本栏目尚未翻译。以下内容按原文以英文显示。 阅读英文版本.
Resources: power and water · updated 3 Nov 2026
Data centre power and water: what new capacity needs and how cooling changes it
How much extra electricity the coming wave of data centres could need, how electricity and water use differ by type of data centre, and how the way heat is removed changes both. Forecasts are labelled as forecasts with their scenario ranges, and every figure links to its source.
Vendor-neutral. Figures come from agencies, grid operators, surveys, operators' own reports and public planning documents. No product or brand is ranked.
Confirmed
415 TWh
global data centre electricity use in 2024, about 1.5% of world use (IEA estimate)
Source: IEA (opens in a new tab)
Projected
945 TWh
forecast for 2030 in the IEA Base Case, just under 3% of world use
Forecast. Source: IEA (opens in a new tab)
Projected
580 TWh
top of LBNL's 2028 range for the US (325 to 580 TWh, up to 12% of US use)
Forecast. Source: LBNL (opens in a new tab)
Confirmed
22%
share of Ireland's metered electricity used by data centres in 2024
Source: CSO (opens in a new tab)
Key takeaways
- The IEA expects global data centre electricity use to roughly double from about 415 TWh in 2024 to about 945 TWh in 2030 (Base Case forecast). Its 2035 cases run from about 700 to over 1,700 TWh.
- The US and China account for nearly 80% of the forecast growth to 2030. In the US, LBNL forecasts 325 to 580 TWh by 2028 and EPRI 196 to 404 TWh by 2030.
- Every MW of IT load running all year uses 8,760 MWh before overhead. PUE adds from under 10% (best hyperscale fleets) to over 50% (average and older sites).
- Cooling choices trade electricity against water. Evaporative and cooling tower systems often have a lower PUE but can use litres of water per kWh. Dry and air-cooled systems use almost no water but more electricity.
- Plant noise scales with fan speed, number of units and distance. In the UK, BS 4142 compares the plant noise with the background level at the nearest homes.
In plain English: how 1 GW of IT becomes X GW of grid supply and Y GW of heat
PUE is the site's total electricity divided by the electricity used by the IT equipment. So grid supply = IT load × PUE. Almost all of that electricity ends up as heat: the servers turn nearly 100% of their power into heat, and so do the losses in transformers, UPS and cables and the work done by the cooling plant (Open Compute Project (opens in a new tab), LBNL best practices guide (opens in a new tab)). So heat to remove ≈ grid supply, minus any heat exported for reuse, for example to a district heating network.
| PUE | Grid supply (X) | UPS losses at 95% | Cooling and other overhead | Heat to remove (Y) |
|---|---|---|---|---|
| 1.1 | 1.10 GW | 0.053 GW | 0.047 GW | about 1.10 GW |
| 1.3 | 1.30 GW | 0.053 GW | 0.247 GW | about 1.30 GW |
| 1.5 | 1.50 GW | 0.053 GW | 0.447 GW | about 1.50 GW |
UPS losses use one double-conversion UPS at 95% efficiency, which LBNL says modern units reach or beat (up from 85% to 90% in the 1990s). LBNL models UPS efficiencies of 90% to 99% for hyperscale sites and 77% to 85% for small edge rooms. Transformer and cable losses are not split out here, so they sit in the overhead column. PUE values of 1.1, 1.3 and 1.5 are examples across the range on this page, for instance Google's 1.09 fleet figure and the Uptime 2025 average of 1.54.
Purple: UPS losses (0.053 GW). Our calculation.
Purple: UPS losses (0.053 GW). Our calculation.
Purple: UPS losses (0.053 GW). Our calculation.
The same 1 GW of IT with each cooling configuration
Grid supply and heat rejected both follow PUE, so the cooling choice changes them. Water follows WUE. All figures are our calculation from the LBNL ranges on this page for 1 GW of IT running all year: grid supply = 1 GW × PUE, heat rejected ≈ grid supply, overhead = PUE − 1, water = 8,760 GWh × WUE.
- Overhead (orange)
- 0.40 GW
- Site water a year
- about 0
Our calculation from LBNL's PUE 1.40 to 2.30 and WUE about 0 L/kWh for this configuration, 1 GW of IT running all year. Water at the best case uses the low WUE and at the worst case the high WUE.
| Configuration | Grid supply, GW | Heat rejected, GW | Cooling and other overhead, GW | Chiller compressor at ASHRAE minimum, GW | Site water, million m³ a year | Noise, sourced example |
|---|---|---|---|---|---|---|
| Air cooling with air-cooled chillers | 1.40 to 2.30 | about 1.40 to 2.30 | 0.40 to 1.30 | 0.34 | about 0 | Air-cooled chiller about 85 to 86 dB(A) sound power per unit |
| Air cooling with water-cooled chillers and cooling towers | 1.30 to 1.77 | about 1.30 to 1.77 | 0.30 to 0.77 | 0.16 | 19.3 to 32.4 | Cooling tower about 102 to 108 dB(A) sound power per unit, by fan speed |
| Evaporative or adiabatic: outside air with adiabatic cooling | 1.02 to 1.28 | about 1.02 to 1.28 | 0.02 to 0.28 | not applicable | 0.4 to 13.6 | No neutral source found |
| Dry coolers with adiabatic assist | 1.07 to 1.35 | about 1.07 to 1.35 | 0.07 to 0.35 | not applicable | 0.4 to 4.4 | Adiabatic cooler about 93.5 dB(A) sound power per unit (one UK assessment) |
| Free-cooling dry coolers with air-cooled chiller back-up | 1.25 to 2.00 | about 1.25 to 2.00 | 0.25 to 1.00 | not applicable | about 0 | Dry cooler about 79 to 90 dB(A) sound power per unit, by fan speed |
| Direct-to-chip liquid with dry coolers | 1.04 to 1.15 | about 1.04 to 1.15 | 0.04 to 0.15 | not applicable | about 0 | Dry cooler about 79 to 90 dB(A) sound power per unit, by fan speed |
Compressor power assumes all 1 GW of IT heat passes through chillers running at the ASHRAE 90.1-2019 full-load minimum (COP 2.96 air-cooled, 6.28 water-cooled), with no free cooling. It is a worst-case illustration, which is why it can exceed the overhead at the low end of the PUE range. Noise examples are per unit, from the planning assessments in the noise section, not per GW.
