EVOPEDIA

← all systems · see it in The System

Data centre

A building or campus purpose-built to house computing: 'a facility for storing, managing, and disseminating digital information, including computer systems and associated components, IT infrastructure, equipment to maintain the proper conditions for the operation of electronic equipment' (src:wp:data-center). At scale its engineers treat it as one machine: 'warehouse-sized computing systems, made up of thousands of computing nodes, their associated storage hierarchy and interconnection infrastructure' (src:fan-2007), fed by a medium-voltage grid connection and cooled by air, water or liquid loops.

Properties published scienceour proposal

Sourced fields, and our own readings.

aliasesData centre (ML accelerators, power)
born at1940s (ENIAC computer rooms); contested to the 1980s or 1997-2000
born precisiondecade
boundaryInside: the site, the building shell and its rooms (white space with racks; electrical, mechanical and network rooms), the power train from the utility feed through transformers, UPS, generators and PDUs to the rack circuits, the cooling plant (chillers, cooling towers, economisers, fans, pumps, liquid loops), the network entry (meet-me rooms), fire and security systems, the sensors and building controls, and the people and software that run the facility. Outside: the electrical grid and its generators, the water supply, the outside air, the internet and carriers beyond the meet-me room, the tenants and their software, the communities, regulators and markets around it. The servers are inside the building but their owners may be outside it (colocation).
nested levelBuilt from computer hardware (servers, storage, network gear) in racks, with electrical and mechanical plant; embedded in the electrical grid, the water and air of its site, the internet, the economy of the companies that own and rent it, and the local community and state that permit it. A hyperscale data centre is the physical layer of a cloud.
scaletechnological

Measures (33) published science

water used by a small 1 MW data centre with evaporative coolingabout 25.5 million litres per year
internal noise levelup to 96 dBA
projects blocked or delayed by local opposition (US, to 2025)64 billion US dollars
US data-centre direct water consumption 202366 billion (21.2 billion in 2014) litres
characteristic clock: grid frequency settling after a large load lossabout 4 min
characteristic clock: grid fault durations that trigger UPS transfer42 to 66 ms
voltage-disturbance counting threshold3 disturbances within 1 min
characteristic clock: carbon-aware planning cycle1 (curves set day-ahead; flexible work done within 24 h)
characteristic clock: mean facility age11.3 years (December 2024)
load lost at once in the 2024 Eastern Interconnection incidentabout 1,500 MW
grid frequency peak after the 2024 load loss60.047 Hz
US data-centre electricity use 2023176 (4.4% of US electricity) TWh
US data-centre share of electricity 2028 (projected)6.7 to 12.0 % of US electricity
US data-centre electricity 2014-2016 (flat era)about 60 TWh per year
US average PUE1.6 (2014) to 1.4 (2023); 1.15-1.35 projected 2028 ratio
Google fleet PUE 2008 to 20131.21 to 1.12 ratio
hyperscale PUE range reported1.25 to 1.18 ratio
traditional data-centre PUE2.23 (2010) to 1.93 (2020) ratio
plant-model PUE prediction error0.004 +/- 0.005 (0.4% at PUE 1.1) PUE units
gap between achieved and theoretical aggregate peak power7-16% (cluster); almost 40% (whole data centre)
building cost per watt of critical power10 to 20 US dollars per watt (2007)
PDU rating75 to 200 kW
server utilisation, typical10 to 50 % of maximum
energy saved per degree of setpoint raised2 to 5 % of energy consumption
chilled water temperature, usual7 to 10 degrees C
on-site water evaporated per kWh of server energyabout 1 to 9 L/kWh
out-of-spec hours, southeast United Statesabout 3,000 (a quarter to a third of the year) hours per year
rack size10 to 80 nodes per rack; 20 to 60 racks per PDU count
Ireland: data centres' share of metered electricity 202321 (5 in 2015, 18 in 2022) %
air-cooled server outlet temperature35 to 45 degrees C
trend in out-of-spec hours (tropics)up to 2 hours per day per decade
US data-centre indirect water footprint 2023nearly 800 billion (4.52 L/kWh) litres
economisation saving in cooling energy (suitable climates)20 to 50 %

Open contentions (3) open contention

Where sources or readings disagree. Both stay on the map.

Partners and parts published science

Built from / builds (10)

Sits inside / holds (8)

Trades with (valence) (15)

expressed by (26)

floor filled by (21)

Kinds (7)

Notes published science

Layers

The ways it takes in its world, where they meet, and how it acts back, each from L1 (where it meets the world) up.

heat out: from the chips through air or liquid to the outside world

  1. L1: Heat made in the chips, power supplies and power conversion
  2. L2: Server-level cooling: air drawn across hot metal (fans, cold aisles) or liquid to the chip (direct-to-chip, immersion)
  3. L3: The room loop: CRAH units and chilled water at 7-10 C carrying heat to the plant; or outside air in free-cooling hours
  4. L4: Heat rejection: cooling towers evaporating water, dry coolers, chillers' condensers, economisers; a small share recovered for building or district heating

the boardroom (shared floors)

