AI data centre footprints get reported in units that don’t line up: terawatt-hours, megawatts, watt-hours per query, litres and square kilometres. A 100 to 750 MW facility, electricity doubling by 2030, water quadrupling by 2028: all big, but hard to compare. The fix is to anchor everything to one typical facility, then work out which numbers matter. For the wider picture, start with the data centre backlash overview.
How much electricity, water, and land does a typical hyperscale data centre actually use? #
A single hyperscale AI facility draws roughly 100 to 750 MW and moves millions of litres of cooling water a day. One gigawatt is roughly the output of one nuclear reactor, enough for about one million homes, so 100 MW covers around 100,000 households. The land adds up too: UN researchers put the global data centre footprint at about 6,900 square kilometres, and Meta’s Hyperion campus in Richland Parish, Louisiana spans 3,650 acres, four times the size of Central Park.
Water runs on a similar scale. Large facilities average about 5 million gallons a day, and Google’s Council Bluffs site withdrew 3.9 million gallons a day and consumed 2.8 million in 2024. Withdrawal is what you take; consumption is what you don’t give back, and evaporative cooling can consume up to 85% of what it withdraws.
Totals are contested because studies count differently. UN researchers put last year’s data centre consumption at 448 TWh and 4.5 trillion litres of water. The trade-off sits in cooling: evaporative systems are water-intensive but energy-efficient, dry and closed-loop systems use less water but more power. AI facilities top the range because the GPU-dense racks that define them need more power and aggressive cooling than legacy facilities.
How does AI data centre energy demand compare with traditional data centres? #
AI is not yet the majority of data centre power. The IEA puts AI-focused facilities at about 155 TWh, roughly 0.5% of global electricity, with traditional facilities consuming about two-thirds of the total.
Conventional racks sit around 20 kW, while AI training racks demand 40 to 100 kW or more. That density forces the shift from air to liquid cooling.
The ongoing load is inference. About 90% of AI’s power use comes from serving live queries. Training is a one-time cost: GPT-3 used about 1.3 GWh, next-generation models 50 to 70 GWh. A typical ChatGPT-style query runs 0.24 to 0.34 Wh, and a long agentic request up to 50 Wh. That’s the Jevons paradox: as each query gets cheaper, usage grows, so totals still climb.
The picture changes again once you split data centres by scale and density rather than workload.
How do hyperscaler facilities differ from enterprise colocation in resource footprint? #
The divide that matters is scale and density. Hyperscale, by IBM’s definition, means 5,000 or more servers, over 10,000 square feet and 100+ MW of power. Colocation is the other end: multi-tenant space leased to many customers, the Equinix and Digital Realty model.
Colocation sites tend to be smaller and air-cooled; hyperscale AI campuses are bespoke builds that connect thousands of GPUs with liquid cooling. Hyperscalers consolidate load in a few rural campuses; colocation clusters in metros for low-latency access. Brookings researchers found hyperscale counties gain telecom jobs while colocation counties often don’t.
For your procurement decisions, a hyperscale campus concentrates grid and water pressure in one community; colocation spreads a smaller footprint across a metro.
How do you assess the energy and water footprint risk of your AI infrastructure providers? #
Ask for per-site numbers, not portfolio averages: PUE, WUE, CUE. PUE divides total facility energy by the energy that reaches the IT equipment; the industry average is 1.8, and the best hyperscale sites report below 1.04. WUE measures litres of water used per kWh delivered to the IT equipment, and CUE applies the same ratio to carbon.
Separate direct from indirect water too: Indirect water, used to generate the electricity, ran about 12 times direct use in the US in 2023.
Geography is the risk amplifier. Data centres use about 5% of US electricity but more than 20% of Ireland’s. The US load is projected to climb toward 12% of national electricity by 2028, from 4.4% in 2023, and Phoenix’s planned facilities could push cooling-water use up 870%. Northern Virginia’s Data Center Alley already hosts nearly 600 facilities, Texas’s ERCOT is tracking roughly 226 GW of large-load requests, and a Chilean court suspended a Google data centre over its 1.9 billion gallon annual water draw. E-waste from short server refresh cycles is a secondary footprint, and the same grid pressure shows up in ratepayer bills.
Those same per-site metrics are what you use to test a provider’s sustainability claims.
What should you look for when evaluating hyperscaler sustainability and community commitments? #
Test every pledge against three things: whether the clean energy is genuinely new supply, when and where it arrives, and whether the commitment binds at the specific site. Renewable power purchase agreements don’t always match load location or hour, and on-site generation is often infeasible at scale.
Water-positive pledges get the same treatment. Google, Microsoft and Meta have committed to replenishing more water than they consume, but the test is whether restoration happens in the same basin. Meta’s goal is to restore more water than it consumes in the watersheds where it operates.