Grid supply and heat rejected are about the same number, because almost all electricity ends up as heat. Our calculation.
Sources: Lawrence Berkeley National Laboratory: 2024 United States Data Center Energy Usage Report (opens in a new tab) (20 Dec 2024), Open Compute Project: Data centers heat reuse 101 (opens in a new tab) (23 Jun 2023).
Show the data as a table
| Configuration | Best case, GW | Worst case, GW |
|---|---|---|
| Air cooling with air-cooled chillers | 1.40 | 2.30 |
| Air cooling with water-cooled chillers and cooling towers | 1.30 | 1.77 |
| Evaporative or adiabatic: outside air with adiabatic cooling | 1.02 | 1.28 |
| Dry coolers with adiabatic assist | 1.07 | 1.35 |
| Free-cooling dry coolers with air-cooled chiller back-up | 1.25 | 2.00 |
| Direct-to-chip liquid with dry coolers | 1.04 | 1.15 |
Does the cooling plant reject all of it?
- Heat from the IT halls, and from the chillers' own compressors, goes to the main heat rejection plant. A chiller rejects the cooling load plus its compressor work (see the heat rejection basis below).
- Some heat is handled separately. UPS and battery rooms often have their own coolers, for example the battery room coolers in a published Scottish assessment (SEPA, DataVita assessment (opens in a new tab)). Outdoor transformers lose their heat straight to the air. So the main plant handles most, not always all, of the grid supply.
- Heat exported for reuse leaves the site without passing through the heat rejection plant.
Design peak, average and connection capacity
- Plant and grid connections are sized for the design peak on the hottest day, with spare units for redundancy, while the yearly energy follows the average. Planning noise models, for example, leave out the redundant units that stand by (SEPA, DataVita assessment (opens in a new tab)).
- Average use sits well below connected capacity. NESO puts Great Britain's data centres at 7.6 TWh a year from 2.4 GW connected, which is an average of about 0.87 GW, or about 36% of the connection (our calculation, NESO via DCD (opens in a new tab)). LBNL assumes 50% average utilisation when it turns US energy into GW (LBNL (opens in a new tab)). In Texas, ERCOT counted 4,132 MW in use at the September 2026 peak against 9,456 MW approved to energise (ERCOT (opens in a new tab)).
1. Extra power demand from upcoming capacity
All figures after 2024 below are forecasts or scenarios, shown as published with their ranges. They use different scopes and methods, so compare within a chart, not across charts.
The IEA did not state a single 2035 Base Case number on the page we used, so only the three 2035 sensitivity cases are shown.
Source: IEA: Energy and AI: energy demand from AI (opens in a new tab) (10 Apr 2025).
Show the data as a table
| Year and case | TWh |
|---|---|
| 2024 estimate | 415 |
| 2030 Base Case | 945 (forecast) |
| 2035 Headwinds | 700 (forecast) |
| 2035 High Efficiency | 970 (forecast) |
| 2035 Lift-Off | over 1,700 (forecast) |
The IEA says the US and China account for nearly 80% of global growth to 2030. The figures for Europe and Japan are the IEA's 'more than 45' and 'around 15' TWh.
Source: IEA: Energy and AI: energy demand from AI (opens in a new tab) (10 Apr 2025).
Show the data as a table
| Region | TWh added 2024 to 2030 (forecast) | Growth |
|---|---|---|
| United States | 240 | +130% |
| China | 175 | +170% |
| Europe | 45 | +70% |
| Japan | 15 | +80% |
LBNL turns its 2028 range into 74 to 132 GW of average demand, assuming 50% utilisation. The IEA adds around 240 TWh (up 130%) for the US between 2024 and 2030.
Sources: Lawrence Berkeley National Laboratory: 2024 United States Data Center Energy Usage Report (opens in a new tab) (20 Dec 2024), US Department of Energy: DOE releases new report evaluating increase in electricity demand from data centers (opens in a new tab) (20 Dec 2024), EPRI: Powering Intelligence: analyzing artificial intelligence and data center energy consumption (opens in a new tab) (28 May 2024).
Show the data as a table
| Estimate or scenario | TWh |
|---|---|
| LBNL 2023 estimate | 176 |
| LBNL 2028 low to high (forecast) | 325 to 580 |
| EPRI 2030 low (forecast) | 196.3 |
| EPRI 2030 moderate (forecast) | 214.0 |
| EPRI 2030 high (forecast) | 296.4 |
| EPRI 2030 higher (forecast) | 403.9 |
The NESO figures reach us through DCD's report of Future Energy Scenarios 2025. Ireland's 2025 and 2034 figures are EirGrid forecasts with currently contracted demand.
