How a data centre works

Loading the model, about 2 MB

How a data centre works

A guided tour

How a data centre works

Every time you load a page, send a message or ask an AI a question, a computer in a building like this one does the work. Almost nobody gets inside. This is a composite of a new one, built from the published standards, which you can scroll through and then walk.

A warehouse full of computers

The phrase belongs to Google's engineers, who wrote that the computer that matters now "no longer resembles a pizza box or a refrigerator, but a warehouse full of computers".1 A data centre is a building that keeps thousands of those computers fed with electricity, cool air and fibre, and keeps them running through anything: a power cut, a heatwave, a digger through the cable.

At the end of 2025 there were 1,360 of the very largest, run by the cloud companies, holding about half the world's computing capacity, with almost 800 more in the pipeline.2 They are fenced, guarded, watched by thermal cameras and entered through iris scans,3 so this one is nobody's in particular. Every dimension in it comes from a published standard or an operator's own document, and the notes at the end say which.

How big it is

The white box is 112 metres long and 58 wide. A football pitch for international matches is between 100 and 110 metres long,4 so the building is a pitch and a bit, and its two halls between them would hold about 1,200 racks. That makes it a modest one. Google's site at Waltham Cross in Hertfordshire is 33 acres;5 Microsoft's consented buildings at Skelton Grange in Leeds run to about 500,000 square feet across three halls;6 and the four buildings Google has proposed at Thurrock would have up to 130,500 square metres of floor space and stand up to 26 metres tall.7 The biggest campuses are many boxes like this side by side.

Most of the site is not building. It is the substation, the generator yard, the chiller yard and the fence. That is the first thing to understand about a data centre: the computers are a minority of what is there.

Journey one: a request

It starts in the street

Say you tap a headline on your phone. The phone asks the network where that website lives, gets back an address, and sends off a request for the page.8 The request leaves as radio, becomes light in a glass fibre somewhere near your home, and travels at about 200,000 kilometres a second, two-thirds of the speed of light in a vacuum.9

Fibres reach this building underground, from a manhole in the road, and by a second route from the far side of the site, so that one digger cannot cut both.

Journey one: a request

The meet-me room

The fibres surface in a small locked room near the front of the building, the meet-me room, where the telecoms companies' cables meet the building's own. Big carrier hotels have two of them, for the same reason there are two fibre routes.10

Most of the world's internet traffic changes hands in rooms like this, at exchanges where networks plug into one another. The London Internet Exchange has around 900 members and carried a record 12 terabits a second during January's football;11 Frankfurt's DE-CIX set a record of 29.5 terabits a second during this summer's World Cup.12

Journey one: a request

The spine

From the meet-me room the fibre runs overhead, in a yellow tray, into the hall and to a row of network cabinets: the spine. Modern halls are wired so that any rack can talk to any other at full speed. Meta's engineers described the ambition as making "the entire data center building one high-performance network",13 and Google's as building networks the size of a building out of ordinary switch chips, with more than a petabit a second of bandwidth across the middle.14 The spine is where your request is handed to the row it is looking for.

Journey one: a request

Down the cold aisle

The yellow tray carries the fibre along the front of the racks. Racks are the wardrobes. A standard cabinet is 600 mm wide, 1,200 mm deep and two metres or so tall,15 and holds its equipment in a 19-inch frame in slices of 44.45 mm called rack units, or U.16 The floor tiles are 600 mm squares,17 which is why everything in a data centre lines up: a rack is one tile wide and two deep, and this aisle is three tiles across. The standards ask for at least a metre in front of a rack and prefer 1.2.18

Journey one: a request

The top of the rack

At the top of the cabinet sits a switch. Every server in the rack plugs into it, and it plugs up into the spine. Meta's 2014 design gave each of these top-of-rack switches four 40-gigabit uplinks for a rack of 10-gigabit servers,13 and the numbers have grown several times since. Your request comes down the fibre into the switch, then along a short cable to one machine.

Journey one: a request

The server

This is the machine that answers you. A two-unit server is about 48 centimetres wide, 9 tall and 77 deep,19 and inside are two processors under their heatsinks, rows of memory either side of them, a row of fans behind the front bezel and two power supplies at the back. It finds the page, assembles the response and sends it back the way it came, as a stream of small packets.8

Between London and New York the round trip takes about 56 milliseconds in theory and around 70 in practice;9, 20 within Europe, about 14.20 Everything else in this building exists so that this box can do that, all day, without stopping.

