Perspectives

Artificial intelligence is accelerating at pace and is unlocking extraordinary leaps in productivity. While the new technology increases data processing capacities, recognises usage patterns, and outlines carbon pathways, the boom in productivity comes with growing environmental challenges. AI has a rapidly expanding energy, water, and carbon footprint.

As organisations integrate AI into everyday operations, sustainability can no longer be an afterthought. If the future of business is AI, AI must be sustainable.

AI models come with large energy demands. Depending on the size of the model, training a model can consume electricity comparable to the use of 100 UK homes in a year (footnote). However, today, most emissions come not from training but from using the models. Every query sent to an LLM is being processed and responded to. Each inference has emissions, and there are millions or billions of them every day, as we come into a habit of the continuous running of AI models.

Data centres are already responsible for an estimated 2–3% of global electricity consumption, and this is rising sharply. Google recently reported a 51% increase in emissions from 2019–2024, driven primarily by AI expansion. Independent analyses suggest the true figure may be closer to 65%. Other major companies, like Microsoft, Meta, and Amazon, have also quietly delayed, softened, or reinterpreted their net-zero pathways as AI demand accelerates faster than renewable energy deployment.

AI is not just energy-intensive, it is also resource-intensive. Data centres require enormous volumes of water for cooling and then there is land use, heat waste, and heavy hardware turnover. The processors in these data centres contain rare minerals that have further externalities through mining.

At the Oxford Centre, we approach AI sustainability the same way we approach buildings: whole-life carbon.

1. Embodied Carbon

Servers, chips, batteries, cooling systems, and the buildings that house them all have significant embodied carbon. With hardware cycles as short as 3–5 years, this embodied carbon repeats itself far more quickly than in other building typologies.

2. Operational Carbon

This is where most of AI’s impact lies. Continuous queries, cooling operations, and uninterruptible power systems all contribute to growing operational emissions, as models become more capable and more frequently used.

3. Indirect Carbon

The global supply chain for semiconductors, data centre construction, and global logistics adds further emissions not usually captured in AI discussions. Responsible AI adoption needs to look beyond performance metrics and integrate environmental performance.

Despite its footprint, AI can be one of the most powerful tools for sustainability.

It is already optimising grid management, accelerating renewable integration, improving energy modelling for buildings, and reducing material waste through predictive design. Tools like CalCO₂, developed by Equitis, show how AI can help organisations quantify, understand, and reduce carbon footprints with far greater precision.

The challenge is ensuring more decarbonisation impact than emissions, so that the carbon ROI of AI is positive.

To ensure AI supports rather than undermines sustainability commitments, organisations should:

• Map the whole-life carbon of their AI usage
• Choose cloud providers based on sustainability performance, not only cost
• Use smaller, more efficient models where possible
• Require transparency on data centre energy sources
• Integrate AI governance into existing ESG frameworks
• Align AI adoption with standards: ISO 14064, RICS WLCS, SBTi, and emerging EU requirements

AI will be one of the defining technologies of the next decade. The question is whether it will also become one of the defining sources of emissions. We have the opportunity to seize the moment to build AI systems that are powerful, responsible, and sustainable.

At the Oxford Centre, we believe the future of technology can be a sustainable one.