Sources: DCD, reporting NESO Future Energy Scenarios 2025: UK grid operator: high uncertainty on data center growth, 30 to 71 TWh by 2050 (opens in a new tab) (18 Jul 2025), Central Statistics Office (Ireland): Data Centres Metered Electricity Consumption 2024: key findings (opens in a new tab) (10 Jun 2025), Commission for Regulation of Utilities (CRU), citing EirGrid: CRU publishes its decision on new electricity connection policy for data centres (opens in a new tab) (12 Dec 2025).
Show the data as a table
| Place and year | TWh |
|---|---|
| GB now (NESO, 2.4 GW connected) | 7.6 |
| GB 2050 (NESO range) | 30 to 71 (forecast) |
| Ireland 2024 (CSO, metered) | 7.0 |
| Ireland 2025 (EirGrid) | 9.4 |
| Ireland 2034 (EirGrid) | 14.6 |
China and Europe: the IEA forecasts around 175 TWh more in China (up 170%) and more than 45 TWh more in Europe (up 70%) by 2030 than in 2024. We have not found a national China forecast we could check, so none is shown. See also China and Europe and the UK. What that electricity costs on the wholesale market, country by country: European electricity prices by country.
Pipeline and grid queues
Requests to connect are not demand. Many are speculative or the same project at several sites. These figures show pressure on grids, not what will be built.
Every point is ERCOT's own count of large-load interconnection requests (75 MW or more), as published by ERCOT or reported from its filings. Requests are not a forecast of demand: many are speculative or for the same project at several sites.
Sources: The Texas Tribune: Data center explosion creates planning problems for Texas (opens in a new tab) (30 Oct 2025), ERCOT: Board item 16.2: System Planning and Weatherization Update (large loads as of 18 November 2025) (opens in a new tab) (2 Dec 2025), Utility Dive: Facing an estimated 474 GW of interconnection requests, Texas hits pause on data centers (opens in a new tab) (5 Aug 2026).
Show the data as a table
| Date | GW requested |
|---|---|
| Sep 2024 | 56 |
| Dec 2024 | 63 |
| Oct 2025 | 205 |
| Nov 2025 | 226 |
| Jun 2026 | 474 |
US utility pipelines
- Dominion Energy (Virginia (PJM)): over 54 GW, pipeline in contracting stages. As of 30 Oct 2026. Source: The Motley Fool (opens in a new tab)
- American Electric Power (11 states incl. Ohio, Texas, Virginia): 84 GW, contracted load through 2031. As of 28 Oct 2026. Source: Utility Dive (opens in a new tab)
- Southern Company (Georgia, Alabama, Mississippi): 17 GW, contracted large load. As of 30 Jul 2026. Source: Southern Company (opens in a new tab)
More on US grids: US data centre power and energy supply pathways.
2. Electricity and water by data centre type
How we work it out. One MW of IT load running all year uses 1 × 8,760 hours = 8,760 MWh of IT energy. The facility uses that × PUE. Site water is IT energy × WUE: 8,760,000 kWh × WUE (L/kWh) = 8,760 × WUE cubic metres per MW per year. Full load is an upper bound, real sites run below it.
| Type | PUE range | Site WUE (L/kWh) | Facility MWh per MW-year | Water m³ per MW-year | Rack density |
|---|---|---|---|---|---|
| Hyperscale | 1.08 to 1.17 Meta 2024 fleet 1.08, Google 2025 fleet 1.09, Microsoft FY25 fleet 1.17 | 0.19 to 0.27 Meta 2024 0.19, Microsoft FY25 0.27 L/kWh | 9,461 to 10,249 | 1,664 to 2,365 | Mostly 10 to 30 kW per rack across surveyed sites (Uptime 2025, all types) |
| Colocation (large) | 1.30 to 1.48 Uptime 2025: many new high-latitude sites 1.3 or better, new sites average 1.48, sites of 20 MW or more average 1.44 | about 0.36 No colocation-only figure published. LBNL US average for all data centres, 2023, just over 0.36 L/kWh | 11,388 to 12,965 | about 3,154 | Mostly 10 to 30 kW (Uptime 2025) |
| Enterprise (in-house) | 1.54 to 2.30 Uptime 2025 global average 1.54, LBNL midsize air-cooled systems up to about 2.3 | about 0 to 3.70 LBNL midsize: about 0 with air-cooled chillers, up to about 3.7 with water-cooled chillers | 13,490 to 20,148 | 0 to 32,412 | Mostly 10 to 30 kW (Uptime 2025) |
| Edge and small rooms | 1.45 to 2.55 LBNL small spaces (edge, branch, SMB): about 1.45 to 2.55 | about 0 to 4.10 LBNL small spaces: about 0 with DX or air-cooled chillers, up to about 4.1 with water-cooled chillers | 12,702 to 22,338 | 0 to 35,916 | Not published in our sources |
| AI training (liquid-cooled) | 1.04 to 1.25 LBNL liquid-cooled IT: about 1.04 to 1.25 | about 0 to 2.45 LBNL liquid-cooled IT: about 0 with dry coolers, up to about 2.45 with cooling towers | 9,110 to 10,950 | 0 to 21,462 | approximately 120 kW (GB200 NVL72) to up to 142 kW (GB300 NVL72) per rack, from NVIDIA documentation |
Sources: Uptime Institute: Global Data Center Survey 2025 (opens in a new tab) (30 Jul 2025), DCD, reporting the Uptime Institute 2025 survey: Roughly one-third of data center owners and operators doing AI training or inference (opens in a new tab) (30 Jul 2025), Lawrence Berkeley National Laboratory: 2024 United States Data Center Energy Usage Report (opens in a new tab) (20 Dec 2024), Google: Power usage effectiveness: Google data centers (opens in a new tab) (accessed 3 Nov 2026), Microsoft: Measuring energy and water efficiency for Microsoft datacenters (opens in a new tab) (accessed 3 Nov 2026), Meta: 2025 Environmental Data Index (opens in a new tab) (1 Oct 2025). AI rack figures: NVIDIA, approximately 120 kW (GB200 NVL72, rack power consumption) and up to 142 kW (GB300 NVL72, full rack power requirement), from our IT equipment library. LBNL ranges are read from its Figure 4.4 and rounded, so treat them as approximate.
