Updated Sustainable Tools for AI: Frameworks and Best Practices
Artificial intelligence now appears in business software, public services, and everyday operations across UK companies. However, its environmental costs are growing alongside its adoption. Data centers that power AI consume substantial electricity and water, and new measurement tools are helping businesses understand and manage these impacts.
For UK SMEs, this matters for three reasons. First, corporate carbon reporting requirements increasingly cover digital infrastructure and cloud services. Second, procurement teams are facing questions about AI's resource use from clients, investors, and supply chain audits. Third, choosing the wrong model can create unnecessary costs in energy, water, and computing contracts.
Consequently, a new set of tools and frameworks has emerged to measure AI's environmental footprint. These resources help businesses select lighter models, improve governance, and make environmental performance visible to developers and procurement teams. This article explains what these tools do, what the latest data shows, and what UK businesses should consider when deploying AI systems.
Data center electricity demand is rising sharply
AI models run in data centers that require electricity for training, inference, storage, and cooling. According to the International Energy Agency, global data center electricity consumption rose 17% in 2025. Meanwhile, electricity use by AI-focused data centers increased by 50% in the same period.
The IEA projects that total data center electricity demand will rise from approximately 485 terawatt-hours in 2025 to about 950 TWh by 2030. That represents roughly 3% of global electricity demand. A separate IEA assessment estimated that data centers consumed 415 TWh in 2024, or about 1.5% of global electricity, and could reach approximately 945 TWh by 2030.
Individual facilities are also growing in scale. A typical AI-focused data center can consume as much electricity as 100,000 households. The largest facilities under construction could consume around 20 times that amount.
These figures matter for UK businesses in two ways. First, they signal rising costs for cloud services and AI platforms as providers manage growing electricity bills. Second, they indicate that AI use will increasingly appear in carbon reporting and supply chain audits, particularly for companies subject to Streamlined Energy and Carbon Reporting or preparing for PPN 06/21 compliance.
Water stress adds a second layer of risk
Energy accounting can obscure water-related impacts. Cooling systems may consume water directly, while power generation requires substantial water indirectly. Semiconductor production also adds to the resource footprint of AI hardware.
Nearly 40% of identified AI-dedicated data center capacity is located in areas experiencing high or extremely high baseline water stress, according to an IEA assessment. Water may be used directly for cooling and indirectly through electricity generation and semiconductor manufacturing.
The IEA has emphasized that energy and water efficiency must be assessed together. Reducing one can increase the other. For example, some cooling approaches may reduce electricity use but increase water consumption. Water-saving systems may require more electricity.
This trade-off makes location-specific analysis essential. A model's carbon footprint varies according to the electricity mix used by the data center. Its water impact depends on climate, cooling technology, grid composition, and local water stress. Consequently, a company that reports only a global average may miss the most significant local risks.
Tools for estimating energy and emissions from closed models
The Trellis tracker identifies a growing ecosystem of tools intended to make AI's environmental impacts more measurable. These resources address a central problem: companies often know how much they spend on AI services, but not how much energy, water, or carbon is associated with individual prompts, coding sessions, or business workflows.
CLEER, the Closed-model Latent Energy Estimation Range, was developed by Sustainable AI Group. It estimates the energy and carbon impacts of proprietary models whose operators do not publish complete environmental data. The tool uses information from open models to estimate the likely performance of closed systems.
Carbonlog, developed by Cnaught, connects with Anthropic's Claude and estimates emissions generated during coding sessions. The tool is designed to help teams see how AI-assisted development affects power consumption and emissions.
The AI Energy Score is hosted by Hugging Face and developed with Salesforce. This resource compares the electricity use and efficiency of AI models. It is intended to support more informed technical choices and sustainability-focused procurement.
Greenpixie, a UK-based company, offers tools for comparing the emissions associated with leading AI models. Its customers include Mastercard and Unilever. The Tokenomics Foundation, backed by organizations including Accenture and IBM, is working on open approaches to measuring token-related cost, energy, and other economic metrics.
Additionally, the Green Software Foundation has launched 14 courses for business leaders, developers, and policymakers through its Green Software & AI Academy. The program aims to integrate environmental measurement and mitigation into AI and software development from the outset.
Selecting smaller models can reduce environmental impact
One of the most direct ways to reduce AI's environmental impact is to avoid using a large, energy-intensive model when a smaller one can perform the task adequately. The Trellis tracker highlights BetterTech's Frugal AI Switchboard, which catalogs open-source alternatives to frontier models. The resource includes information about associated energy use and carbon emissions for applications such as chatbots, coding, writing, data analysis, and image generation.
Model selection should therefore consider more than accuracy. Relevant criteria include the model's size and computational requirements, the energy required for inference rather than only training, the number of users and expected prompt volume, and whether the task requires a frontier model or can be handled by a smaller system.
Furthermore, organizations should assess the model's ability to operate efficiently on available hardware and the provider's disclosure of energy, emissions, and water data. The IEA notes that electricity use per AI task is declining rapidly as hardware and software become more efficient. However, rising adoption and the growth of energy-intensive applications could offset those efficiency gains.
Efficiency improvements alone may not reduce total environmental impact if organizations substantially increase the number and complexity of AI tasks. Consequently, choosing the right model requires understanding both unit efficiency and total expected demand.
Corporate governance is beginning to incorporate environmental criteria
Some companies are incorporating environmental criteria into AI governance and procurement. According to Trellis, SAP evaluates the expected energy and emissions impacts of new projects and has adopted practices such as using smaller models. Salesforce has added energy-use and emissions information to fact sheets for its machine-learning models, covering both training and operational use.