  1. L5: Facility monitoring and control: the building management system and its sensors, setpoints for chillers, towers, pumps and fans, PUE and WUE, UPS and protection logic, fire detection and physical security
  2. L6: Operations: capacity and power provisioning (oversubscribing measured headroom), workload placement and consolidation, power capping, carbon-aware hourly capacity curves, remote operation
  3. L7: The operator: the company's decisions on what to build where, what to sell or run, which energy to buy, and how to answer regulators and neighbours

the data centre as facility engineers and operators describe it: the site's power, water, air and fibre up to the operator's purpose (reader's view)

  1. L1: The site and what reaches it: a medium-voltage feed from a utility substation, municipal or reclaimed water, outside air, fibre from carriers, land with its climate and neighbours
  2. L2: The intake plant: transformers, UPS (static battery or flywheel DRUPS), generators and the automatic transfer switch; chillers, cooling towers and economiser intakes; the network entry (meet-me rooms, border routers)
  3. L3: Distribution: two independent sides (A and B), each UPS feeding PDUs of 75-200 kW paired with static transfer switches, then branch circuits to racks; chilled-water loops and hot/cold aisles; the data-centre network fabric
  4. L4: The white space: racks of servers, storage and network gear (and GPU-accelerated AI servers) doing the computing, mostly between 10 and 50% of peak utilisation
  5. L5: Facility monitoring and control: thousands of sensors, the building management system, setpoints for chillers, towers and pumps, PUE and WUE tracking, UPS and protection logic, fire detection and physical security
  6. L6: Operations: capacity and power provisioning within the power budget, workload placement and server consolidation, power capping, carbon-aware hourly capacity curves, remote (lights-out) operation
  7. L7: The operator's purpose: an enterprise, colocation provider or hyperscaler decides what to build where, provides space, power, cooling, security and interconnection to tenants or runs its own cloud, buys energy, and answers to regulators and neighbours

way in: Demand for space, power and computing from tenants and the operator's own services

  1. L1: Tenants and services needing computing capacity
  2. L2: Contracts set a negotiated load, enforced by penalties
  3. L3: The operator must deliver each branch circuit's rated load and make sure the upper levels can carry it

way in: Equipment health: errors, failures, drive and memory faults

  1. L1: Component faults: latent sector errors, DRAM errors, disk failures, rising with temperature
  2. L2: Drive self-reporting (SMART) and error logs
  3. L3: Field data collected across production fleets

way in: Grid carbon intensity and price signals

  1. L1: Carbon per kWh on the grid varies by hour and by place
  2. L2: Next-day carbon-intensity forecasts are gathered
  3. L3: Day-ahead demand prediction models are trained

way in: Grid power (voltage and frequency at the feed)

  1. L1: The utility's medium-voltage feed, its voltage and frequency
  2. L2: UPS and protection relays watch the voltage; a disturbance hands the load to batteries or a flywheel at once; some sites count dips (three in a minute) and leave the grid
  3. L3: A and B sides, PDUs and static transfer switches route power to the racks
  4. L4: Server power supplies draw it, more or less with utilisation

way in: Heat and humidity inside the rooms

  1. L1: Heat from the equipment in the room air or coolant
  2. L2: Inlet temperature, dew point and humidity sensors read against the recommended envelope
  3. L3: The building management system gathers thousands of sensors
  4. L4: Sensor readings combined into plant-level variables (totals, means) that models and operators use

way in: Neighbours and public opinion

  1. L1: Residents and groups: concerns over water, bills, noise, land and sovereignty

way in: Outside air: temperature, humidity, wet bulb, wind

  1. L1: The weather at the site
  2. L2: Economiser intakes and dampers take outdoor air in directly or through a heat exchanger
  3. L3: Hours when outdoor heat and humidity exceed the inlet limits are counted as out-of-spec
  4. L4: Wet bulb, dry bulb, enthalpy, humidity and wind enter the plant model with the IT load

way in: People at the perimeter and doors

  1. L1: People approaching the site
  2. L2: Badge and fingerprint readers, mantraps, guards and cameras
  3. L3: Access logged at entrance, inner rooms and cabinets
  4. L4: Cabinet locks networked through intelligent PDUs

way in: Rules: permits, tax incentives, tariffs, moratoria, disclosure

  1. L1: Laws, incentives, tariffs and moratoria in the host jurisdiction

way in: Smoke and heat from a fire

  1. L1: Overheating and fire in equipment or plant
  2. L2: Smoke detectors

way in: Water supply for cooling

  1. L1: Municipal, potable or reclaimed water at the site
  2. L2: Make-up water into cooling towers where air crosses wet media and water evaporates
  3. L3: Metered as water usage effectiveness (litres per kWh), where measured

way in: Work arriving: requests and jobs over the network

  1. L1: Traffic over fibre from carriers, exchange points and cable landings
  2. L2: Border routers and cross-connects; two or more upstream providers
  3. L3: The data-centre network fabric carries it to servers
  4. L4: Servers run it: training at high, steady utilisation with transients; inference bursty with user demand
  5. L5: Power capping, accelerator scheduling and rack storage shape the electrical load the work makes

way out: Building and siting facilities

  1. L1: New buildings, substations and connections appear, which can take years
  2. L7: The operator decides to add capacity and where

way out: Fire suppression

  1. L1: Gas (or mist outside the main room) floods the space
  2. L5: Detection triggers suppression

way out: Flexible work moved in time (carbon-aware capacity curves)