Journey two: the power

From the grid

The second journey starts at the pylons. A large data centre connects to the high-voltage grid rather than the local network, and the connection is the hard part of building one. National Grid's new substation at Uxbridge Moor, a 400 kV and a 132 kV substation side by side, is being built partly to serve requests from data centres for around 1.8 gigawatts of new capacity, connecting more than a dozen of them.21 The International Energy Agency says waiting times for a grid connection "can be as long as five to ten years in many jurisdictions".22

Journey two: the power

The substation

The lines land in a fenced yard, where transformers step the voltage down for the building. From here on everything comes in pairs. Engineers write redundancy as N plus something, where N is the equipment you need to carry the load:23 most designs are N+1 or N+2, and operators have been drifting from the first to the second.24 The Uptime Institute's tiers grade the result. Tier III means any component can be taken out for maintenance without stopping the computers; Tier IV means a single failure does not stop them either.25

Journey two: the power

Switchgear

Inside, the switchgear room is rows of steel cabinets that divide the supply and can isolate any part of it. It is dull to look at and the most important room in the building for keeping the lights on. Some operators go further and run their racks on 48-volt direct current: Google told the Open Compute Project in 2016 that feeding 48 volts to the motherboard and converting once was "at least 30% more energy efficient".26

Journey two: the power

The batteries

The uninterruptible power supply sits between the grid and the computers: rows of battery cabinets and the electronics that switch to them. If the grid fails, the batteries carry the load for the seconds or minutes it takes the generators to start. Uptime describes the runtime as "usually short — worst case, 15 minutes",27 and the rotary kind rides through for 10 to 30 seconds.23 Nearly half of operators now use lithium-ion batteries here.28 It is not an idle safety net: the most common power-related outages come from faults in UPS systems, transfer switches and generators themselves.29

Journey two: the power

The generators

Beyond the west wall stand six diesel generators in containers, each with a fuel tank in its belly, and a bulk tank beside them. The Uptime standard requires at least 12 hours of fuel at full load on site,30 and they are run regularly to prove they will start: US rules allow emergency generators up to 100 hours a year of testing.31 Uptime's verdict in 2019 was blunt: "Few — very few — can even contemplate operating production environments without diesel generators."32

Journey two: the power

The busway

Back inside, the power reaches the hall along the amber busways: solid conductors in a casing, hung over each row, with tap-off boxes every few racks. Two run over every row, one from each side of the electrical plant, so a rack keeps its second feed when the first is lost. They are the reason there are no cables trailing across the floor.

Journey two: the power

Into the server

Each rack has its own power strips, and each server has two power supplies, one plugged into each feed. Here the electricity ends its journey, at a few hundred watts per machine. Add the racks up and the sums get large. The typical rack in Uptime's 2025 survey drew almost 9 kW,33 and if every standard rack in this hall drew that, the 560 of them would need about 5 MW. A two-building data centre approved on appeal in Hertfordshire in 2025 was for 96 MW.34

Journey three: the heat

Where the heat begins

Almost all the electricity that reaches a processor leaves it as heat, and the third journey is about getting rid of it. The heatsink takes the heat into the air; the fans push the air out of the back of the server. A processor can now draw 300 to 400 watts, and at that point ordinary air-cooled servers "become more difficult", in the words of the engineers' body ASHRAE; a 50 kW rack can spend 5 kW just on its fans.35

Journey three: the heat

The hot aisle

The racks stand back to back, so all the exhaust blows into one aisle, and that aisle is closed in: doors at the ends, clear panels above, up to the ceiling. The point is that hot and cold air never mix. The US Department of Energy's design guide describes the contained hot aisle as becoming simply "a heat exhaust", returning air "often 85°F or higher", about 29°C, with a rise across a server of 10 to more than 40°F.36 Schneider's engineers put the saving from containing the hot aisle rather than the cold at 43% of annual cooling energy.37 It is warm in here, on purpose.

Journey three: the heat

Above the ceiling

Lift the lid. The hot air rises out of the top of each contained aisle into the void above the ceiling, the return plenum, and drifts back towards the north wall. Nothing about this is exotic. It is a chimney.

Older halls pushed cold air up through perforated tiles from under a raised floor, and most existing data centres still work that way.38 New halls increasingly deliver air from the side or above, which the energy department notes reduces the layering of temperatures in the cold aisles.36

Journey three: the heat

The fan wall

The air arrives at the fan wall: a bank of large fans and cooling coils built into the north side of the hall. Chilled water runs through the coils; the air gives up its heat to the water and comes back into the hall cold. How cold is a matter of standards. ASHRAE recommends server intake air between 18 and 27°C, and allows 15 to 32 for ordinary equipment,39 which is why a modern hall is not the fridge people imagine. Google runs cold aisles at up to 27°C and reported "no detrimental effect on equipment".40

Journey three: the heat

Chilled water

The water goes out to the yard. The chillers here are air-cooled: fans draw outside air over their coils and the water goes back inside a few degrees colder. The alternative is a cooling tower, which sheds the heat by evaporating water, and that choice is the heart of the water question. Evaporation is the more energy-efficient of the two; air-cooled chillers "use no water" but "use more energy".41 In Uptime's 2025 survey air-cooled chillers were the most common way of rejecting heat, at 35% of facilities, with evaporative towers next at 22%.42

Journey three: the heat

Into the air

On the roof, the dry coolers: more coils, more fans. In a cool climate the outside air is cold enough for much of the year to do the whole job without the chillers, a mode engineers call free cooling; the energy department notes that "many data centers in cool climates use only economizer" cooling.36 This is the end of the third journey. Heat that began in a chip a few metres away leaves as slightly warmer air over Hertfordshire, or wherever this is.