Okta's sustainability team has worked with engineering and information-technology strategists to develop guidance on when employees should use AI. The company encourages the use of models with lower energy consumption. GitLab has developed AI vendor disclosure guidelines for contracts and created a screening tool to assess environmental questions in requests for proposals.
These examples point toward a broader shift: AI sustainability is moving from an informal technology concern to a procurement, risk-management, and governance issue. For UK businesses, this has practical implications. Procurement teams may need to request environmental data from AI vendors. Sustainability managers may need to incorporate AI into carbon inventories. IT teams may need to track prompt volumes and model usage across departments.
A practical governance process should assess a proposed AI use case before deployment. It should document expected energy, emissions, water, and hardware requirements. It should identify who can require redesign or mitigation. Environmental criteria should be incorporated into AI logs, impact assessments, and provider disclosures.
Grid composition shapes emissions outcomes
The IEA projects that data center emissions will increase as electricity demand grows. Under its projections, data center emissions could reach approximately 350 million tonnes by 2035. The energy mix supplying data centers will strongly influence the outcome.
The IEA projects that renewable generation will meet nearly half of additional data center electricity demand through 2030. Natural gas, coal, and nuclear power also contribute to supply growth. Renewables currently provide approximately 27% of electricity consumed by data centers globally, while natural gas supplies about 26% and nuclear power about 15%.
Therefore, the location and operating practices of data centers matter. The IEA recommends better demand forecasting, improved grid-connection planning, and more flexible operation of data center loads and on-site energy assets. Flexible computing can allow some workloads to shift away from periods of grid stress or regions with more carbon-intensive electricity.
The overall climate impact is complex. Data centers remain a relatively small share of total global emissions, but they are among the fastest-growing sources of electricity demand. Meanwhile, AI applications could help reduce emissions in other sectors. The IEA estimates that widespread adoption of existing AI applications could enable emissions reductions equivalent to approximately 1.4 billion tonnes of carbon dioxide in 2035 in one scenario.
That potential is not automatic. It depends on whether AI applications produce genuine efficiency gains rather than simply adding new demand. For UK businesses, this means weighing the environmental cost of deploying AI against the potential savings it may deliver in energy management, logistics, or operations.
Summary of key data points
- Global data center electricity use increased by 17% in 2025, while AI-focused data centers saw a 50% increase in electricity consumption.
- Data centers consumed approximately 415 TWh of electricity in 2024, representing about 1.5% of global electricity demand.
- Projected data center electricity consumption could reach about 945 to 950 TWh by 2030, representing roughly 3% of global electricity demand.
- A typical AI-focused data center consumes as much electricity as 100,000 households, while the largest facilities under construction could consume around 20 times that amount.
- Nearly 40% of identified AI-dedicated data center capacity is located in areas experiencing high or extremely high baseline water stress.
- Data center emissions could reach approximately 350 million tonnes by 2035 under IEA central projections.
- Renewables currently provide approximately 27% of electricity consumed by data centers globally, while natural gas supplies about 26% and nuclear power about 15%.
- Widespread adoption of existing AI applications could enable emissions reductions equivalent to approximately 1.4 billion tonnes of carbon dioxide in 2035 in one scenario, though this depends on genuine efficiency gains.
Practical steps for UK businesses deploying AI
A credible sustainable AI program should begin with measurement and then connect measurement to operational decisions. First, inventory where AI is already being used across products, operations, and employee workflows. Record model type, task, prompt volume, inference frequency, and computing location.
Second, prefer the smallest model that meets accuracy, reliability, and safety requirements. Use batch processing, caching, and retrieval strategies to avoid unnecessary computation. Set rules for when generative AI is appropriate and when conventional software is sufficient.
Third, include energy, emissions, water, and hardware disclosures in vendor contracts. Require providers to explain whether reported figures cover training, inference, cooling, and embodied emissions. Use model-efficiency dashboards and coding-session estimators to identify high-impact workloads.
Additionally, consider local water stress and grid carbon intensity when evaluating data center locations. Report both efficiency improvements and total AI demand, since lower energy per task does not guarantee lower aggregate consumption.
For businesses working towards carbon reporting compliance under PPN 06/21, AI use may need to be included in emissions inventories. Similarly, companies subject to Streamlined Energy and Carbon Reporting should consider whether digital infrastructure falls within their reporting boundary.
Training can also help. The SBS Academy offers courses on carbon measurement and environmental management systems. These can help teams understand how to incorporate AI impacts into existing reporting processes.
Ultimately, sustainable AI depends on disciplined decisions about when AI is necessary, which model is sufficient, and whether the expected benefits justify the resource demand. No single calculator or framework can capture every impact, particularly when proprietary providers withhold operational data. However, the tools now available can help businesses make more informed choices.
Where to find authoritative guidance
The International Energy Agency publishes detailed analysis on data center energy demand and AI's electricity consumption. Its reports include global projections and regional breakdowns.
The Green Software Foundation provides open-source tools, training, and standards for measuring and reducing software emissions. Its resources include courses for developers and business leaders.
UK businesses subject to carbon reporting requirements can find guidance on the government's Streamlined Energy and Carbon Reporting page. This includes information on reporting boundaries and calculation methodologies.
For organizations preparing for public sector tenders, the Procurement Policy Note 06/21 sets out requirements for carbon reduction plans. Understanding how AI fits into these plans may become increasingly important for suppliers.
Finally, the Trellis tracker maintains an updated directory of sustainable AI tools and frameworks. It includes vendor information, use cases, and technical specifications for measurement and model-selection resources.