Hyperscale WUE figures are fleet withdrawal or consumption figures as each company defines them. LBNL also estimates about 4.52 L of water used off-site per kWh, in generating the electricity, for the average US data centre in 2023. That is not included above.
Lower is better. A PUE of 1.5 means 0.5 kWh of overhead for every kWh used by IT.
Sources: Uptime Institute: Global Data Center Survey 2025 (opens in a new tab) (30 Jul 2025), Lawrence Berkeley National Laboratory: 2024 United States Data Center Energy Usage Report (opens in a new tab) (20 Dec 2024), Google: Power usage effectiveness: Google data centers (opens in a new tab) (accessed 3 Nov 2026), Microsoft: Measuring energy and water efficiency for Microsoft datacenters (opens in a new tab) (accessed 3 Nov 2026), Meta: 2025 Environmental Data Index (opens in a new tab) (1 Oct 2025).
Show the data as a table
| Type | PUE low | PUE high |
|---|---|---|
| Hyperscale | 1.08 | 1.17 |
| Colocation (large) | 1.30 | 1.48 |
| Enterprise (in-house) | 1.54 | 2.30 |
| Edge and small rooms | 1.45 | 2.55 |
| AI training (liquid-cooled) | 1.04 | 1.25 |
3. How heat removal changes electricity and water
Almost all the electricity used by IT ends up as heat that has to be removed. Systems that evaporate water reject that heat with less fan and compressor work, so PUE falls but water use rises. Dry and air-cooled systems use almost no water on site but need more electricity, especially in hot weather. LBNL notes there is often a trade-off between low PUE and low site WUE. Liquid cooling at the chip runs at warmer water temperatures, so more hours can use free cooling.
Low end: best practice in a favourable climate. High end: poor practice in a hot, humid climate.
Source: Lawrence Berkeley National Laboratory: 2024 United States Data Center Energy Usage Report (opens in a new tab) (20 Dec 2024). Read from Figure 4.4 and rounded.
Show the data as a table
| Method | Data centre size in LBNL | PUE low | PUE high |
|---|---|---|---|
| Air cooling with air-cooled chillers | midsize | 1.40 | 2.30 |
| Air cooling with water-cooled chillers and cooling towers | midsize | 1.30 | 1.77 |
| Evaporative or adiabatic: outside air with adiabatic cooling | large-scale | 1.02 | 1.28 |
| Dry coolers with adiabatic assist | large-scale | 1.07 | 1.35 |
| Free-cooling dry coolers with air-cooled chiller back-up | midsize | 1.25 | 2.00 |
| Free cooling: outside air economiser, no evaporation | large-scale | 1.02 | 1.35 |
| Direct-to-chip liquid with dry coolers | liquid-cooled IT | 1.04 | 1.15 |
| Direct-to-chip liquid with waterside economiser and towers | liquid-cooled IT | 1.05 | 1.25 |
Site water only. Water used at power stations is not included.
Source: Lawrence Berkeley National Laboratory: 2024 United States Data Center Energy Usage Report (opens in a new tab) (20 Dec 2024). Read from Figure 4.4 and rounded.
Show the data as a table
| Method | WUE low | WUE high |
|---|---|---|
| Air cooling with air-cooled chillers | 0.00 | 0.00 |
| Air cooling with water-cooled chillers and cooling towers | 2.20 | 3.70 |
| Evaporative or adiabatic: outside air with adiabatic cooling | 0.05 | 1.55 |
| Dry coolers with adiabatic assist | 0.05 | 0.50 |
| Free-cooling dry coolers with air-cooled chiller back-up | 0.00 | 0.00 |
| Free cooling: outside air economiser, no evaporation | 0.00 | 0.00 |
| Direct-to-chip liquid with dry coolers | 0.00 | 0.00 |
| Direct-to-chip liquid with waterside economiser and towers | 1.65 | 2.45 |
Immersion cooling is not in LBNL's modelled set and we found no independent, published fleet PUE or WUE for it, so it is left out of the charts and calculator. Learn more in the technology guides for direct liquid cooling and other cooling technologies.
Per 1 GW of heat rejected, not 1 GW of IT
Heat rejection plant is sized for the heat it must reject, which is roughly the IT load plus the cooling plant's own power and the electrical losses. For a chiller, heat rejected = cooling load + compressor work. With COP = cooling load ÷ compressor power, the heat rejection ratio = 1 + 1 ÷ COP.
- Air-cooled chiller at the ASHRAE 90.1-2019 minimum of 10.1 EER (Btu/Wh) at full load: COP = 10.1 ÷ 3.412 = 2.96. Heat rejection ratio = 1 + 1 ÷ 2.96 = 1.338. Per 1 GW rejected: 747 MW of cooling load and 253 MW of compressor power.
