Over the last year, across our AI in education communities, we have continued to have many conversations around AI and sustainability. Within our discussions, members have shared how questions and concerns around the environmental impact of AI are increasing. With some also encountering resistance amongst staff and students to the use of AI tools on environmental grounds.
To approach AI critically and have these important conversations with colleagues and students requires an informed understanding of the environmental impact of AI, however getting to the facts is challenging. A lack of transparency from AI developers remains a key part of this issue, but we have more insight into the impacts now than ever before.
The United Nations University (UNU) has recently released a report presenting a thorough assessment of impacts and recommendations: Environmental Cost of AI’s Energy Use Carbon, Water and Land Footprints.
The UNU’s report considers a fuller picture of the environmental impact, not just considering energy use but also water use, land use, e-waste, human labour, and more. It still contains estimates and projections based on what data is available, for example estimating the costs of training newer models based on data from earlier models.
Even though the data isn’t exact, we find the insights drawn are useful to informing our individual and organisational approaches to AI and sustainability in education. Here, we bring together some of the key points from the UNU’s report as well as other recent publications on AI and environmental sustainability.
We also refer members to newly released practical guidance from the Environmental Association for Universities and Colleges EAUC and Jisc: AI and environmental sustainability in post-16 education
Key insights on environmental costs
Training input is very costly and inference costs far exceed training costs
The UNU report makes clear that while yes, training models has incredibly high energy costs, the cost of the continuous use of those models still manages to exceed this by magnitudes. They estimate that energy use at the “inference stage”- all of us using the tools and systems built on those models, accounts for approximately 80-90% of total AI energy consumption.
This figure shows that use of AI has significantly more impact than the training stage, this might be surprising to some as earlier conversations focused on the intensive energy costs of training AI models. This emphasises the need to critically evaluate if, when and how we use AI. Both as individual users, and as organisations, this is an area we do have an amount of control over through our personal and institutional approaches to AI use.
Responsibility for inference costs doesn’t sit only with the end user however, the responsibility of developers for these costs has also been raised. A recent literature review of sustainable AI discourse, finds that the deployment phase seems to be a particular blind spot when it comes to sustainability. Development choices could make a great deal of impact to inference costs, for instance by making it easier to control and limit AI features within tools – ensuring that AI use is an intentional choice and not a default. This is not currently typical however, there is hope that pressure from consumers could have some influence.
All AI is not equally impactful
Importantly, the report also highlights the key differences between costs of different types of AI models and systems. That the costs increase between non- generative and generative AI systems, and then further escalate through text, image and video generation is perhaps not too surprising. However, the figures they provide illustrate the scale of that escalation and warn particularly of the resource intensiveness of complex video generation.
“A typical ChatGPT-style query is about 200 times more energy-intensive than basic text classification; long GPT-style responses can cost around 1,000 times more than text classification; and AI-generated images can require roughly 1,450–2,000 times the energy of text classification. Video is the new energy frontier: a single short AI-generated video can draw as much electricity as 200,000 spam classifications or hundreds of AI-generated images.”
The sustainable AI literature review highlighted too that few sources considered whether AI is valuable enough to be worth the cost. With a lot of discourse having an air of inevitability about AI development, and a focus on finding technical solutions to the costs of AI rather than limiting or cancelling AI development.
Alongside this is the proposition that AI can present meaningful solutions to the climate crisis, which can be used as justification for continuing AI development despite the environmental cost. Ketan Joshi’s report The AI Climate Hoax: Behind the Curtain of How Big Tech Greenwashes Impacts, analyses these claims and finds that the AI development which incurs the highest environmental cost – generative AI systems, are largely not those which are being used to create sustainability solutions.
It is vital that the variation in AI technology and the related environmental impacts is understood, and that this informs the choices we make in whether and how to use different types of AI.
Data centres aren’t only about AI but they’re accounting for more and more of their energy demand
Data centres, the buildings and physical infrastructure required to support our global digital systems, are central to conversations around AI and sustainability. The UNU’s report outlines the considerable energy, water and land footprints of data centres, all of which are set to accelerate as AI development increases demand for energy. They refer to Gartner’s projection that the proportion of data centre electricity use attributable to AI workloads will double by 2030 from 20% to 40%. The UNU estimates this to be 378 TWh, which they state would meet the residential electricity needs of all 1.3 billion people in Sub-Saharan Africa for over 2 years.
The location of data centres is a key aspect too. Technically, the environmental impact of a data centre can be mitigated by its location – What land is it built on? What energy sources does it use?
In the UK there are just under 500 data centres (as of 2025 data), and there are increasing concerns around how new developments will support expansion of the fossil fuel industry and jeopardise national climate targets. It has been recently reported that over 100 newly proposed data centres have requested gas connections, meanwhile Buckinghamshire development plans include an on-site gas turbine.
Globally, it is the US which dominates with almost half of all data centres worldwide (4,165 in 2025). In recent years we have seen the expansion of data centres strongly protested by US residents due to the stress caused to local water and energy infrastructure as well as direct environmental damage. The Brokovich data centre reporting project provides a tracking map of AI data centre expansion across the US and the efforts of the public to resist new builds.
Countries with fewer or no data centre infrastructure will experience less direct environmental impact, but they can also be left without access to the benefits – including having a stake in AI development. There is a potential tension emerging between AI sustainability and AI sovereignty here, as access and control of AI capability becomes a national issue and more nations look to build the infrastructure needed to enter the AI development space.
What are our responsibilities?
Each of these points emphasise the need to approach AI use critically. In doing this though we have different responsibilities as individuals and as part of organisations. In some cases we will feel like our impact is negligible and that others bear more responsibility. This should not prevent us though from doing our best to affect positive change where we can.
The UNU report outlines roles and associated responsibilities to make real change, from developers to governments, international bodies to individuals.
Perhaps the most relevant here is that for individuals and organisations:
“Users and enterprises should recognize their everyday influence over AI’s environmental impact. They should choose low-energy options when these meet the need and escalate to generative models only when they offer real value. Organizations should adopt energy or token budgets, encourage concise use, and default to smaller or on-device models. Procurement should incorporate criteria linked to environmental footprints.”
These recommendations focus on the reduction of the costs of the inference stage as well as influencing the market through procurement practices.
The UNU recommendations are pitched at a high level, combining individual and organisational responsibilities. The AI and environmental sustainability in post-16 education guide discusses practical advice for responsibilities across more granular, sector specific levels. These include general advice for all as well as role specific recommendations for students, staff (including teaching and learning, research, IT, professional services), senior leaders and governors.
Staying up to date
New research and data becomes available regularly in this space, Jisc’s digital sustainability newsletter is a valuable resource for keeping up to date with new developments and covers all things digital sustainability including AI. It was an essential resource to this piece too.
Sustainability also continues to be a regular topic across our AI in Education communities. Get involved with our communities to join those discussions and explore solutions together with peers from across the sector.
Find out more by visiting our Artificial Intelligence page to explore publications and resources, learn more about our communities and sign up for our AI Literacy training.
For regular updates from the team sign up to our mailing list.
Get in touch with the team directly at AI@jisc.ac.uk