A few operators pipe it somewhere useful instead. Meta's Odense data centre sends 165,000 megawatt-hours of heat a year into the city's district heating, enough for around 9,000 households,43 and Queen Mary University of London heats part of its Mile End campus from 39 racks.44

Journey three: the heat

The row that is different

One pair of rows is taller, wider and plumbed. These are racks for training and running AI models, and air cannot cool them. Nvidia's GB200 NVL72 rack draws about 120 kW by Nvidia's guide and 132 by others',45, 46 roughly fifteen times the survey's typical rack, and most of that heat is carried off by water flowing through cold plates bolted to the chips.45 A coolant distribution unit at the end of the row exchanges that water with the building's chilled-water loop. The sixty racks in these two rows would draw more than the other 560 in the hall put together.

Operators asked when liquid cooling becomes necessary mostly said above 20 kW a rack;42 more than 80% of facilities still have no rack above 30.33

Energy and water

The numbers, carefully

Data centres used about 415 terawatt-hours of electricity in 2024, around 1.5% of the world's total, by the International Energy Agency's estimate, and the agency expects that to roughly double by 2030.47 The same report says there is "substantial uncertainty about both their current and future consumption".47 Water is harder still to pin down, because different people measure different things. The section below sets out what is known, who measures it how, and what is not known.

Walk it

Take control and go where you like: the gate, the cold aisle, inside the hot aisle, the liquid-cooled row, the fan gallery, the battery room, the substation, the roof. Keyboard on a desktop, two joysticks on a phone. Turn on X-ray to see the air moving and the three journeys running.

Hall B is empty on purpose: the second hall of a phased build usually is, for a while.

Energy and water: what the numbers say

Figures about data centres are contested partly because they are hard to gather and partly because people quietly measure different things. This section gives the main published estimates, says who produced them and on what basis, and flags where they cannot be compared.

Electricity

The most-cited global figure is the IEA's: about 415 TWh in 2024, around 1.5% of world electricity, rising in its base case to around 945 TWh, just under 3%, by 2030.47 The agency's own alternative cases span roughly 670 to 1,260 TWh for 2030 and 700 to 1,720 TWh for 2035, and it notes that reporting by operators "remain[s] limited", so even the starting point carries uncertainty.47 Its 2026 update put 2025 consumption at 485 TWh, up 17% in a year, and kept the 2030 figure at about 950 TWh.22

You will also see a larger, older number: 460 TWh in 2022. That comes from the IEA's Electricity 2024 report and covers "data centres, cryptocurrencies, and artificial intelligence" together, with cryptocurrency mining alone accounting for around 110 TWh of it.48 The two figures use different definitions and should not be chained into a growth rate.

National figures vary widely. In the United States, Lawrence Berkeley National Laboratory estimated 176 TWh in 2023, 4.4% of the country's electricity, and a 2028 range of roughly 325 to 580 TWh;41 its 2026 update puts data centres at between 9.5% and 15.3% of US electricity by 2030.49 In Ireland, data centres took 21% of metered electricity in 2023, up from 5% in 2015.50 For the EU, the Commission's research centre estimated 45 to 65 TWh in 2022, 1.8% to 2.6% of the total.51 For Britain the best public figure I could find is a 2022 grid operator note putting data centres at 2.5% of consumption, which also observes that no central authority in Great Britain monitors the sector.52

Efficiency

The industry's standard efficiency measure is PUE, power usage effectiveness: everything the building draws divided by what the computers draw, so 1.0 would mean no overhead at all. Uptime's 2025 survey put the industry average at 1.54, essentially unchanged for six years, with facilities built since 2020 averaging 1.48.33 Google reports 1.09 across its fleet,53 Microsoft 1.17.54 Those two are not comparable with each other or with the survey: Google counts substations and offices in its overhead, Microsoft counts only sites it owns that have run for a year, and PUE says nothing about water or about how efficiently the computers themselves are used.33

Water

Water is where the definitions matter most. The industry metric, WUE, counts litres of water used on site per kilowatt-hour of computing, for cooling and humidification, and deliberately excludes the water used to generate the electricity.55 Operators then report three different quantities: water withdrawn, water discharged back, and water consumed, which is the difference, mostly evaporation.56, 41

MeasureFigureSource and year
Google data centres, water consumed7,787 million US gallons in 2024; 10,523 million in 2025, as electricity use grew 37%Google environmental reports56, 57
Microsoft fleet WUE0.27 litres per kWh in fiscal 2025, down from 2.3 in the early 2000sMicrosoft54, 58
US data centres, direct on-site consumption66 billion litres in 2023LBNL41
US data centres, indirect via electricity generationnearly 800 billion litres in 2023LBNL41

Two things follow. First, most of the water attributable to data centres is used at the power station, not the data centre: a 2021 study put about three-quarters of the US sector's operational water footprint on the indirect side.59 Second, water and energy trade against each other. Evaporative cooling uses water and saves electricity; dry cooling uses no water and more electricity. Google puts the energy saving from water cooling at about 10%,60 LBNL says plainly that "there are tradeoffs between low PUEs and low site WUEs",41 and Google says it chose air cooling for Waltham Cross because of local water stress.56 A low water figure and a low energy figure from the same building are worth a second look.