- Water-cooled centrifugal chiller of 600 tons or more at the ASHRAE minimum of 0.560 kW/ton: COP = 3.517 ÷ 0.560 = 6.28. Heat rejection ratio = 1.159. Per 1 GW rejected at the tower: 863 MW of cooling load and 137 MW of compressor power.
These are code minimum ratings at standard test conditions (ASHRAE 90.1-2019, Table 6.8.1-3 (opens in a new tab)). Real plants run at part load and other temperatures, and fans and pumps come on top. Our calculation.
How much water a cooling tower evaporates per GW
Evaporating water absorbs its latent heat. NIST gives 2,555.5 kJ/kg for saturated vapour and 125.73 kJ/kg for liquid water at 30 °C, so the latent heat is 2,429.8 kJ/kg (NIST (opens in a new tab)). If all 1 GW left as latent heat: 1,000,000 kJ/s ÷ 2,429.8 kJ/kg = 412 kg/s, or about 1,482 m³/h. The engineering rule of thumb uses 1,000 Btu/lb (2,326 kJ/kg), which gives about 1,548 m³/h. The same rule is usually written as evaporation of about 1% of the circulating flow for every 10 °F (5.6 °C) of cooling range (engineering reference, Wiley (opens in a new tab), US EPA WaterSense (opens in a new tab)). Some heat leaves the tower as sensible heat, so these are upper estimates.
Towers also lose water as blowdown, which removes the minerals left behind, and as drift. The US Department of Energy gives make-up = evaporation + blowdown + drift, and blowdown = evaporation ÷ (cycles of concentration − 1), with drift and leaks counted in blowdown (DOE FEMP (opens in a new tab)). Many systems run at 2 to 4 cycles, and 6 or more may be possible (DOE FEMP (opens in a new tab)). We found no neutral figure for drift on its own, so it is not shown separately.
| Cycles of concentration | Evaporation, m³/h | Blowdown, m³/h | Make-up, m³/h | Make-up, million m³ a year | Per 1 MW of heat, m³/h |
|---|---|---|---|---|---|
| 2 | 1,482 to 1,548 | 1,482 to 1,548 | 2,963 to 3,095 | 26.0 to 27.1 | 2.96 to 3.10 |
| 3 | 1,482 to 1,548 | 741 to 774 | 2,222 to 2,322 | 19.5 to 20.3 | 2.22 to 2.32 |
| 4 | 1,482 to 1,548 | 494 to 516 | 1,975 to 2,064 | 17.3 to 18.1 | 1.98 to 2.06 |
| 6 | 1,482 to 1,548 | 296 to 310 | 1,778 to 1,857 | 15.6 to 16.3 | 1.78 to 1.86 |
Each method per 1 GW of heat rejected
For whole sites we use the LBNL PUE and WUE ranges and our approximation that heat rejected ≈ facility electricity. So per 1 GW of heat: IT supported = 1,000 ÷ PUE MW, overhead = 1,000 × (1 − 1 ÷ PUE) MW, and site water = IT energy × WUE. Divide by 1,000 for 1 MW of heat. We found no neutral, published fan and pump power ratios for dry coolers, so their electricity comes from the PUE range too.
| Method | IT supported, MW | Overhead electricity, MW | Overhead, GWh a year | Site water, m³/h | Site water, million m³ a year |
|---|---|---|---|---|---|
| Air cooling with air-cooled chillers | 435 to 714 | 286 to 565 | 2,503 to 4,951 | about 0 | about 0 |
| Air cooling with water-cooled chillers and cooling towers | 565 to 769 | 231 to 435 | 2,022 to 3,811 | 1,243 to 2,846 | 10.89 to 24.93 |
| Evaporative or adiabatic: outside air with adiabatic cooling | 781 to 980 | 20 to 219 | 172 to 1,916 | 39 to 1,520 | 0.34 to 13.31 |
| Free-cooling dry coolers with air-cooled chiller back-up | 500 to 800 | 200 to 500 | 1,752 to 4,380 | about 0 | about 0 |
| Direct-to-chip liquid with dry coolers | 870 to 962 | 38 to 130 | 337 to 1,143 | about 0 | about 0 |
The more efficient the cooling, the more of each GW of heat comes from useful IT load.
Source: Lawrence Berkeley National Laboratory: 2024 United States Data Center Energy Usage Report (opens in a new tab) (20 Dec 2024). Our calculation.
Show the data as a table
| Method | IT low, MW | IT high, MW |
|---|---|---|
| Air cooling with air-cooled chillers | 435 | 714 |
| Air cooling with water-cooled chillers and cooling towers | 565 | 769 |
| Evaporative or adiabatic: outside air with adiabatic cooling | 781 | 980 |
| Free-cooling dry coolers with air-cooled chiller back-up | 500 | 800 |
| Direct-to-chip liquid with dry coolers | 870 | 962 |
From LBNL WUE ranges. Dry and air-cooled methods use about 0 on site. Compare with the tower physics above.
Source: Lawrence Berkeley National Laboratory: 2024 United States Data Center Energy Usage Report (opens in a new tab) (20 Dec 2024). Our calculation.
Show the data as a table
| Method | Low, m³/h | High, m³/h |
|---|---|---|
| Air cooling with air-cooled chillers | 0 | 0 |
| Air cooling with water-cooled chillers and cooling towers | 1,243 | 2,846 |
| Evaporative or adiabatic: outside air with adiabatic cooling | 39 | 1,520 |
| Free-cooling dry coolers with air-cooled chiller back-up | 0 | 0 |
| Direct-to-chip liquid with dry coolers | 0 | 0 |
Noise per GW of heat is not shown. It depends on how many units are used, their fan speed and the layout, and we found no neutral source that gives it per GW. Per-unit levels from planning assessments are in the noise section.
