Policy paper

Compute Evidence Annex: changes corrected (23 April 2026)

Updated 23 April 2026

Breakdown of the changes published in Compute Analyticial Annex PDF document on 20 April 2026. The changes were made to the original content published on 17 July 2025.

Changes made

Within the executive summary:

• Rising demand is partly balanced by more efficient models and hardware. Energy use per ChatGPT prompt has fallen by a factor of 10 since 2023. Considering this alongside potential increases in AI demand and the UK’s ambitious targets for electricity grid decarbonisations, the UK’s cumulative 10 year greenhouse gas emissions from AI compute could range from 34 to 123 MtCO₂ – this is around 0.9-3.4% of the UK’s projected total emissions over the 10 year period. If successful, the UK’s grid decarbonisation plans would help to reduce emissions from data centres towards the bottom end of this range.

Within the environmental impacts chapter:

• Using DSIT’s AI environmental impacts model (Annex C), we estimate that UK greenhouse gas emissions from AI compute over the 10 years from 2025 to 2035 could range from 34 to 123 MtCO₂ (Figure 7). This is around 0.9-3.4% of the UK’s projected total emissions over the 10 year period. This is driven by indirect emissions from the power generation supplying and constructing electricity-intensive data centres and largely depends on how quickly the UK decarbonises its energy grid and how fast AI adoption grows. The current government plan is that by 2030, the UK power system will see clean sources produce at least 95% of Great Britain’s generation, which, if successful, would mean emissions from AI compute would be towards the bottom of this range. Figure 8 shows that these emissions are expected to be falling by the end of this period as the electricity system continues to decarbonise.

AI data centres also place heavy demands on water, primarily for cooling: 

• Up to 40% of a data centre’s total energy consumption can be for cooling. Factors like increased energy requirements for GPUs, higher density server arrangements, and AI processes running continuously makes it more of a challenge for AI data centres specifically. This has spurred investment into new cooling solutions, such as using liquid immersion cooling where servers are submerged in coolant and can reduce energy use by 30% compared to traditional air cooling. However, a recent study found that training GPT-3 in Microsoft US data centres (which are more efficient) consumed around 700,000 litres of water. Projected annual water consumption from AI compute in the UK in 2035 could range between 0.1–0.5 trillion cubic metres depending on demand and decarbonisation assumptions (Figure 9). This is primarily driven by direct water consumption from cooling compared to indirect water consumption used in electricity generation to power data centres.

In Annex C on methodology:

• DSIT commissioned Cambridge Econometrics to support our assessment the environmental impacts arising from the operation of AI systems, with projections extending to 2035. This involved a Cambridge Econometrics secondee working within the DSIT team on developing key aspects of the modelling, but with DSIT maintaining overall responsibility for the outputs.

• Direct environmental impacts include the consequences from the development and operation of AI systems. While indirect environmental impacts refer to those associated with the construction and production of essential AI hardware and datacentres. The model does not capture end‑of‑life emissions from hardware disposal or upstream impacts from critical mineral extraction, reflecting current data limitations.

• Compute demand projections are taken directly from the McKinsey Compute Demand and Supply model, while environmental coefficients and future trajectories are sourced from a combination of academic literature, grey literature, and government projections. Where future trajectories were not well evidenced in the literature, scenario‑specific assumptions were used to inform projections.

• A key input to the AI Environmental Impacts Model is the number of GPUs needed to meet a specific level of AI demand. This is then converted into electricity consumption based on the assumptions within the McKinsey Compute Demand and Supply model. The electricity demand trajectories for the four scenarios are shown in Figure 10. The High, Medium, and Low scenarios combine high, medium, and low assumptions for both GPU and non-GPU electricity demand. The High with decarbonisation scenario combines high GPU electricity demand with lower non-GPU electricity demand, reflecting assumptions about more efficient cooling and networking technology in that scenario.

• Total electricity demand is then multiplied by a series of quantitative environmental coefficients, including the carbon emissions intensity of UK electricity generation, to assess the size of direct and indirect environmental impacts in terms of GHG emissions, water consumption, and material and land use. As a key driver of the modelled outputs, the scenario assumptions for the emissions intensity of UK electricity generation are described below and shown from 2025-2035 in Figure 11:
  
  

  • Medium scenario: This scenario uses the most recent emissions factors published by the Department for Energy Security and Net Zero for use in UK government policy development.
    
  • Low scenario: Assumes the lowest compute demand projection and the “CCC Balanced Pathway” emissions intensity assumptions. In this scenario, the share of renewables in UK energy production follows the path projected in the CCC 6th Carbon Budget, which is broadly in line with the UK government’s current target to reduce grid carbon intensity to below 50gCO2e/kWh by 2030.
    
  • High scenario: Assumes the highest compute demand projection and that the emissions intensity projections follow the NESO’s “Falling Short” scenario, where the UK will not meet Net Zero by 2050.
    
  • The high with decarbonisation scenario: Combines the high compute demand projections with the low emissions intensity assumptions from the “CCC Balanced Pathway” scenario

Why changes were needed to the original content

The original content (published on 17 July 2025) referring to DSIT’s environmental impacts model has been updated to reflect new analysis. We keep analysis under routine review to ensure it reflects the most up to date assumptions and analysis.

Content has been updated the following places in the original document:

• bullet point 6, end of page 3 into page 4 
• end of page 15
• top of page 16