Nature Positive Role of the Technology Sector 2025

Text generation ~1.3-28.7 WhSummarization ~0.8-34.8 WhImage generation ~187-1,640 Wh Range of single-prompt energy use for specific tasks (in watt-hours): Source: Hugging Face. (n.d.). AI Energy Score. https://huggingface.co/spaces/AIEnergyScore/Leaderboard. Power demand for cooling increases in line with server rack densities. Power use effectiveness (PUE) is a common metric for understanding data centre power demand for cooling. The metric is calculated by dividing the total energy needed to operate a facility by the total energy used for computing, with typical values today around 1.4. In an ideal state, a data centre would have a PUE of 1, meaning 100% of energy used by the facility goes towards computing. Many popular data centre jurisdictions are adopting minimum efficiency regulations.183 While helpful, PUE as a metric must be taken in context. Some organizations are proposing alternative, more holistic metrics for data centre operators to consider, such as the Data Centre Resource Effectiveness metric, or DCRE, from The Green Grid.184 Assessing how day-to-day use of AI contributes to data centre power demand is a new challenge for individuals and businesses. While analysing the energy demands of different AI models can be challenging with proprietary company data, a US company called Hugging Face developed and launched, in partnership with Salesforce, its AI Energy Score project to provide standardized energy ratings across various tasks for open- source and closed AI models submitted. Some tasks and their range of single-prompt energy use are shown in Figure A4. Data centres – single-prompt energy use for specific tasks FIGURE A4 This energy use adds up across the estimated 2 billion+ daily prompts across all AI models.185 With 15 of the top 20 most-used AI models being closed source and these model providers not submitting models for testing, there remains a gap in understanding the real-world implications of AI energy use.186 Finally, to address energy infrastructure constraints, some data centre operators are building captive, behind-the-metre solutions principally to meet their own energy needs as opposed to supplying the grid. While this approach may enable faster development and provide flexibility for grid operators during peak hours, it can further contribute to nature loss depending on the sources used. Water use Water use in data centres is a growing area of concern. Historically, electric-powered air cooling could meet the requirements of data centres with server rack power densities of 20 kW or less.187 As rack densities have increased, direct-to-chip water cooling has become the preferred method given its higher capability.188 Data centre cooling typically involves two steps: server and facility-level cooling (to remove heat from computing equipment) and facility heat rejection (to remove heat from the facility); water may be required across both. Liquid cooling methods can be extremely water intensive, with even a small 1 MW data centre able to consume 25.5 million litres of water annually using evaporative cooling.189 On average, a hyperscale facility can use 2.1 million litres of water per day, while a retail facility may use around 68,000.190 With an average of 60% of water use consumed by evaporation and the remaining 40% going into local wastewater systems, managing water use is a key area for reducing data centres’ impacts on local water supply.191 Water use effectiveness (WUE), which reflects the annual litres of water used for humidification and cooling divided by the total annual kWh used to power IT equipment,192 can be a helpful if imperfect metric for tracking this nature impact. Figure A5 includes various technologies for rejecting heat from data centres and their range of WUE. As with PUE, WUE may not always equate to nature impact. A data centre with a high WUE in a region with low water stress is less concerning than a similar data centre in a region with substantial water stress, but generally a lower WUE figure close to zero is better. Nature Positive: Role of the Technology Sector 60