  1. L1: The power drawn from the grid follows
  2. L4: Flexible jobs run in the greener hours
  3. L6: Next day's hourly capacity curves set for flexible work

way out: Generator exhaust and plant noise to the neighbourhood

  1. L1: Nitrogen oxides, fine particles and noise reach neighbours
  2. L5: Backup generators are started (outage, test)

way out: Heat and water vapour rejected to the outside

  1. L1: Heat and evaporated water leave to the air
  2. L3: Chillers, CRAHs, pumps move the heat
  3. L5: Setpoints for chilled water, towers and fans are set

way out: Results back out over the network

  1. L1: Packets over carriers' fibre
  2. L3: The fabric and border routers send results out
  3. L4: Servers compute
  4. L6: Placement and scheduling decide which servers do the work

way out: The load the grid sees (drawn, shifted, or dropped to backup)

  1. L1: The grid sees load drop (and return, or not, until manual reconnection)
  2. L2: UPS, flywheel or generator takes the load
  3. L5: Protection logic decides to ride through or transfer to backup

way out: Waste heat to nearby buildings

  1. L1: Heat into the building or district heating system
  2. L5: Workload placement tuned to raise the heat captured at the cooling coils

Sources (23) published science

Patterson et al. 2021, Carbon Emissions and Large Neural Network Training (arXiv:2104.10350)
Fan X, Weber W-D, Barroso LA (2007). Power Provisioning for a Warehouse-sized Computer. ISCA '07. doi:10.1145/1250662.1250665
Gao J (2014). Machine Learning Applications for Data Center Optimization. Google white paper (behind its links)
Karamperidou C, Casselman JW, Cleveland SB, et al. (2026). Limitations to air free cooling in data centers under rising heat and humidity. Scientific Reports 16. doi:10.1038/s41598-026-56926-3 (PMC13522486) (behind its links)
NERC (2025). Incident Review: Considering Simultaneous Voltage-Sensitive Load Reductions (Large Load Loss) (behind its links)
Radovanovic A, Koningstein R, Schneider I, et al. (2022). Carbon-Aware Computing for Datacenters. IEEE Transactions on Power Systems. arXiv:2106.11750 (behind its links)
Heering J, Voeten E (2026). How sovereign control, decarbonization and energy costs shape public support for data centers. Nature Communications. doi:10.1038/s41467-026-76500-9 (PMC13550595) (behind its links)
Data center (Wikipedia, via Wikimedia Enterprise snapshot 2026-11-01) (behind its links)
Google data centers (Wikipedia, via Wikimedia Enterprise) (behind its links)
OVHcloud (Wikipedia, via Wikimedia Enterprise) (behind its links)
Mytton D (2021). Data centre water consumption. npj Clean Water 4:11. doi:10.1038/s41545-021-00101-w (behind its links)
Tao Y, Gao P (2025). Global data center expansion and human health: A call for empirical research. Eco-Environment & Health. doi:10.1016/j.eehl.2025.100157 (behind its links)
Shehabi A, Newkirk A, Smith SJ, et al. (2024). 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory LBNL-2001637 (behind its links)
Barroso LA, Hölzle U (2007). The Case for Energy-Proportional Computing. IEEE Computer 40(12):33-37. doi:10.1109/MC.2007.443 (behind its links)
El-Sayed N, Stefanovici I, Amvrosiadis G, Hwang AA, Schroeder B (2012). Temperature Management in Data Centers: Why Some (Might) Like It Hot. SIGMETRICS '12. doi:10.1145/2254756.2254778 (behind its links)
Li P, Yang J, Islam MA, Ren S (2023, v5 2025). Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models. arXiv:2304.03271 (behind its links)
Central Statistics Office Ireland (2024). Data Centres Metered Electricity Consumption 2023 (statistical release, 23 July 2024) (behind its links)
Kuzay M, Demirel E, Yilmaz C, Koirala BP, Heer P (2025). Maximizing waste heat recovery from a building-integrated edge data center. Scientific Reports. doi:10.1038/s41598-025-22498-x (behind its links)
Colocation centre (Wikipedia, via Wikimedia Enterprise) (behind its links)
Schroeder, Pinheiro & Weber, DRAM Errors in the Wild: A Large-Scale Field Study, SIGMETRICS/Performance 2009, doi:10.1145/1555349.1555372 (behind its links)
Power usage effectiveness (Wikipedia, via Wikimedia Enterprise) (behind its links)
Dhakal B, et al. (2026). A Critical Review of Grid Integration Challenges and Flexibility Opportunities for Large-Scale Data Centers. Preprint submitted to Renewable and Sustainable Energy Reviews. arXiv:2610.19374 (behind its links)
Katal A, Dahiya S, Choudhury T (2023). Energy efficiency in cloud computing data centers: a survey on software technologies. Cluster Computing 26(3):1845-1875. doi:10.1007/s10586-022-03713-0 (behind its links)