Artificial intelligence may appear weightless to the employees using copilots and generative platforms. Yet every prompt depends on an expanding physical infrastructure of processors, cooling equipment, power systems and network connections.

That infrastructure is growing at extraordinary speed. The International Energy Agency forecasts that global data centre electricity consumption will more than double to approximately 945TWh by 2030, slightly more than Japan consumes today. Demand is expected to grow by around 15% annually between 2024 and 2030, more than four times faster than electricity use across other sectors combined.

This creates a difficult contradiction for enterprises. AI is being positioned as an engine of productivity and competitive advantage, but the infrastructure supporting it could place corporate climate commitments under mounting pressure. Renewable energy procurement may help address reported emissions, but it cannot automatically eliminate local grid congestion or competition for water.

The question, therefore, is no longer simply whether organisations can obtain enough computing capacity. It is whether they can expand their use of AI without transferring its environmental and infrastructure costs to energy networks and future sustainability targets.

Power is Becoming the Real Constraint

The scale of the challenge is particularly visible in the US, where data centres could consume 11.8% of all electricity by 2030 under Berkeley Lab’s reference scenario. Its other projections range from 9.5% to 15.3%, illustrating how rapidly the position could change depending on AI adoption, server utilisation and infrastructure efficiency.

Accelerated servers – the systems primarily supporting AI – are forecast to increase their electricity consumption by approximately 30% a year to 2030. They could account for almost half the net growth in global data centre demand.

For enterprise leaders, this is not an abstract problem for cloud providers to solve. Rising demand can translate into higher costs, delayed grid connections, restricted capacity and greater scrutiny of corporate environmental claims.

“You wouldn’t sign off a new production line without knowing its running costs, but that’s exactly what most companies are doing with AI,” Adhum Carter Wolde-Lule, Director at Prism Power, told Silicon UK. “Not every workload is equal. Training a large model, running a chatbot and doing routine inference have very different footprints, and often a smaller model does the same job for a fraction of the energy.”

um Carter Wolde-Lule, Director, Prism Power.

The logical response is to make energy, carbon and water part of the business case for every significant AI workload. Organisations already compare solutions according to cost, latency, accuracy and security. Resource consumption should become another decision-making dimension.

Dan Magestro, Chief Data and AI Officer at Sphera, says this is “quickly becoming a capital allocation question, not just an ESG one”. However, many enterprises still lack workload-level visibility and possess only company-wide or facility-wide figures.

This makes it difficult to distinguish a high-value application from an energy-intensive experiment that produces only marginal returns. Enterprises need metrics that connect resource consumption to useful results: transactions processed, design cycles shortened, fraud detected or revenue generated. Tokens per watt and useful compute per unit of resource could ultimately be more informative than the headline size of a model.

The challenge is also organisational. Sandhya Sabapathy, a former FTSE 100 sustainability director and founder of Kaleidoscope, commented that sustainability and infrastructure teams are frequently rewarded for opposing outcomes. “The sustainability function is rewarded for the credibility of the commitment. The infrastructure function is rewarded for capacity, latency and uptime,” she says. Without shared governance, efficiency gains can simply reduce the cost of consuming more computing power rather than place any meaningful constraint on demand.

Cooling Becomes a Strategic Decision

AI is changing the physical design of data centres. Traditional air cooling is reaching its practical limits as more powerful processors are concentrated into increasingly dense racks. Cooling can already account for 33% to 40% of a facility’s total energy consumption, depending on location and power density.

Liquid cooling removes heat closer to the processor and can let operators support more compute in the same physical space. Yet it is not a universal environmental solution. Some configurations reduce fan and chiller energy, while others can increase pumping requirements or move heat to another part of the facility. Operators must also consider water consumption, refrigerants, local climate, and the carbon intensity of electricity.

“For operators scaling high-density AI infrastructure, liquid cooling is one of the most immediate interventions because it addresses a constraint already emerging inside the rack,” says Shahar Belkin, Chief Evangelist for ZutaCore. “Every watt consumed still has to leave the server as heat.”

Shahar Belkin, Chief Evangelist for ZutaCore

Belkin argues that operators must examine the complete thermal chain, including pumps, fans, chilling and heat rejection, rather than relying on a single efficiency metric. A design that uses less water but requires energy-intensive chilling may merely exchange one environmental burden for another.

Local conditions are critical. Closed-loop and waterless cooling systems can be particularly valuable in water-stressed areas. Conversely, a technology that performs well in a cool, water-abundant region may be a poor choice in a drought-prone area or one with grid congestion.

US data centres directly consumed approximately 66 billion litres of water in 2023, according to Berkeley Lab’s 2024 report. Their indirect footprint was substantially larger: the electricity generation serving those facilities consumed almost 800 billion litres. This demonstrates why operators cannot treat energy and water as separate considerations.

Reza Azizian, CEO of Ferveret, says the focus should be on the useful computation produced for every unit of resource consumed. Better heat removal can reduce infrastructure overhead while letting processors run more efficiently. It may also expand the range of locations in which facilities can be built, enabling operators to place compute closer to renewable generation or available grid capacity.

