How much power and water does a data centre actually use?
What building AI infrastructure costs the community around it, and what of that is established.
Last checked: 4 September 2026
The United States is roughly two years ahead of Europe in the rollout of AI data centres. That makes the American debate useful: what is being fought out there now is likely to arrive here later. Below is what can be demonstrated, with the source, and where the figures are disputed.
The electricity bill
The assumption that technology companies bear the full cost of their own rollout does not hold. In the United States, demand for grid connection capacity raises the costs borne by network operators, and those costs reach every customer on that same grid through tariffs.
Quite how large that effect is remains a matter of debate. A PolitiFact fact-check notes that the average residential electricity price in the US has risen by around 42 per cent over the past five years, that data centres are not the sole cause, but that they are a significant driver in the regions where the increase has been steepest. Between March 2021 and March 2026, residential tariffs rose by 94 per cent in Washington DC, 74 per cent in Maryland, 73 per cent in Maine and 58 per cent in New York. Figures in circulation citing increases of several hundred per cent generally refer to wholesale prices, not to a household's bill.
The political response has been substantial. In the first six weeks of 2026, more than three hundred bills concerning data centres were introduced across over thirty American states, including separate tariff classes for large consumers and obligations to bear the full cost of grid connection. In March 2026, several technology companies signed a non-binding declaration undertaking to fund the power for their own facilities and to pay for the associated grid investment.
Whether those measures resolve the problem is not established. In Virginia, the regulator approved a separate tariff class for large data centres, with mandatory fourteen-year contracts, alongside an increase of around eleven dollars a month for an average household.
The consequences do not fall evenly. Low-income households and tenants spend a multiple of their income on energy compared with higher income groups, so the same tariff increase weighs more heavily on them.
Water
Cooling requires water, and the upper bound is considerable. Large data centres can use up to five million gallons a day, around nineteen million litres, which has been compared to the consumption of a municipality of ten to fifty thousand inhabitants. For an average facility the figure is lower: the International Energy Agency estimates the consumption of a typical American 100 MW data centre at around two million litres a day, equivalent to some 6,500 households.
The distinction between those two figures matters. Presenting the upper bound as the average overstates the case by a factor of ten.
Two-thirds of the data centres built or under development worldwide since 2022 are located, according to the IEA, in areas already facing water stress. That is where the pressure lies, not everywhere.
Noise
This is where the most misleading figures circulate, and it is worth taking them apart.
Industrial diesel generators can produce up to around 105 decibels. That is a peak value at source, during testing. Measurements taken on residential properties around the data centres complained of in Northern Virginia came out between 40 and 59 decibels, quieter than a conversation at one metre and within local limits.
The problem, then, is not the volume. It is the permanence. Complaints concern a constant hum that continues day and night, with a strong low-frequency component that is difficult to capture with an ordinary decibel meter. As a result, local noise ordinances do not work on it in practice: they were written for loud parties, not for an industrial installation that never stops. Residents report disturbed sleep, headaches and loss of concentration. In Michigan, one complaint led to a federal class action.
What the other side says
There is a defensible counter-argument, and leaving it out weakens the rest.
Critics of the criticism point out that the measured noise levels fall within existing limits, that the water consumption of data centres represents a limited share of the American freshwater supply, and that widely used estimates come out too high because they include evaporation from reservoirs that would occur without data centres in any case. Counting direct consumption alone gives a figure around half of what is usually cited.
We do not settle that discussion here. What is established is that in 2026 a broad majority of American legislators saw sufficient reason to regulate.
Why this sits differently in Europe
The European framework starts from different premises. Where the American discussion centres on the availability of capacity, European regulation sets requirements on how that capacity is generated and on what happens to the waste heat, above certain thresholds.
That difference is not a safeguard. It does mean that costs now becoming visible in the United States are in part priced in beforehand here, rather than passed on to a community afterwards.
For an organisation deciding today where its AI capacity runs, this is relevant for a reason that has little to do with ecology. Regulation introduced after the fact changes the cost structure of existing contracts. The three hundred American bills from early 2026 concern tariff classes, cost allocation and connection terms. Those are the parameters underneath your own price per unit of compute.
What we were unable to confirm
Two claims that recur in this debate we have not been able to substantiate with a source, and they are therefore absent from this article: the number of permanent jobs a hyperscale data centre creates, and the assertion that such facilities are systematically sited in poorer neighbourhoods. Both are plausible and both are repeated. That does not make them demonstrable.