HOUSE OF THE PEOPLE · CITIZENS ASSEMBLY ON AI

What happens in a single second? References

The sources behind the figures used in the booklet about the data centres.

The booklet was written by Assemble, the organisation that convenes the House of the People, for members of its citizens' assembly on AI. Its wording comes from the third session of the assembly's learning phase on 16 September 2026: the talk "What happens in one second?" and the handout cards members read in small groups.

PAGE 2What goes into that second?

Since 1980, mining companies have produced four times as much of many rare metals as in all of history before then.

From Anatomy of an AI System, which gives no source of its own for this figure. Production records show the same steep rise: world rare-earth output was a few tens of thousands of tonnes a year before the 1970s and is now around 400,000 tonnes a year.

Just 1% of the demand for rare-earth metals is met by recycling old electronics.

The UN also found that only about 22% of electronic waste was properly collected and recycled in 2022.

In cobalt mines in the Congo, some workers are paid the equivalent of one US dollar a day, and children as young as 7 have been found working.

Amnesty International documented children as young as 7 in cobalt mines in the Democratic Republic of the Congo. More than 40 million people worldwide work in small-scale, often hand-dug mining.

A data centre can be the size of several football pitches.

The largest data centres cover many hectares; many in the UK are much smaller.

PAGE 3How much water does AI use?

A single typical AI question uses about 0.26 ml of water, roughly a twentieth of a teaspoon. Google self-reports this figure for a typical text question on Gemini.

This is Google's own figure for a typical (median) text question to its Gemini model. OpenAI says about 0.32 ml for an average ChatGPT question but has not published how it worked this out. Both are the companies' own figures.

AI data centres worldwide could need about 1 trillion litres of water a year by 2028.

A projection by the bank Morgan Stanley, with a range of roughly 0.6 to 1.5 trillion litres. The research note itself is not public, so we link to reporting of it.

The small number leaves out model training, complex prompts and AI agents; the big number includes water use in making the electricity and the chips.

Google's study counts water used to cool the data centre while answering a prompt. The projection covers all AI data centre water use, including power generation.

Data centres take about 5 million litres a day from England's public water supply: under 0.04% of it.

From a 2026 study for the water market regulator MOSL. The 0.04% is our own sum: 5 million out of about 14 billion litres a day. Water UK's evidence to Parliament in April 2026 gives a slightly higher figure, about 6.6 million litres a day. There is no general public reporting requirement for data centre water use; see the note on page 4.

Almost 3 billion litres a day are lost through leaky pipes in England and Wales.

Water companies' own figures for 2023/24, which add up to more than a trillion litres a year. This figure covers England and Wales; the supply and data centre figures are for England.

Data centres could use up to 70 million litres a day by 2030 (worst case).

Water UK, which represents the water companies, told Parliament in April 2026 that data centres' peak demand could reach about 42 million litres a day by 2030, and about 77 million in a worst case where new sites use water like the least efficient ones today. The booklet rounds this down to 70 million. On an average day, use would be about half the peak. The Thames Water plans first listed here as the source do not contain this figure.

The shortfall forecast does not fully account for new data centres.

Water companies' long-term plans, which feed into the national picture, were drawn up before the recent growth in data centres. Thames Water says its plan included only an early estimate for data centre demand, and that it is now working with data centre operators on better figures.

PAGE 4How much electricity do data centres use?

Putting cars and heating on the grid means the UK will need several times more clean electricity than today.

The Climate Change Committee expects electricity demand to rise from 274 to 692 terawatt hours by 2050, more than double, with almost all of it from clean sources. Today a large share still comes from gas, so clean supply has to grow by more than demand does. The National Energy System Operator's 2025 scenarios have demand roughly tripling. Both forecasts already include some growth in data centres. The four times figure shown in Session 3 came from the Committee's 2019 advice.

In the United States, data centres accounted for about half the growth in electricity use in 2025.

In Slough, data centres used 65% of all the electricity in 2024: almost twice as much as everything else in the town.