OpenAI’s Stargate site in Saline Township, Michigan is planned at 1.4 GW, and Meta’s Louisiana campus draws more than twice the power of New Orleans. Treat the footprint as a reliability and cost question.
What questions should you ask cloud providers about water consumption and community impact during procurement? #
Procurement questions fall into three areas: water, grid, and the community deal, and they force real answers before you sign.
On water, request site-level WUE, gallons per day, the source, recycled share and a basin-specific replenishment plan. Few operators publish per-facility figures, so portfolio aggregates conceal local harm.
On the grid, ask for the site’s interconnection queue position and who pays for transmission upgrades. ERCOT has flagged about $14.9 billion in planned transmission projects through 2030.
On the community deal, ask about tax incentives, job commitments and any NDAs tied to the site. A review of 31 Virginia municipalities found 80% had non-disclosure agreements with local officials. Microsoft has said it will stop using NDAs with local governments for data centre development. Cross-check every answer against independent data, because community opposition and moratoriums usually trace back to a site deal never disclosed.
An AI data centre is a city-scale consumer of electricity, water and land, and its footprint has to be proven site by site before any pledge is believed. Power draw, cooling water and land are now operational and financial questions. For how that footprint reaches your ratepayer bill, see the Virginia electricity tax and subsidy trade-off; for the wider picture, the full backlash overview pulls the whole cluster together.
Frequently Asked Questions #
Is it true that one ChatGPT query uses a bottle of water?
No, that claim overstates the research. The figure that popularised it measured roughly 500 millilitres of water across a full conversation of 20 to 50 exchanges, not a single prompt. The real footprint varies widely with the cooling system, data centre location and model size, which is why per-query water figures are so easy to misread.
Why do AI data centres need water if they run on electricity?
Mostly for cooling. Servers and GPUs convert electricity into heat, and water removes that heat efficiently. Many facilities use evaporative cooling, where water absorbs heat and evaporates, and some also use water for humidity control or in the power stations that supply the grid. Closed-loop and dry systems reduce water but often use more electricity, so the choice is a trade-off.
What is the difference between water withdrawal and water consumption?
Withdrawal is the total volume a facility takes from a source such as a river or aquifer. Consumption is the portion that is not returned because it evaporates or is otherwise used up. The gap matters because evaporative cooling can consume up to 85% of the water it withdraws, so a facility can report large withdrawals while returning very little to the original basin.
How many homes could a single hyperscale data centre power?
A useful rule of thumb is that one gigawatt, roughly one nuclear reactor, can supply about one million homes. On that basis a 100 MW facility could cover around 100,000 homes and a 750 MW AI campus closer to 750,000. These are approximations, because real demand varies by hour and season, but they show why a single campus now competes with a small city for power.
Do data centres give the water back after they use it?
Not the same water. Evaporative cooling turns a large share of withdrawn water into vapour that leaves the site rather than returning to the river or aquifer. Water-positive pledges from Google, Microsoft and Meta aim to replenish a basin, but the meaningful test is whether restoration happens in the same catchment that supplied the facility, not through credits somewhere else.
Why are hyperscale AI campuses built in rural areas instead of big cities?
They need land, power and willing hosts at a scale cities rarely offer. A campus like Meta’s Hyperion covers 3,650 acres in Richland Parish, Louisiana, and demands hundreds of megawatts. Rural sites provide large parcels, access to transmission and often tax incentives, while metro areas such as Northern Virginia are already congested with hundreds of facilities competing for the same grid.
Does making data centres more efficient actually lower their total energy use?
Not necessarily. This is the Jevons paradox: when efficiency makes AI cheaper to run, usage tends to grow, so total consumption can still rise even as each unit of work uses less energy. Power usage effectiveness keeps improving, but the industry’s overall electricity demand is still projected to double by 2030 because the volume of computing keeps expanding.
Are data centres causing water shortages in places like Phoenix and Texas?
It is more accurate to say they add pressure to basins that are already stressed. Phoenix’s projected cooling-water growth and Texas grid constraints are real, but a facility’s impact depends on its specific source, cooling system and local hydrology. This is why site-level withdrawal and consumption data, rather than a broad regional claim, is the only way to judge actual risk.
What is a good PUE score for a data centre?
PUE, or power usage effectiveness, divides a facility’s total energy use by the energy that reaches the IT equipment. A lower number means less overhead. The industry average sits around 1.8, so a facility at or below that is typical, while the best hyperscale sites report below 1.04. Always ask for a per-site figure, not a portfolio average that hides weaker locations.
Why do data centres replace servers so often, and what happens to the old hardware?
AI hardware advances quickly, so operators refresh servers on short cycles to stay competitive on performance and power. That turnover creates a secondary e-waste footprint, with retired GPUs and racks entering recycling or disposal streams. It is a smaller share of the footprint than electricity and water, but it is increasingly flagged in lifecycle assessments and worth asking providers about.