Calculator: electricity and water for your IT load or heat rejected
- IT energy per year
- 87,600 MWh
- Facility electricity per year
- 122,640 to 201,480 MWh
- Overhead (cooling, power losses, other)
- 35,040 to 113,880 MWh
- On-site water per year (from WUE)
- about 0 litres (humidification only)
Formulas. IT energy (MWh) = MW × 8,760 hours × load share. Facility electricity = IT energy × PUE. Overhead = IT energy × (PUE − 1). Water (litres) = IT energy (kWh) × WUE (L/kWh).
For air cooling with air-cooled chillers we use PUE 1.40 to 2.30 and WUE about 0 L/kWh, the midsize range read from LBNL (December 2024), Figure 4.4. This is a screening estimate, not a design figure. At 100% load the result is an upper bound.
Where facility power goes
For every 100 kWh a site takes from the grid, 100 ÷ PUE kWh reach the IT equipment and the rest is overhead. Two real reference points: Google's 2025 fleet PUE of 1.09 and the Uptime 2025 survey average of 1.54.
kWh per 100 kWh from the grid. Our calculation from the published PUE. Source: Google (opens in a new tab).
kWh per 100 kWh from the grid. Our calculation from the published PUE. Source: Uptime Institute (opens in a new tab).
Inside the IT share, the IEA says servers take around 60% of a modern data centre's electricity on average, storage around 5% and networking up to 5% (IEA (opens in a new tab)).
Data centre noise levels: dry coolers, cooling towers, adiabatic coolers, chillers and generators
Two different measures are used. Sound power (LwA) is the noise a machine emits, whatever the distance. Sound pressure (LpA) is what you hear at a stated distance. Both are in dB(A). The figures below come from published UK and US planning and permit noise assessments for real data centres. They are examples, not typical values for every product, and no brand is ranked.
All sound power (LwA). Fan speed matters: the dry coolers in one assessment are 11 dB quieter at the lower night-time fan speed, and the cooling towers in another are 6 dB quieter at two-thirds speed. The adiabatic cooler figure is the consultant's calculation from 65.5 dB(A) measured at 10 m. Each row comes from a different site and product, so compare orders of magnitude, not brands. The axis starts at 50 dB(A).
Sources: SEPA permit consultation (DataVita DV1 plant noise impact assessment): Appendix J1: noise assessment (opens in a new tab) (6 Aug 2026), Environment Agency permit consultation (Amazon Data Services UK, Didcot): Environmental permit application: noise impact assessment (opens in a new tab) (30 Sept 2025), East Whiteland Township planning file (B. Laing Associates for the Green Fig Land Company data centre, Exton, Pennsylvania): Supplemental sound analysis with barrier, letter to the Director of Planning (opens in a new tab) (15 Mar 2024), UK planning file (E3P noise impact assessment for the Erin Landfill data centre, Chesterfield): Noise Impact Assessment, Erin Landfill Data Centre, report 51-854-R1 (opens in a new tab) (22 Oct 2025).
Show the data as a table
| Plant | Sound power LwA, dB(A) | Source |
|---|---|---|
| Dry cooler, night fan speed (17 °C) | 79 | SEPA permit consultation (DataVita DV1 plant noise impact assessment) |
| Dry cooler, near-peak day fan speed (24 °C) | 90 | SEPA permit consultation (DataVita DV1 plant noise impact assessment) |
| Air-cooled chiller | 85 to 86 | SEPA permit consultation (DataVita DV1 plant noise impact assessment) |
| Air handling unit or fan wall intake (50% to 70% fan speed) | 80 to 88 | SEPA permit consultation (DataVita DV1 plant noise impact assessment) |
| CRAC unit | 81 | Environment Agency permit consultation (Amazon Data Services UK, Didcot) |
| Air handling unit exhaust | 89 | Environment Agency permit consultation (Amazon Data Services UK, Didcot) |
| Standby generator (enclosed) | 96 to 103 | SEPA permit consultation (DataVita DV1 plant noise impact assessment) and Environment Agency permit consultation (Amazon Data Services UK, Didcot) |
| Cooling tower, 2/3 fan speed | 102 | East Whiteland Township planning file (B. Laing Associates for the Green Fig Land Company data centre, Exton, Pennsylvania) |
| Cooling tower, full speed | 108 | East Whiteland Township planning file (B. Laing Associates for the Green Fig Land Company data centre, Exton, Pennsylvania) |
| Adiabatic cooler | 93.5 | UK planning file (E3P noise impact assessment for the Erin Landfill data centre, Chesterfield) |
| Transformer | 64 | SEPA permit consultation (DataVita DV1 plant noise impact assessment) |
| Plant | Sound pressure, dB(A) | Distance | Source |
|---|---|---|---|
| Standby generators, attenuated | 70 to 75 | 1 m | SEPA, DataVita assessment (opens in a new tab) |
| Emergency generator package (design target) | 80 | at the package surface | Environment Agency, Didcot assessment (opens in a new tab) |
| Transformer | 75 | 1 m | Environment Agency, Didcot assessment (opens in a new tab) |
How distance and the number of units change it
- Distance: away from a compact source in the open, sound pressure falls by about 6 dB each time the distance doubles (HSE RR1223 (opens in a new tab)). Worked example from the 75 dB(A) at 1 m generator above: about 69 at 2 m, 63 at 4 m, 57 at 8 m, 51 at 16 m and 45 at 32 m, before screening or ground effects. Formula: LpA(r) = LpA(1 m) − 20 × log10(r).