The grid

The constraint operators talk about most is not electricity in the abstract but the connection. The IEA estimates about a fifth of the data centre capacity projected for 2030 is at risk of delay for lack of grid access, and reports connection queues of five to seven years in the UK.47 Britain's system operator began clearing a queue of more than 700 GW of connection requests in December 2025, about four times what it says the country needs by 2030, with data centres among the demand it kept in the pipeline.61

Reading the figures

Four views

Rendered from the same model in Blender, with the racks placed by the same layout file that places them in the browser.

Aerial view of the composite data centre: a long white building with a substation to the east, generators to the west and a chiller yard to the north, inside a fenced site
The site from the south-east.
A cold aisle in the data hall: dark racks either side, translucent containment above, busways and a yellow fibre tray overhead, the fan wall at the far end
The cold aisle, looking north to the fan wall.
A server pulled out of its rack on rails, showing two processors under heatsinks, memory modules and a row of fans
The server, out on its rails.
The roof and the chiller yard: dry coolers along the roof edge, four chillers on hardstanding, pipes into the building
The roof and the chiller yard.

Notes and sources

Every figure above links to one of these. Each was fetched and read on 7 September 2026; the quotation in the note is from the document itself. Where the only reachable source was secondary, the note says so.

  1. Luiz André Barroso, Jimmy Clidaras and Urs Hölzle, The Datacenter as a Computer: An Introduction to the Design of Warehouse-Scale Machines, second edition, Morgan & Claypool, 2013, via Google Research. https://research.google/pubs/pub41606/
  2. Synergy Research Group, "Hyperscale Operators to Account for 67% of all Data Center Capacity by 2031", 7 April 2026: 1,360 hyperscale data centres at the end of 2025, 48% of worldwide capacity, almost 800 more in the known pipeline. https://www.srgresearch.com/articles/hyperscale-operators-to-account-for-67-of-all-data-center-capacity-by-2031
  3. Google, "Advancing security", Google Data Centers, accessed 7 September 2026. Describes the anti-climb fence, vehicle barriers, thermal cameras, guard patrols, badges and iris scans. https://datacenters.google/advancing-security/
  4. IFAB, Laws of the Game, Law 1, The Field of Play: for international matches the touchline is 100 to 110 m and the goal line 64 to 75 m. https://www.theifab.com/laws/latest/the-field-of-play/
  5. Debbie Weinstein, "Our $1 billion investment in a new UK data centre", Google, 18 January 2024: "The 33-acre site" at Waltham Cross, Hertfordshire. https://blog.google/around-the-globe/google-europe/united-kingdom/google-1-billion-investment-in-a-new-uk-data-centre/
  6. Harworth Group, "Harworth and Microsoft receive resolution to grant planning consent for hyperscale data centre... at Skelton Grange, Leeds", 27 April 2026: "c.500,000 sq ft across three data centre halls". https://harworthgroup.com/2026/04/27/harworth-and-microsoft-receive-resolution-to-grant-planning-consent-for-hyperscale-data-centre-and-il-space-at-skelton-grange-leeds/
  7. Thurrock Data Centre, Design and Access Statement, April 2025 (submitted to Thurrock Council): "up to four data centre buildings ... with a total gross external area (GEA) of up to 130,500 sqm"; "ground level plus up to three storeys, with a maximum height of 26m". https://docs.planning.org.uk/20250625/107/SW5PMMQGMBJ00/6pf6ry2uvn99m7s9.pdf
  8. MDN Web Docs, "How the web works", Mozilla, last modified 20 July 2026. The browser finds the server's address through DNS, sends an HTTP request, and the server returns the files "as a series of small chunks called packets". https://developer.mozilla.org/en-US/docs/Learn_web_development/Getting_started/Web_standards/How_the_web_works
  9. Ilya Grigorik, "Primer on Latency and Bandwidth", High Performance Browser Networking, O'Reilly: "the speed of light in fiber is around 200,000,000 meters per second"; table 1-1 gives a New York to London round trip of 56 ms. https://hpbn.co/primer-on-latency-and-bandwidth/
  10. Telehouse, "Telehouse West Data Centre in London": "Connectivity to 2 meet me rooms offering true diversity directly connectable from each tenant floor". https://www.telehouse.net/data-centre-services/uk/london/telehouse-west/
  11. LINX, "About the London Internet Exchange" (approximately 900 members), and Mark Jackson, "Football Drives London Internet Exchange to Hit 12Tbps of Global Data Traffic", ISPreview, 27 January 2026, reporting LINX's announcement of "a new global internet traffic peak of over 12.069Tbps". https://www.ispreview.co.uk/index.php/2026/01/football-drives-london-internet-exchange-to-hit-12tbps-of-global-data-traffic.html
  12. DE-CIX press release, "World Cup Drives Global DE-CIX Internet Traffic to a New Record High of 29.5 Terabits per Second", 8 July 2026, via Telecom Ramblings. https://newswire.telecomramblings.com/2026/07/world-cup-drives-global-de-cix-internet-traffic-to-a-new-record-high-of-29-5-terabits-per-second/
  13. Alexey Andreyev, "Introducing data center fabric, the next-generation Facebook data center network", Meta Engineering, 14 November 2014: "Each TOR currently has 4 x 40G uplinks, providing 160G total bandwidth capacity for a rack of 10G-connected servers." https://engineering.fb.com/2014/11/14/production-engineering/introducing-data-center-fabric-the-next-generation-facebook-data-center-network/
  14. Arjun Singh et al., "Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google's Datacenter Network", SIGCOMM 2015: "more than 1Pbps of bisection bandwidth". https://research.google/pubs/jupiter-rising-a-decade-of-clos-topologies-and-centralized-control-in-googles-datacenter-network/
