Loughborough University (UK)

Prioritising sustainable AI for AI growth. A key issue here is that AI infrastructures are struggling to keep pace with both the short-term data demand and longer-term growth aspirations. Just as an example, the organisation, Gartner, predicts that by 2027 40% of AI data centres will suffer from operational limitations owing to power shortages, which is an alarming thought. No country or organisation is infallible to this 'data demand vs. infrastructure' issue, in the United States for instance it was recently reported by the World Economic Forum that $64 billion dollars of data centre projects have been halted or delayed due to local opposition since 2023. One opportunity to address infrastructure limitations is through more effective engagement with sustainable data practices across three AI infrastructure pillars: compute, network and storage. The better management of data waste specifically promises capacity efficiencies that can help to realise the AI growth agenda.

Safe, secure and trustworthy AI;AI capacity-building;

Sustainable AI could fall under 'responsible AI' consistent with the OECD.AI and this is closely tied to AI capacity building. The sustainability--capacity building nexus is a critical issue globally as we're seeing an increasing emphasis on AI Sovereignty, and a mindset of 'we'll do it ourselves' leading many countries to seek to further their own AI agenda within their own geographies. This is playing a contributing role to increasing data needs for the development and deployment of AI systems globally, with observers suggesting this is a big factor leading the world to a projected data creation explosion over the next decade. With renewable electricity resources being finite, AI sustainability practices and AI growth (capacity building) must go hand-in-hand, they are not mutually exclusive as is often assumed

It would be sensible to identify 'sustainable AI for growth' as its own thematic area given the centrality of this to AI growth. Data is the cornerstone of AI development and deployment. As reported recently by Gartner, data storage could account for around 36% of the total estimated data centre electricity use by 2030, so about double the amount in 2022. Of this stored data, a recent use-case from AWS showed up to 80% of an organization's primary business data was dark data, which equated to petabytes (1000's of terabytes) in size. Such data is created in training models, which remains stored but not used again. Several questions are then raised: • AI systems are inherently data intensive, and this increases the costs involved, but do we need to store all the data that is generated in the training and deployment of AI models? • What are the minimum viable data required for training? • How is data managed throughout the AI lifecycle? • Are redundant data like training data disposed of or at the very least placed in the most cost-effective storage option? Addressing questions like these under the theme of sustainable AI would increase capacity by reducing waste, decrease costs to organisation, and minimise the AI carbon footprint.

The global challenges highlighted by the Gartner and the World Economic Forum, namely: - by 2027 40% of AI data centres will suffer from operational limitations owing to power shortages. - in the United States $64 billion dollars of data centre projects have been halted or delayed due to local opposition since 2023. Are being felt in many countries, including the UK, when it comes to planning and implementting AI infrastructure growth ambitions, e.g. more AI data centres are needed but where can they be located? Can the grid support the extra electricity demands? How to overcome resistance from local communities for planned pylons and energy infrastructure changes? etc.

The AI Dialogue can advance international cooperation on AI governance by promoting key activities to support sustainable AI for AI growth by governments that might include, but are not limited to, - Integrating dark-data accounting into public-sector digital procurement and IT governance. - Establishing minimum data-efficiency requirements for publicly funded digital systems (e.g., evidence of data minimisation, archival and waste policies). - Incentivising the adoption of circular-economy approaches to data: such as retention limits, low-energy storage, data lifecycle audits. - Providing digital-carbon reporting frameworks for SMEs and local governments that are simple and accessible for future economic growth. - Finally, national training programmes for digital carbon literacy across government, industry, and academia is important to building education, which plays a key role.

There is tremendous value in bringing together AI ecosystem actors by way of multi-stakeholder alliances, that can be formal or informal and can advance knowledge by sharing tools and practices. Bridging government–industry–academia through technical working groups and community forums provides a means for open discussion and debate. The OECD.AI catalogue of tools is a great example of a digital community coming together to share best practices, tools and use-cases. So to is the Coalition for Sustainable AI. Generally speaking, the data industry remains quite opaque, this is changing, but by bringing people together and providing safe spaces for people to challenge assumptions, question the status quo, and raise new AI viewpoints the co-creation of new ideas and ways forward can emerge. Through collaboration we can develop and promote appropriate and relevant tools to increase public awareness and equip individuals with the skills to harness AI technologies while mitigating their footprint, thereby fostering greater digital inclusion and reducing the digital divide.

Providing information, training and technical assistance to promote capacity building in the public sector is key, but so too are the - Promotion of certification schemes for digital sustainability tools to build trust and consistency. - Embed good data ethics and digital-emissions reporting. - Sponsoring skills pipelines, training for CIOs, data engineers, sustainability leaders. - Lead by adopting tools in public services first; demonstrate value through use cases and open dashboards. - As well as engaging with the development of standards for measuring digital emissions and impacts, such as the development of the IEEE P7100.

Policy level understanding and knowledge is often not shared at the level of organisations and, thus, this makes policy execution extremely difficult. For example, AI compute resources can be general-purpose, serving both AI and non-AI workloads, or specifically tailored for AI, yet this distinction is often not clearly made. This issue is further complicated by the lack of standardized and validated data on AI-related compute, as emphasized by the OECD.AI expert group on AI Compute and Climate. The expert group highlight a range of related challenges when it comes to capturing compute resource needs and commitments; the group's Insights from preliminary survey results of organizations revealed the following: • 31% reported that they do not measure how much AI compute they have. • 20% reported that they did not know whether they measure AI compute. • 52% of respondents reported challenges accessing sufficient AI compute. When asked about the percentage of their organisation's total annual costs spent on AI compute: • 37% reported that they did not know. • 5% reported no annual costs spent on AI compute. • 26% reported 10-40% of costs. • 3% reported that AI compute costs were 50% or more of annual costs. Though the findings serve only as an illustration given the limited sample size, it is clear that issues of transparency, understanding, and knowledge pertaining to AI compute resources are real. This presents significant challenges in attempts to capture the real energy and fiscal costs of AI compute its subsequent impact on the environment.

AI education is lacking for many. Close partnerships with academia can help to address this issue and, by increasing knowledge and understanding of users, engagement levels will increase. It is difficult for organisations to engage when they do not understand the issues fully. As an example, The 2025 OECD D4SME Survey of SMEs revealed: - 67% of non-users are unsure about how to use generative AI or the risks involved, compared to 42% of users, though this suggests that still a substantial share of users remains uncertain about the risks. - Additionally, three quarters of non-users worry about harmful content, compared to 52% of users, suggesting concerns lessen with use. Top concerns for both groups include copyright and legal issues, inaccurate information, with around 80% viewing these as barriers. - 83% of non-users versus 70% of users are concerned about data privacy. - the environmental impact of AI does not feature in the top concerns outlined by respondents, which presents a barrier to sustainable AI adoption

The research on Digital Decarbonisation at Loughborough University: https://www.lboro.ac.uk/research/digital-decarbonisation/