- Number of units: two identical machines running together are 3 dB louder than one, not twice the decibels (HSE (opens in a new tab)). In general the increase is 10 × log10(n): about +6 dB for 4 units and +10 dB for 10 units.
- Fan speed and temperature: dry coolers run their fans faster on hot days, which is why planning assessments model near-peak temperatures (SEPA, DataVita assessment (opens in a new tab)).
Planning limits and guidance
- UK, BS 4142:2014+A1:2019. The rating level of the plant (including corrections for tones or impulses) is compared with the background level at homes. A difference of around +10 dB or more is likely to indicate a significant adverse impact, and around +5 dB an adverse impact, depending on context (Environment Agency guidance on BS 4142 (opens in a new tab)).
- WHO Guidelines for Community Noise (1999). Outdoor living areas: 55 dB LAeq over 16 hours for serious annoyance and 50 dB for moderate annoyance. Outside bedrooms at night: 45 dB LAeq over 8 hours and 60 dB LAmax. Inside bedrooms: 30 dB LAeq (WHO (opens in a new tab)).
Cooling towers and adiabatic coolers (added 3 Nov 2026): the cooling tower rows are from a US planning file for a two-building data centre with 20 towers at grade, where running the three towers nearest homes at two-thirds speed cut their sound power by 6 dB and a planned 40 ft barrier wall was modelled to bring boundary levels down to 48.5 to 59.3 dB(A) (East Whiteland Township planning file (opens in a new tab)). The adiabatic cooler row is from a UK planning noise assessment, which also used 1 dB of added uncertainty in its model (E3P for Erin Landfill data centre (opens in a new tab)). Still left out: low-noise or EC fan dry cooler variants, which we found only in manufacturers' datasheets.
Sources
- IEA: Energy and AI: energy demand from AI (opens in a new tab) published 10 Apr 2025, accessed 3 Nov 2026
- Lawrence Berkeley National Laboratory: 2024 United States Data Center Energy Usage Report (opens in a new tab) published 20 Dec 2024, accessed 3 Nov 2026
- US Department of Energy: DOE releases new report evaluating increase in electricity demand from data centers (opens in a new tab) published 20 Dec 2024, accessed 3 Nov 2026
- EPRI: Powering Intelligence: analyzing artificial intelligence and data center energy consumption (opens in a new tab) published 28 May 2024, accessed 3 Nov 2026
- DCD, reporting NESO Future Energy Scenarios 2025: UK grid operator: high uncertainty on data center growth, 30 to 71 TWh by 2050 (opens in a new tab) published 18 Jul 2025, accessed 3 Nov 2026
- Commission for Regulation of Utilities (CRU), citing EirGrid: CRU publishes its decision on new electricity connection policy for data centres (opens in a new tab) published 12 Dec 2025, accessed 3 Nov 2026
- Central Statistics Office (Ireland): Data Centres Metered Electricity Consumption 2024: key findings (opens in a new tab) published 10 Jun 2025, accessed 3 Nov 2026
- Uptime Institute: Global Data Center Survey 2025 (opens in a new tab) published 30 Jul 2025, accessed 3 Nov 2026
- DCD, reporting the Uptime Institute 2025 survey: Roughly one-third of data center owners and operators doing AI training or inference (opens in a new tab) published 30 Jul 2025, accessed 3 Nov 2026
- Google: Power usage effectiveness: Google data centers (opens in a new tab) accessed 3 Nov 2026
- Microsoft: Measuring energy and water efficiency for Microsoft datacenters (opens in a new tab) accessed 3 Nov 2026
- Meta: 2025 Environmental Data Index (opens in a new tab) published 1 Oct 2025, accessed 3 Nov 2026
- SEPA permit consultation (DataVita DV1 plant noise impact assessment): Appendix J1: noise assessment (opens in a new tab) published 6 Aug 2026, accessed 3 Nov 2026
- Environment Agency permit consultation (Amazon Data Services UK, Didcot): Environmental permit application: noise impact assessment (opens in a new tab) published 30 Sept 2025, accessed 3 Nov 2026
- Environment Agency permit consultation (Virtus, Saunderton): Stage 3 acoustic report (opens in a new tab) accessed 3 Nov 2026
- Environment Agency (GOV.UK): Method implementation document (MID) for BS 4142 (opens in a new tab) accessed 3 Nov 2026
- World Health Organization: Guidelines for community noise (opens in a new tab) published 1 Apr 1999, accessed 3 Nov 2026
- Health and Safety Executive: Noise (adding sound levels) (opens in a new tab) accessed 3 Nov 2026
- Health and Safety Executive: RR1223: noise control in woodworking, review and illustration of control methods (opens in a new tab) accessed 3 Nov 2026
- ASHRAE: Addendum x to ANSI/ASHRAE/IES Standard 90.1-2019, Table 6.8.1-3 liquid-chilling packages minimum efficiency (opens in a new tab) published 4 Feb 2022, accessed 3 Nov 2026
- NIST Chemistry WebBook: Thermophysical properties of water, saturation properties at 20, 30 and 40 °C (opens in a new tab) accessed 3 Nov 2026
- Wiley (engineering reference): Appendix E: Cooling water calculations (evaporation from a tower heat balance) (opens in a new tab) accessed 3 Nov 2026
- US Department of Energy (FEMP): Cooling towers: understanding key components and how to improve water efficiency (opens in a new tab) published 1 Feb 2011, accessed 3 Nov 2026
- US Department of Energy (FEMP): Best Management Practice 10: cooling tower management (opens in a new tab) accessed 3 Nov 2026
- US EPA WaterSense: WaterSense at Work, Section 6.3: cooling towers (opens in a new tab) published 1 May 2023, accessed 3 Nov 2026
- Open Compute Project: Data centers heat reuse 101 (opens in a new tab) published 23 Jun 2023, accessed 3 Nov 2026
- LBNL Center of Expertise for Energy Efficiency in Data Centers: Best practices guide for energy-efficient data center design (opens in a new tab) published 1 Jul 2025, accessed 3 Nov 2026
- East Whiteland Township planning file (B. Laing Associates for the Green Fig Land Company data centre, Exton, Pennsylvania): Supplemental sound analysis with barrier, letter to the Director of Planning (opens in a new tab) published 15 Mar 2024, accessed 3 Nov 2026
- UK planning file (E3P noise impact assessment for the Erin Landfill data centre, Chesterfield): Noise Impact Assessment, Erin Landfill Data Centre, report 51-854-R1 (opens in a new tab) published 22 Oct 2025, accessed 3 Nov 2026
Pipeline and grid queue figures come from our US data centre power data, each with its own source above. Spotted an error? Tell us.