  15. Vertiv VR3300 42U cabinet, "2000mm x 600mm x 1200mm (height, width, and depth)", product listing at Dell; Rittal 42U cabinet datasheet, 2000 mm high, 800 mm wide, 1200 mm deep, via Thomas-Krenn. The model's racks are 600 by 1,200 mm and 2.2 m tall, the height of a 48U cabinet. https://www.thomas-krenn.com/en/products/rack-server/technical-datasheet/p.server-cabinet-rittal-42u
  16. The 19-inch rack standard is EIA-310, now CEA/ECIA-310-E (2005). The standard itself is paywalled; the figures here (482.6 mm panel width, 44.45 mm rack unit) are from the Wikipedia article "19-inch rack", which cites it, and a rack manufacturer's summary. https://en.wikipedia.org/wiki/19-inch_rack
  17. Kingspan Access Floors, raised access floor datasheets to EN 12825, September 2021: "Panel Size: 600mm x 600mm". Dell's PowerMax site planning guide gives the US equivalent: "Best practice is to use 24 x 24 inch heavy-duty, concrete-filled steel floor tiles." https://www.raisedfloor.co.uk/wp-content/uploads/2021/09/Kingspan-Raised-Access-Floors-RFS-BS-EN-Datasheets-Sept-2021.pdf
  18. ANSI/TIA/EIA-942, as summarised in Anixter's standards reference guide: "A minimum of 1 m (3 ft) of front clearance shall be provided"; "A front clearance of 1.2 m (4 ft) is preferable"; "Align front or rear of cabinets with edge of floor tiles". The current edition is TIA-942-C (May 2024), which I have not read. https://web.anixter.com/AXECOM/AXEDocLib.nsf/(UnID)/09DF0A00E09C409886257177006611B4/$file/ANSI-TIA-EIA-942.pdf
  19. Dell PowerEdge R760 spec sheet: width 482 mm, height 86.8 mm, depth 772.13 mm with bezel. https://www.delltechnologies.com/asset/en-in/products/servers/technical-support/poweredge-r760-spec-sheet.pdf
  20. Verizon Business, "Monthly IP Latency Data", July 2026: Europe 14.4 ms, trans-Atlantic 70.3 ms, trans-Pacific 111.8 ms round trip on Verizon's network. https://www.verizon.com/business/terms/latency/
  21. Isabella Harris, "Uxbridge Moor work started on largest new substation by Iver", Slough Observer (Local Democracy Reporting Service), 2025, reporting National Grid: "two substations – one 400kV and one 132kV"; data centre requests "will require around 1.8GW of new capacity"; "over a dozen new data centres". National Grid's own press release exists but could not be fetched. https://www.sloughobserver.co.uk/news/25219128.uxbridge-moor-work-started-largest-new-substation-iver/
  22. IEA, Key Questions on Energy and AI, 2026: data centre electricity use "grew by 17% in 2025"; "roughly doubling from 485 TWh in 2025 to 950 TWh in 2030"; "Waiting times for grid connections can be as long as five to ten years in many jurisdictions." https://iea.blob.core.windows.net/assets/3179f7f8-01f6-4dd6-bffa-c9f7b73f1dc9/KeyQuestionsonEnergyandAI.pdf
  23. Kevin Heslin, "Myths and Misconceptions Regarding the Uptime Institute's Tier Certification System", Uptime Institute Journal, 20 July 2017. Defines N as "the number of components that are minimally required to meet the load demand"; notes diesel rotary UPS "ride-through times of between 10-30 seconds" and that the Tier Standard sets no minimum ride-through. https://journal.uptimeinstitute.com/myths-and-misconceptions-regarding-the-uptime-institutes-tier-certification-system/
  24. Andy Lawrence, "Why data center operators are investing in more redundancy", Uptime Institute Journal, 26 October 2020: "designers mostly favoring either N+1 or N+2 configurations" and "a steady three-year shift from N+1 to N+2". https://journal.uptimeinstitute.com/datacenter-operators-redundancy/
  25. Uptime Institute, "Tier Classification System": Tier III is "Concurrently maintainable"; Tier IV adds "Fault tolerance... When equipment fails, IT operations will not be affected." https://uptimeinstitute.com/tiers
  26. John Zipfel, "Google joins Open Compute Project to drive standards in IT infrastructure", Google Cloud, 10 March 2016: the 48 V rack "was at least 30% more energy efficient". https://cloud.google.com/blog/products/compute/google-joins-open-compute-project-to-drive-standards-in-it-infrastructure
  27. Chris Brown, "Lithium Ion Batteries for the data center", Uptime Institute Journal, 11 November 2019: UPS runtime "is usually short — worst case, 15 minutes — before power is restored either by the utility or the engine generator system." https://journal.uptimeinstitute.com/lithium-ion-batteries-in-the-data-center/
  28. Daniel Bizo, "Startups brew new chemistries for fresh battery types", Uptime Institute Journal, 1 September 2021: "Nearly half of operators have adopted the technology [lithium-ion] for their centralized uninterruptible power supply (UPS) plant". https://journal.uptimeinstitute.com/startups-brew-new-chemistries-for-fresh-battery-types/
  29. Jacqueline Davis, "Data center operators will face more grid disturbances", Uptime Institute Journal, 28 June 2023: "The most common power-related outages result from faults with uninterruptible power supply (UPS) systems, transfer switches and generators." https://journal.uptimeinstitute.com/data-center-operators-will-face-more-grid-disturbances/
  30. Kevin Heslin and Keith Klesner, "Fuel System Design and Reliability", Uptime Institute Journal, 18 June 2014: "The Tier Standard: Topology requires this 12-hour fuel storage minimum for all Tiers at 12 hours of runtime at 'N' load". https://journal.uptimeinstitute.com/fuel-system-design-reliability/
  31. Lamont Fortune, "Diesel Exhaust After Treatment for Data Centers", Uptime Institute Journal, 19 May 2014: "The EPA does allow emergency equipment to be run for system maintenance, exercising and testing purposes, up to 100 hours per year." https://journal.uptimeinstitute.com/diesel-exhaust-treatment-data-centers/
  32. Andy Lawrence, "Data centers without diesel generators: The groundwork is being laid", Uptime Institute Journal, 2 December 2019. https://journal.uptimeinstitute.com/data-centers-without-generators-the-groundwork-is-being-laid/
