As demand for AI infrastructure accelerates, are data centres becoming fundamentally incompatible with corporate net-zero commitments, or can innovation keep their environmental impact under control?

“I don’t think AI infrastructure is inherently incompatible with net zero, but it is incompatible with the idea that efficiency gains alone will solve the problem. We need to look at absolute resource use, not simply whether each generation of hardware is more efficient than the last. If AI demand grows faster than grids can decarbonise, corporate targets will come under real pressure.

“The answer is a combination of cleaner power, better chips, smarter cooling, flexible workloads and, critically, more disciplined decisions about what we use AI for. Net zero cannot become an accounting exercise that sits separately from digital strategy. If AI is now core infrastructure, then its energy, carbon and water requirements must be designed into corporate sustainability plans from the outset.”

Should organisations assess the business value generated by an AI workload against its energy, carbon and water footprint before deciding whether to deploy it?

“Yes. We should stop treating every technically possible AI use case as automatically worth deploying. The relevant question is: what economic or social value does this workload create relative to the resources it consumes?

“That does not mean reducing every decision to a single carbon number. It means bringing energy, carbon, water and hardware impacts into the same investment conversation as cost, productivity, risk and revenue. Some AI workloads may justify a significant footprint because they create substantial value; others may be marginal conveniences with disproportionate resource demands.

“Organisations already apply hurdle rates to capital. We need an equivalent discipline for digital infrastructure. Sustainable AI starts with asking whether a workload should exist at all, then making the workloads we do choose as efficient as possible.”

Which technologies offer the most immediate opportunity to reduce the environmental impact of AI data centres: liquid cooling, more efficient chips, battery storage, microgrids or smarter workload management?

“The quickest gains are likely to come from a combination of more efficient chips and smarter workload management, because both attack the problem at the source: how much compute we use and when we use it.

“Liquid cooling can be extremely important for high-density AI systems, but its environmental benefit depends on the facility and local conditions. Batteries and microgrids are valuable, particularly for flexibility and resilience, but they are infrastructure investments rather than a substitute for reducing demand. In the current era of climate change, we should not search for a single winning technology. AI data centres are systems. The best results will come from optimising the whole system (model design, hardware, cooling, scheduling, power supply and location) rather than improving one component, which risks just shifting the environmental burden somewhere else.”

Liquid cooling can reduce water consumption at some facilities but may increase electricity demand elsewhere. How should operators evaluate trade-offs between energy efficiency, water use and local environmental conditions?

“Operators need to move beyond a single efficiency metric. A lower PUE is not automatically a better environmental outcome if it comes at the cost of higher water stress, and low water use is not automatically better if it materially increases electricity demand on a carbon-intensive grid.

“The correct unit of analysis is the whole system in its local context. That means looking at energy, carbon and water together, including when electricity is consumed, where water comes from, local water stress and the consequences of heat rejection. Operators should use location-specific lifecycle and resource assessments and make the trade-offs explicit. Sustainability is not about optimising one number. It is about operating within multiple environmental constraints simultaneously and avoiding solutions that simply move an impact from one part of the system to another.”

Can on-site renewables and battery storage realistically provide the reliable power required by AI infrastructure, or will data centres remain dependent on national grids and fossil-fuel backup?

“On-site renewables and batteries can make a meaningful contribution, but for most large AI facilities they are unlikely to provide continuous, reliable power on their own. Data centres require substantial firm capacity; solar and wind are variable, while batteries are currently much better at shifting electricity across hours than solving every long-duration resilience challenge.

“So national grids will remain central, which makes grid decarbonisation part of any AI sustainability agenda. The more interesting opportunity is to perhaps make data centres better grid participants: locating capacity where clean power is available, matching flexible workloads to periods of abundant renewable generation, investing in storage and supporting additional clean generation. Fossil backup in this case should not be treated as inevitable forever, but replacing it requires a genuine resilience strategy rather than simply purchasing more renewable certificates. Effectively, we should use AI data centres to help phase out fossil fuels if possible.”

How can data centre operators avoid monopolising limited electricity-grid capacity or water resources at the expense of local businesses, communities and essential services?

“The starting point is to recognise that electricity and water are shared infrastructures, not simply private inputs that can be purchased by the highest bidder. Data centre development therefore has to be planned at system level with utilities, local government and communities.

“Operators should be transparent about projected demand, fund the infrastructure upgrades their facilities require, avoid development in already constrained locations and use flexibility wherever workloads allow it. They should also demonstrate local value and avoid externalising infrastructure costs onto communities around them. Simple examples include waste-heat re-use, helping to install and fund community energy schemes or reducing energy arrangements for nearby public facilities. In some places, hard limits may be appropriate. Digital infrastructure is important, but so are homes, hospitals and local industry. Resilience means designing a system in which those needs can coexist, rather than allowing one sector to crowd out everything else.”

What information should cloud and AI providers disclose so that enterprise customers can accurately understand the carbon, energy and water consequences of their workloads?

“Enterprise customers need workload-level transparency, not broad corporate sustainability statements. At minimum, providers should disclose the energy consumed by a workload, its associated carbon impact, direct and indirect water use, and the embodied impact of hardware where that can reasonably be allocated.

“Customers should also be able to understand where and when the workload ran, the energy mix supporting it, and how renewable-energy claims, offsets or other accounting mechanisms have been calculated. Crucially, methodologies need to be consistent enough to allow meaningful comparison between providers. Companies cannot properly manage what they cannot properly measure. If cloud and AI services are becoming a material part of an organisation’s environmental footprint, the environmental information associated with those services should ultimately become as accessible, comparable and auditable as the financial bill. This needs standards, accounting standards and new measurement frameworks to be agreed.”

What practical steps should business leaders take now to ensure that expanding their use of AI does not undermine their organisation’s sustainability targets?

“Business leaders should make sustainability part of AI governance now, before usage becomes too embedded to manage. Start by establishing a baseline: which AI workloads are running, what they cost, what value they create and what their energy, carbon and water impacts are. Then introduce a deployment gate that asks whether AI is actually necessary and whether a smaller model or less compute-intensive approach can achieve the same outcome. Procurement matters too: demand workload-level environmental information from cloud and AI suppliers and favour providers that can demonstrate credible low-carbon infrastructure.

“Finally, connect AI growth to existing carbon budgets and sustainability targets rather than treating it as a separate technology programme. The leadership question is not simply how quickly we can adopt AI, but whether digital transformation creates lasting value without undermining the systems on which it depends.

“The risk is that we frame this as a data-centre engineering problem when it is really a question about how we choose to use digital resources. Better chips, cooling and clean power all matter, but ultimately sustainability requires us to ask what we are using the compute for, what value it creates, and whether that value justifies the resources consumed. The most sustainable unit of compute is still the one you discover you didn’t need.”

Dr Catherine Mulligan, Founder of Digital Crossroads.

Catherine Mulligan, PhD, FHEA, is an internationally recognised expert in digital technologies for sustainability and resilience. She is also a visiting academic at Imperial College London and founder of Digital Crossroads, where her work focuses on ensuring digital transformation supports economic, social and environmental resilience in practice, not just in theory. Furthermore, she is a former Fellow of the World Economic Forum and a member of the United Nations Secretary-General’s High-level Panel on Digital Cooperation.