Data centres used 1.3 TWh, against about 0.7 TWh for Slough's roughly 52,000 households, shops, schools and other businesses. Slough has 29% of Britain's data centre electricity use; London and the South East together have 77%. These are new official statistics and are likely to be underestimates.

Slough's data centres use 29% of Britain's data centre electricity use, running on 416 megawatts (MW) of capacity in 2025.

Figures from the data centre market analysts DC Byte, supplied by the landowner SEGRO and published by Slough Borough Council. The figures also show capacity up from 115 MW in 2017. Capacity is the most the buildings are set up to draw, not what they use all the time. The 2025 figure covers part of the year.

Big projects can use a national fast-track planning route where developers no longer have to consult local communities before applying. This has been the case since July 2026.

Data centre developers can ask to use this route, and ministers decide. There is no set size threshold, and only a few have used it so far. Most still go through local council planning, where the public can comment on applications.

About 500 data centres exist in the UK, the second highest in the world by one count, and almost 200 more are planned.

The ranking depends on the source. Statista (November 2025) counts 499 UK sites, second after the United States (4,165) and just ahead of Germany (487). Cloudscene (2025, used in Stanford's 2026 AI Index) counts 523, third behind the United States (5,427) and Germany (529). These count sites, not capacity. Counts also depend on what is counted as a data centre: the government's own statistics matched over 2,400 sites, including small ones. In September 2026, 173 schemes were in planning and 24 more had been announced.

No law in England makes data centres report their water use publicly. Mains water is metered, but private sources are not counted.

Water companies know how much mains water they supply and bill (about 5 million litres a day to data centres in England). Some companies report voluntarily. The government said in June 2025 that there is currently no requirement for data centres to report water use. Water from the mains is metered and billed, so water companies know it, but that is not public reporting. Water from private abstraction, rainwater and non-drinking sources, and the water used in generating a site's electricity, may not be captured. Licensed abstractors report to the Environment Agency under their licence, but that is not specific to data centres. Data centres in the EU, including Ireland, must report water use under EU rules; UK sites do not. The 5 million litres a day is public supply to non-household customers in England only, worked out from an annual figure, and six sites account for about two-thirds of it. A government "Data Centre Covenant" published on 28 October 2026 is reported to ask the largest sites to report water use annually, as a condition of government support rather than a law. We have only read press summaries of it, so check the Treasury text before relying on it.

A typical AI chip trains new models for up to two years, does general work until about six, then is used secondhand, then landfilled.

Google writes off its servers over six years, and Amazon cut some of its own from six years to five in 2025, citing faster progress in AI. The analysts Bernstein say AI chips can run profitably for about six years. Critics, including the investor Michael Burry, argue that their real working life is closer to two or three years. The companies disagree on how long chips last.

The four stages (0 to 2 years training, 2 to 6 years general use, 6 to 10 years secondhand, over 10 years landfill) come from Session 3 and are a simplified model, not a measured statistic. The two to six years is mostly how fast companies write chips off in their accounts, not how long a chip physically works, and firms choose it, which is why critics say it flatters profits. Chips are usually moved from training to cheaper jobs rather than thrown away. Some older chips are still rented out years later: CoreWeave says it has an A100 contract running into 2029, and used A100 and H100 servers are being sold on. That evidence comes mostly from interested parties and could change if demand slows. We found no good data on how much retired data-centre equipment is resold, recycled or landfilled, or support for the over-10-years stage.

Only about a fifth of the world's electronic waste is properly recycled.

The UN found that 22% of electronic waste was properly collected and recycled in 2022.

The planned data centres would be worth around £100bn if all were built, but only about £5bn is built or under way.

From construction data firm Barbour ABI, August 2026. About £2.6bn of data centres have been completed and about £2.3bn are on site. Another £12.7bn have full planning permission, £24.5bn outline permission, £40.6bn are planning applications and £23.2bn are at the pre-application stage. This is the value of planned building, not money committed. It is also separate from the government's claim of over £100bn private investment in the whole AI sector.