  33. Uptime Institute, Global Data Center Survey 2025, July 2025: "The average of the modal densities in our total sample reaches almost 9 kW"; "More than 80% of the operators responding to our survey say their facility has no racks above 30 kW"; the industry average PUE "had a weighted average of 1.54", "the sixth consecutive year that this headline figure has virtually stood still"; facilities built since about 2020 average 1.48. PUE, the report notes, "excludes important elements of efficiency from its scope, such as facility water use or IT efficiency." https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2025.Annual.Survey.Report.pdf
  34. Appeal decision APP/P1940/W/24/3346061, land off Bedmond Road, Abbots Langley, Ministry of Housing, Communities and Local Government, 12 May 2025: "a data centre of up to 84,000 sqm (GEA) delivered across 2no. buildings" with "the provision of 96 MW of capacity". https://www.slough.gov.uk/downloads/file/5107/cd-7-02-appeal-decision-relating-to-land-off-bedmond-road-abbotts-langley-dated-12-may-2025-reference-3346061-
  35. ASHRAE TC 9.9, "Emergence and Expansion of Liquid Cooling in Mainstream Data Centers", white paper, 2021: "As socket power moves through 300 W toward 400 W, standard 1U and 2U servers become more difficult from an air-cooled standpoint"; "In a 50 kW rack, the fan power translates to be at least 5 kW." https://www.ashrae.org/file%20library/technical%20resources/bookstore/emergence-and-expansion-of-liquid-cooling-in-mainstream-data-centers_wp.pdf
  36. US Department of Energy, Federal Energy Management Program, Best Practices Guide for Energy-Efficient Data Center Design, revised July 2024. Sources for: alternating hot and cold aisles; the contained hot aisle becoming "a heat exhaust" with return air "often 85°F or higher" and a rise across a server "from 10°F to more than 40°F"; overhead air delivery reducing stratification; "many data centers in cool climates use only economizer (air and/or water)"; dry coolers; cold plates and coolant distribution units. https://www.energy.gov/sites/default/files/2024-07/best-practice-guide-data-center-design_0.pdf
  37. John Niemann, Kevin Brown and Victor Avelar, "Impact of Hot and Cold Aisle Containment on Data Center Temperature and Efficiency", APC by Schneider Electric, White Paper 135. A vendor paper: "hot-aisle containment over cold-aisle containment can save 43% in annual cooling system energy cost". https://www.facilitiesnet.com/whitepapers/pdfs/APC_011112.pdf
  38. Magnus Herrlin, "Thermal Guidelines and Temperature Measurements in Data Centers", Lawrence Berkeley National Laboratory, 15 September 2020: the open raised-floor hall with underfloor cooling "represents a vast majority of all data centers". https://datacenters.lbl.gov/sites/default/files/FINAL%20Thermal%20Guidelines%20and%20Temp%20Measurements%209-15-2020.pdf
  39. ASHRAE TC 9.9, "2021 Equipment Thermal Guidelines for Data Processing Environments", reference card from Thermal Guidelines for Data Processing Environments, fifth edition: recommended dry-bulb 18 to 27°C for classes A1 to A4; allowable 15 to 32°C for class A1. https://www.ashrae.org/file%20library/technical%20resources/bookstore/supplemental%20files/therm-gdlns-5th-r-e-refcard.pdf
  40. Google, "Google's Green Data Centers: Network POP Case Study", 2011: "ASHRAE recommends cold aisle temperatures of up to 27°C/81°F, which we've found to have no detrimental effect on equipment." https://static.googleusercontent.com/media/www.google.com/en//corporate/datacenter/dc-best-practices-google.pdf
  41. Arman Shehabi et al., 2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory, December 2024. Sources for: US data centres "reaching 176 TWh by 2023, representing 4.4% of total U.S. electricity consumption"; 2028 range "roughly 325 and 580 TWh"; direct water use, "the 66-billion-liter total" in 2023; indirect water "nearly 800 billion liters"; "While air-cooled chillers use no water, they use more energy"; "there are tradeoffs between low PUEs and low site WUEs"; the definition of consumption as water "permanently removed from the immediate water cycle due to evaporation". https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf
  42. Uptime Institute, "DLC adoption remains slow and steady: Uptime Institute Cooling Systems Survey 2025", Data Report 181, July 2025: air-cooled chillers the most common heat rejection system (35%), evaporative cooling towers 22%, dry coolers 6%; asked at what rack density liquid cooling becomes necessary, most respondents (n=786) chose 20 kW or more. https://intelligence.uptimeinstitute.com/sites/default/files/2025-07/UI%20Field%20181_Data%20center%20cooling.pdf
  43. Meta, "Meta's Odense Data Centre", information sheet, February 2025: "165,000 megawatt hours of free surplus heat from the data centre's server halls is delivered to the local district heating system", for "up to ~9,000 households in Odense". https://datacenters.atmeta.com/wp-content/uploads/2025/02/Metas-Odense-Data-Center.pdf
  44. Queen Mary University of London, "Queen Mary University of London data centre waste heat to provide hot water and heating for campus", 9 December 2024: 39 racks averaging 10 kW, feeding the Mile End district heating network. http://www.qmul.ac.uk/news/latest-news/2024/se/queen-mary-university-of-london-data-centre-waste-heat-to-provide-hot-water-and-heating-for-campus.html
  45. NVIDIA, DGX GB Rack Scale Systems User Guide: Hardware, updated 3 March 2026: "The rack power consumption is approximately 120kW"; 18 compute trays each with two Grace CPUs and four Blackwell GPUs, cooled "through the cold plates that are attached to the CPUs and the GPUs in the tray". https://docs.nvidia.com/dgx/dgxgb200-user-guide/hardware.html
