# Green AI hosting and data centres

> How to choose where AI runs: grid carbon intensity, PUE and WUE, renewable energy claims, water use and siting, with questions to ask hosts.

Source: https://greenai.guide/practice/hosting-and-data-centres/  
Updated: 2026-11-19

## In short

Where AI runs can matter as much as how it runs. Choose regions with low-carbon electricity, ask hosts for PUE, WUE and location-based carbon figures rather than just renewable claims, and remember that a clean data centre does not make wasteful software efficient.

## Key points

- The IEA estimates data centres used about 415 TWh in 2024, around 1.5% of global electricity, and projects about 945 TWh by 2030.
- Grid carbon intensity varies by country, region and hour, so location and timing change the carbon cost of the same work.
- PUE measures building overhead, not useful work; the Uptime Institute's 2025 survey put the average at 1.54, barely changed for six years.
- Renewable claims based on certificates do not change the physical grid; ask for location-based figures too.
- Water matters: ask for WUE and whether it includes water used to generate the site's electricity.

## Why does where AI runs matter?

**Because the same computation can have very different carbon, water and grid impacts depending on the data centre
and the electricity behind it.** Patterson and colleagues found that the choice of data centre, processor and
model together can change the carbon footprint of training by up to around 100 to 1,000 times.

The scale is growing. The International Energy Agency estimates data centres used about 415 TWh of electricity in
2024, around 1.5% of global consumption, and projects that this roughly doubles to about 945 TWh by 2030 in its base
case, with AI the largest driver.

## How do I compare grid carbon intensity?

**Look up the carbon intensity of the grid in each candidate region, for the hours your work will run.** Carbon
intensity, measured in grams of CO₂ equivalent per kilowatt-hour, depends on the mix of generation: hydro, nuclear,
wind and solar are low; coal and gas are high. It varies between countries, between regions within them, and from
hour to hour.

- **Data sources:** Electricity Maps covers many countries; in Great Britain, the National Energy System Operator's
  Carbon Intensity API gives forecasts for 14 regions.
- **Use location-based figures** for decisions about your software. The Green Software Foundation's SCI
  specification uses location-based grid intensity and excludes certificates and offsets.
- **Match the time.** Annual averages hide the difference between a windy night and a still winter evening.
  Flexible jobs can be moved to cleaner hours; see
  [efficient training and inference](/practice/efficient-training-and-inference/).

## What do PUE, WUE and other metrics tell me?

**Each metric answers one narrow question; ask for several, and know what each leaves out.**

| Metric | What it measures | What it doesn't tell you |
|---|---|---|
| [PUE](/glossary/#pue) (power usage effectiveness) | Total facility energy ÷ IT equipment energy; 1.0 is the minimum | Whether the servers do useful work; carbon; water |
| WUE (water usage effectiveness) | Litres of water per kWh of IT energy; "site" or "source" (including water used to generate the electricity) | Local water stress |
| CUE (carbon usage effectiveness) | Carbon per kWh of IT energy | Efficiency of the software |
| ERE / ERF (energy reuse) | Benefit of reusing waste heat | Anything if folded into PUE instead |
| Server utilisation | How busy the IT equipment is | Carbon intensity |

Two warnings:

- **PUE has stalled.** The Uptime Institute's 2025 survey reported a weighted average PUE of 1.54, the sixth year in
  which it barely changed. And PUE can look excellent while servers sit nearly idle.
- **PUE cannot go below 1.0.** If a host claims a figure under 1.0 because it reuses heat, ask for heat reuse as a
  separate ERE or ERF figure, as The Green Grid intends.

In the EU, data centres with at least 500 kW of installed IT power must report energy, PUE, WUE, energy reuse and
renewable energy figures each year to a European database under Delegated Regulation 2024/1364. These reports are a
useful source when you compare EU hosts.

## Can I trust "100% renewable" claims?

**Treat them as a starting point and ask how they are made.** Most rest on certificates or power purchase
agreements matched over a year, not on the electricity the site draws at a given hour.

- The GHG Protocol's Scope 2 guidance distinguishes **location-based** emissions (from the local grid average) and
  **market-based** emissions (reflecting contracts and certificates), and asks companies with access to such
  contracts to report both.
- The Green Web Foundation advises buyers to ask AI vendors for both figures, since they can differ significantly.
- Ask whether renewable supply is **additional** (new generation that would not otherwise exist), **local** to the
  grid the site uses, and **matched hourly** rather than annually.
- Clean electricity is limited. Wasteful software on a "green" host still uses power that others could have used,
  and still needs hardware. A green host is a complement to efficiency, not a replacement.