Common questions
How much extra electricity will new data centres need?
The IEA forecasts global data centre electricity use roughly doubling from about 415 TWh in 2024 to about 945 TWh in 2030 in its Base Case.
Why it matters That growth lands in a few places, so local grids feel it far more than the global share suggests.
- About 415 TWh in 2024
- About 945 TWh in 2030, Base Case forecast
- About 700 to over 1,700 TWh in 2035 across the IEA's other cases
Source: IEA, Energy and AI (April 2025)
How much electricity will US data centres use?
LBNL forecasts 325 to 580 TWh in 2028, up from 176 TWh in 2023. EPRI's scenarios give 196 to 404 TWh in 2030.
Why it matters At the top of LBNL's range data centres would use about 12% of US electricity.
- 176 TWh in 2023, 4.4% of US use
- 325 to 580 TWh in 2028, forecast
- 196.3 to 403.9 TWh in 2030, EPRI scenarios
Source: LBNL (December 2024), EPRI (May 2024)
How much heat does 1 GW of IT produce?
Roughly the whole grid supply, which is the IT load times PUE. At a PUE of 1.3, 1 GW of IT needs about 1.3 GW from the grid and about 1.3 GW of heat has to be removed.
Why it matters Almost all electricity used on site ends up as heat, unless some is exported for heat reuse.
- Grid supply = IT load × PUE
- Chiller heat rejection ratio = 1 + 1 ÷ COP
- A tower evaporates up to about 1,480 to 1,550 m³ an hour per GW of heat
Source: Open Compute Project, ASHRAE 90.1-2019, NIST, our calculation
How much electricity does 1 MW of IT load use in a year?
8,760 MWh of IT energy if it runs flat out all year, plus overhead set by PUE.
Why it matters At a PUE of 1.5 the site draws 13,140 MWh, at 1.1 it draws 9,636 MWh.
- 1 MW × 8,760 hours = 8,760 MWh
- Facility energy = IT energy × PUE
Source: Our calculation, PUE figures from Uptime Institute 2025 and LBNL 2024
How much water does a data centre use?
It depends mostly on the cooling. LBNL's US average was just over 0.36 litres per kWh of IT energy in 2023, while water-cooled chiller plants can use several litres per kWh and dry systems almost none.
Why it matters 1 MW of IT at 0.36 L/kWh is about 3,150 cubic metres a year.
- US average about 0.36 L/kWh (2023)
- Microsoft FY25 0.27 L/kWh
- Meta 2024 0.19 L/kWh
Source: LBNL (December 2024), Microsoft, Meta 2025 Environmental Data Index
Does liquid cooling save energy and water?
LBNL's modelling puts direct-to-chip liquid cooling at a PUE of about 1.04 to 1.25. With dry coolers it uses almost no water, with cooling towers about 1.65 to 2.45 L/kWh.
Why it matters Warmer coolant allows more hours of free cooling, which cuts compressor work.
- PUE about 1.04 to 1.15 with dry coolers
- PUE about 1.05 to 1.25 with towers
Source: LBNL 2024, Figure 4.4 (approximate readings)
How loud are data centre generators and dry coolers?
In published UK assessments, standby generators emit about 96 to 103 dB(A) sound power and dry coolers about 79 to 90 dB(A), depending on fan speed.
Why it matters What neighbours hear also depends on distance, the number of units and screening.
- Generators 70 to 75 dB(A) at 1 m in one assessment
- About 6 dB less per doubling of distance
- +3 dB for two identical units
Source: SEPA and Environment Agency permit noise assessments, HSE
What noise limits apply to data centre plant in the UK?
Plant noise is usually judged with BS 4142, which compares the plant's rating level with the background level at nearby homes.
Why it matters Around +10 dB above background is likely a significant adverse impact, around +5 dB an adverse impact, depending on context.
- WHO 1999: 45 dB LAeq outside bedrooms at night
- WHO 1999: 50 to 55 dB LAeq in outdoor living areas by day
Source: Environment Agency guidance on BS 4142, WHO Guidelines for Community Noise (1999)