Heat reuse could create an additional benefit, with captured heat supplied to nearby homes, offices or industrial facilities. However, it requires suitable local demand and infrastructure. It should be treated as part of the site strategy from the outset, not added later to improve a project’s sustainability narrative.

Renewables Cannot Solve the Problem Alone

Technology companies have become major buyers of renewable electricity. They accounted for around 40% of corporate renewable power-purchase agreements signed in 2025, according to the IEA. Nevertheless, contractual renewable procurement does not necessarily mean that a facility is physically operating on clean power every hour.

Data centres currently receive approximately 30% of their electricity from coal, 27% from renewables, 26% from natural gas and 15% from nuclear energy. Although renewables are expected to supply nearly half the additional electricity data centres require through 2030, gas and coal together could provide more than 40% of the increase.

On-site solar and wind generation can reduce grid dependency, but their output is variable and frequently insufficient for continuously operating AI infrastructure. Battery storage can bridge short gaps, reduce demand peaks and help operators avoid drawing maximum capacity during periods of grid stress. It cannot yet economically power most hyperscale facilities through prolonged periods of low renewable generation.

Microgrids can combine renewable generation, battery storage, grid supplies and, in some cases, gas-powered generation. This gives operators greater control and resilience, although planning requirements, capital costs and long equipment lead times mean microgrids are generally a medium-term intervention.

Magestro says the outcome will be hybrid for some time. Adaptive grid use, on-site generation and storage can reduce dependence on a single power source, but fossil-fuel backup is likely to remain part of many facilities in the near term.

Operators should also look beyond individual sites. Masahisa Kawashima, IOWN Technology Director at NTT, says photonic networking could allow distributed data centres to operate as a single computing fabric. Enterprises could send workloads to locations where low-carbon electricity and capacity are available while keeping sensitive data under their control.

Masahisa Kawashima, IOWN Technology Director, NTT.

“For enterprises, the focus over the next few years is unlikely to be large-scale model training,” says Kawashima. “Most will instead be fine-tuning models and running inference against their own proprietary data. Photonic networks can connect that data securely to the compute required, wherever it resides. Sustainability can then become a factor in where each AI job is executed, rather than a constraint addressed only when building new infrastructure.”

Smarter workload management offers a more immediate opportunity. Flexible, non-urgent tasks can be scheduled for times or regions with abundant renewable electricity. Enterprises can also select smaller models, limit unnecessary inference, improve hardware utilisation and retire AI applications that do not produce sufficient value.

Sustainability Must Become a Governance Constraint

Engineering innovation will be essential, but it will not resolve the paradox on its own. More efficient processors, cooling systems and power infrastructure lower the resources required for an individual task. They can also make AI cheaper, encouraging organisations to run far more of it.

Johanna Ahola-Launonen, Academy Research Fellow at Aalto University School of Business, says enterprises should challenge the assumption that every commercially viable AI application deserves deployment. “The fact that AI can be extraordinarily valuable is a reason to prioritise its uses, not a reason to treat all AI demand as equally necessary,” she commented. “We need to become much more explicit about which uses of AI are important enough to justify their environmental costs.”

Ahola-Launonen, Academy Research Fellow at Aalto University School of Business.

This is also becoming a question of community consent. Data centres can compete with homes, hospitals and local businesses for finite grid and water resources. Operators need to engage utilities and communities before acquiring land, publish anticipated grid loads and water requirements, and contribute to infrastructure improvements that benefit the wider region.

Enterprise customers can exert pressure through procurement. They should ask cloud and AI providers for workload-level energy, emissions and water data broken down by location and, where possible, time. Annual corporate averages reveal little about whether a workload ran on renewable electricity or fossil-fuel generation during a particular period.

Businesses should also incorporate AI-related emissions into Scope 2 and Scope 3 measurement, even when the initial data is incomplete. Waiting for perfect information could leave several years of growth invisible.

Ozgur Duzgunoglu, Head of Engineering and Design at Telehouse Europe, says organisations should consider staged, modular expansion rather than procuring speculative capacity. This can reduce unnecessary capital expenditure and the embodied carbon locked into infrastructure that may never be fully used.

Ozgur Duzgunoglu, Head of Engineering and Design, Telehouse Europe.

Ultimately, enterprises need someone with the authority to challenge or reject an AI workload when its environmental and commercial costs cannot be justified. Sustainability targets are meaningful only if they influence technology choices before contracts are signed and applications become operationally indispensable.

AI growth and climate commitments are not necessarily incompatible. Liquid cooling, efficient chips, batteries, microgrids, renewable generation and intelligent workload placement can all reduce the impact of each unit of compute. But technology cannot compensate indefinitely for uncontrolled demand.

The enterprises most likely to reconcile AI with sustainability will be those that treat energy, water and carbon as strategic constraints rather than reporting concerns. The winners of the AI race may not be the organisations that secure the most compute, but those that extract the greatest business value from every watt they consume.