## How much water does AI use?

**It depends heavily on cooling design, climate and the electricity source, so ask for figures per site.**

- Google reports 0.26 mL of water for its median Gemini text prompt in 2025, by its own accounting.
- Li and colleagues project that global AI demand may account for 4.2 to 6.6 billion cubic metres of water
  withdrawal in 2027, counting on-site cooling and water used to generate electricity.
- Water is local. A litre in a water-stressed region matters more than a litre where water is plentiful. Ask where
  sites are and how water is sourced.

## Where should new AI capacity go?

**Where clean energy is plentiful, water is not scarce and waste heat can be used.** A 2026 UK green AI sandpit of
mathematicians, computer scientists and industry discussed siting data centres where renewable electricity is
regularly curtailed because the grid cannot take it, so that surplus energy is used rather than wasted. Siting
also involves grid capacity, planning, local communities and data sovereignty; see
[sovereign AI and sustainability](/articles/sovereign-ai-and-sustainability/).

For your own workloads:

- put latency-tolerant work (training, batch inference, evaluation) in the lowest-carbon region that meets your
  legal and security requirements;
- deploy interactive services close to users, preferring the cleaner of equivalent regions;
- allow for the energy and cost of moving large datasets.

## Checklist

- [ ] Grid carbon intensity compared for candidate regions, by hour if possible
- [ ] PUE, WUE (site and source), CUE and heat-reuse figures requested
- [ ] Location-based and market-based emissions both requested
- [ ] Renewable claims checked for additionality, locality and hourly matching
- [ ] Water figures and local water stress considered
- [ ] Flexible workloads placed in low-carbon regions and hours
- [ ] Efficiency work continuing regardless of how green the host is

## What to do next

1. List where your AI workloads run today, including those run by suppliers.
2. Look up each region's grid carbon intensity and note which jobs could move.
3. Send your hosts the questions above; use [sustainable AI procurement](/practice/sustainable-ai-procurement/) for
   contract wording.
4. Put grid carbon intensity on your operations dashboard next to load and cost.
5. Decide what runs centrally and what could run locally; see [edge or cloud](/practice/edge-or-cloud/).
6. Before describing your hosting as green, read [honest green claims](/practice/honest-green-claims/).

## Frequently asked questions

### What is a good PUE?

PUE is total facility energy divided by IT equipment energy, so 1.0 is the theoretical minimum. The Uptime Institute's 2025 survey reported an industry average of 1.54.

### Does a data centre on 100% renewable energy mean my AI is zero carbon?

Not necessarily. Such claims usually rest on annual certificates or contracts. At a given hour the site still draws from the local grid, so ask for location-based emissions and, ideally, hourly matching data.

### How much water do data centres use?

It varies hugely with cooling design and climate. Google reports 0.26 mL of water for a median Gemini text prompt; researchers project global AI water withdrawal of 4.2 to 6.6 billion cubic metres in 2027.

### Should I always pick the greenest region?

For flexible, latency-tolerant work, often yes. For interactive services, weigh carbon against latency, data-protection rules, resilience and the cost of moving data.

## Sources

- [Energy and AI](https://www.iea.org/reports/energy-and-ai), International Energy Agency (2025)
- [Uptime Institute Global Data Center Survey 2025](https://intelligence.uptimeinstitute.com/resource/uptime-institute-global-data-center-survey-2025), Uptime Institute (2025)
- [Water Usage Effectiveness (WUE): a Green Grid data center sustainability metric (White Paper 35)](https://www.thegreengrid.org/system/files/store/WUE_v1.pdf), The Green Grid (2011)
- [ERE: a metric for measuring the benefit of reuse energy from a data center](https://datacenters.lbl.gov/sites/default/files/EREmetric_GreenGrid.pdf), The Green Grid (2010)
- [Measuring the environmental impact of delivering AI at Google scale](https://arxiv.org/abs/2508.15734), Google (arXiv 2508.15734) (2025)
- [Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models](https://arxiv.org/abs/2304.03271), Li, Yang, Islam and Ren (arXiv 2304.03271) (2023)
- [GHG Protocol Scope 2 Guidance](https://ghgprotocol.org/sites/default/files/2023-03/Scope%202%20Guidance.pdf), GHG Protocol (2015)
- [Software Carbon Intensity (SCI) specification](https://sci.greensoftware.foundation/), Green Software Foundation (2024)
- [Commission Delegated Regulation (EU) 2024/1364 on the first phase of a common Union rating scheme for data centres](https://eur-lex.europa.eu/eli/reg/2024/1364/oj), EUR-Lex (2024)
- [Carbon Intensity API](https://carbonintensity.org.uk/), National Energy System Operator (2025)
- [Carbon Emissions and Large Neural Network Training](https://arxiv.org/abs/2104.10350), Patterson et al. (arXiv 2104.10350) (2021)
- [Thinking about using AI? Here's what you can and (probably) can't change about its environmental impact](https://www.thegreenwebfoundation.org/publications/report-ai-environmental-impact/), Green Web Foundation (2024)