  46. Daniel Bizo, "Density choices for AI training are increasingly complex", Uptime Institute Journal, 25 June 2025: the NVL72 "is rated at 132 kW in a tall but otherwise fairly common 19-inch rack"; "the cold plates handle around 100 kW of thermal power, while the air exhaust exceeds 25 kW"; "1.4 metric tons per rack". https://journal.uptimeinstitute.com/density-choices-for-ai-training-are-increasingly-complex/
  47. IEA, Energy and AI, World Energy Outlook special report, April 2025: "Data centres accounted for around 1.5% of the world's electricity consumption in 2024, or 415 terawatt-hours (TWh)"; "set to more than double to around 945 TWh by 2030"; the four cases span roughly 670 to 1,260 TWh in 2030 and "700 TWh to 1 720 TWh" in 2035; "There is therefore substantial uncertainty about both their current and future consumption"; "around 20% of the projected data centre additions by 2030 in our Base Case could be at risk of delay"; table 2.4 reports connection queues of 5 to 7 years in the United Kingdom, partly from trade press. https://iea.blob.core.windows.net/assets/dd7c2387-2f60-4b60-8c5f-6563b6aa1e4c/EnergyandAI.pdf
  48. IEA, Electricity 2024, revised January 2024: "data centres, cryptocurrencies, and artificial intelligence (AI)" together "consumed about 460 TWh of electricity worldwide in 2022"; "In 2022, cryptocurrencies consumed about 110 TWh". https://iea.blob.core.windows.net/assets/6b2fd954-2017-408e-bf08-952fdd62118a/Electricity2024-Analysisandforecastto2026.pdf
  49. Sarah Smith et al., United States Data Center Energy Usage Report: 2025 Update, Lawrence Berkeley National Laboratory, 18 June 2026: data centres "between 9.5 and 15.3% of total U.S. electricity use by 2030". https://escholarship.org/uc/item/33m6w3x0
  50. Central Statistics Office (Ireland), Data Centres Metered Electricity Consumption 2023, 23 July 2024: "The percentage of total metered electricity consumption accounted for by data centres rose from 5% in 2015 to 21% in 2023." https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2023/keyfindings/
  51. George Kamiya and Paolo Bertoldi, Energy Consumption in Data Centres and Broadband Communication Networks in the EU, European Commission Joint Research Centre, 2024: "Data centres in the EU used an estimated 45–65 TWh of electricity in 2022 (1.8–2.6% of total EU electricity use)". https://publications.jrc.ec.europa.eu/repository/handle/JRC135926
  52. Joe Andrews, "Data Centres", National Grid ESO Future Energy Scenarios note, March 2022: "between 400-600 known commercial data centres in Britain", "accounting for 2.5% of the UK's electricity consumption"; "there isn't currently a central authority in Great Britain monitoring the data centre landscape." https://www.neso.energy/document/246446/download
  53. Google, "Power usage effectiveness", Google Data Centers: "a comprehensive trailing twelve-month (TTM) PUE of 1.09 across all our large-scale data centers". Google's overhead includes substations, transformers, water treatment and offices, which is broader than many operators count. https://datacenters.google/efficiency/
  54. Microsoft, "Measuring energy and water efficiency for Microsoft datacenters": fiscal 2025 fleet PUE 1.17 and WUE 0.27 litres per kWh, counted only for data centres Microsoft owns and controls that had been operating for 12 months. https://datacenters.microsoft.com/sustainability/efficiency/
  55. The Green Grid, "Water Usage Effectiveness (WUE): A Green Grid Data Center Sustainability Metric", White Paper 35, 2011: WUE is annual site water use divided by IT equipment energy, in litres per kilowatt-hour; life-cycle water is excluded "for practical considerations". https://www.thegreengrid.org/system/files/store/WUE_v1.pdf
  56. Google, Environmental Report 2025: data centre water consumption 7,787 million gallons in 2024; "We calculate water consumption by subtracting water discharge from water withdrawal"; air cooling chosen at Waltham Cross because of local water stress. https://www.gstatic.com/gumdrop/sustainability/google-2025-environmental-report.pdf
  57. Google, 2026 Environmental Report: data centre water withdrawal 13,562, discharge 3,039 and consumption 10,523 million gallons in 2025; "our total electricity consumption grew by 37% in 2025, up from 27% in 2024"; average annual PUE 1.09. https://sustainability.google/files/google-2026-environmental-report.pdf
  58. Melanie Nakagawa and colleagues, "Inside Microsoft's two-decade push to cut water intensity while scaling for growth", Microsoft, 24 June 2026: water use effectiveness improved "by nearly 90%", from 2.3 litres per kWh in the early 2000s to 0.27 in 2025. https://blogs.microsoft.com/blog/2026/06/24/inside-microsofts-two-decade-push-to-cut-water-intensity-while-scaling-for-growth/
  59. Md Abu Bakar Siddik, Arman Shehabi and Landon Marston, "The environmental footprint of data centers in the United States", Environmental Research Letters 16, 064017, 21 May 2021: "Roughly three-fourths of US data centers' operational water footprint is from indirect water dependencies." https://iopscience.iop.org/article/10.1088/1748-9326/abfba1
  60. Urs Hölzle, "Our commitment to climate-conscious data center cooling", Google, 21 November 2022: "Water-cooled data centers use about 10% less energy and thus emit roughly 10% less carbon emissions than many air-cooled data centers." https://blog.google/company-news/outreach-and-initiatives/sustainability/our-commitment-to-climate-conscious-data-center-cooling/
  61. NESO, "NESO implements electricity grid connection reforms to unlock investment in Great Britain", 8 December 2025: the old connection queue exceeded 700 GW, "around four times what Great Britain requires" in 2030; 99 GW of transmission-connected demand retained, with data centres among the drivers. https://www.neso.energy/neso-implements-electricity-grid-connection-reforms-unlock-investment-great-britain

