodology and AI methodology