About

This is a guided tour of a data centre that does not exist. It is a composite of how new ones are built in the mid-2020s: two data halls either side of a service spine, hot-aisle containment with a fan wall down one side, chilled water to air-cooled chillers and dry coolers, a substation, switchgear, batteries, diesel generators and overhead busways, and one liquid-cooled row of the kind that AI has made necessary. Every dimension is from a published standard or an operator's own document, and the notes above say which. Where I have had to choose between practices, I have said what the alternatives are.

I built it because most people have no idea what a data centre is for or how one works, and the buildings are closed to visitors, so the only way to walk through one is to make it. The model was built in Blender, in conversation with Claude Code, from a single layout file that positions every rack, route and label; the same file drives the web page, so the browser view and the renders cannot drift apart. The page is plain HTML and one JavaScript module on top of three.js, with no build step.

Where the model is a choice, not a fact

How the page works

Scrolling moves the camera between chapters and runs a pulse along whichever journey the chapter belongs to: cyan for the request, amber for the power, coral for the heat. Labels are projected from the model and fade when something is in the way. "Lid off" cuts the building at the ceiling; "Section" cuts it through the hot aisle the journeys pass through; "X-ray" shows the air as particles moving from the cold aisles through the racks and up the chimneys, with all three journeys running; "Scale" draws a football pitch beside the site. "Walk" hands you the camera: WASD or arrow keys, Shift to hurry, F to fly, X for x-ray, L for the lid, and the number keys 1 to 8 to jump between viewpoints. On a phone the two joysticks do the same.

Diary

Cuts, and a better day
The camera used to fly straight through walls between chapters, which was disorientating. Now any move that would cross a wall, or cover more than seventy metres, is a cut: fade to dark, jump, fade back, decided by a ray cast between the two positions when the model loads. Short moves down an aisle are still dollies. The palette was also lifted out of the swamp: fresher grass, a bluer sky, thinner fog, whiter cladding and a little more exposure. The first-load view now keeps the building clear of the title card.
First version
The model, the three journeys, the walk, the energy and water section and the notes, built in a day. The day's rule was that no number goes on the page from memory: three research passes fetched and quoted every source before a word of the text was written, and the figures that could not be verified (the Open Compute rack specification, hall noise levels, the share of new halls on slab floors) were left out rather than guessed. The building's proportions were the easy part; a rack is a rack. The hard part was the water, where every published figure measures something slightly different from every other, so the section says what each one measures rather than picking a winner.

Chris Moran, 2026. Model, page and renders are free to reuse with credit. Quotations are from the sources cited and remain their authors'.