AI can help decarbonise complex systems, but its rapidly growing physical footprint is testing power grids, water supplies and corporate climate commitments. The outcome will depend less on the technology’s promise than on the choices organisations make now.
The paradox behind the promise
Artificial intelligence (AI) is often described as an intangible technology: an algorithm answering a question, forecasting demand or finding a pattern hidden in millions of data points. Yet every AI response has a physical foundation. It is produced by semiconductor-intensive servers housed in data centres, connected through networks, cooled by air or water and supplied with electricity around the clock. That makes AI simultaneously a potentially powerful sustainability tool and a fast-growing source of environmental pressure.
On one side of the ledger, AI can make energy systems more flexible, factories less wasteful, logistics more efficient and climate risks easier to anticipate. On the other, the race to build and deploy ever-larger models is increasing demand for electricity, water, land, minerals and equipment. The International Energy Agency projects global data-centre electricity consumption to rise from roughly 460 terawatt-hours in 2024 to about 945 TWh in 2030. AI is the most important driver of that growth. The central question, therefore, is not whether AI is “green” or “dirty”. It is whether the environmental value created by a particular use of AI is greater than the full footprint required to deliver it – and whether that value would have occurred anyway through a simpler solution.
This distinction matters. Sustainability cannot be established by showcasing an impressive pilot while ignoring the infrastructure behind it. Nor should AI be rejected because it consumes resources. The relevant test is net impact: measurable system-wide benefits minus operational emissions, embodied carbon, water stress, material use and rebound effects.
Where AI can accelerate sustainability
AI’s strongest environmental contribution is its ability to manage complexity at a speed and scale that conventional analysis cannot easily match. Electricity systems are a prime example. As solar and wind become larger shares of generation, grid operators must balance weather-dependent supply with changing demand. Machine-learning models can improve renewable-energy forecasts, predict equipment failures, detect faults and help schedule flexible loads. Better forecasts can reduce the need for costly reserves and curtailment, while predictive maintenance can extend the useful life of turbines, transformers and transmission assets.
Buildings and industry offer another large opportunity. AI-enabled controls can adjust heating, cooling, lighting and industrial processes in response to occupancy, production schedules, weather and energy prices. Computer vision can identify defects earlier, reducing scrap. Digital twins can test process changes before physical equipment is modified. In sectors such as cement, steel, chemicals and food processing, even modest improvements in yield, heat management or downtime can avoid significant energy and material waste when applied across thousands of facilities.
Supply chains can also become more resource-efficient. Companies routinely overproduce because they cannot accurately anticipate demand. AI can improve forecasting, inventory placement, vehicle routing and load consolidation. It can help trace high-risk commodities, detect deforestation signals in satellite imagery and identify suppliers exposed to floods, droughts or heat. Used carefully, these capabilities can reduce empty journeys, spoilage and emergency air freight while strengthening resilience.
Climate science and adaptation may be among the most socially valuable applications. AI can accelerate the processing of satellite data, improve local weather “nowcasting”, map wildfire or flood exposure and translate complex climate information into operational warnings. It can support farmers with crop and irrigation decisions and help cities target inspections or investment. These uses do not replace climate science, engineering or local knowledge; they can make those disciplines faster, more granular and more accessible.
The expanding footprint
The benefits, however, do not erase AI’s direct impacts. Training frontier models is computationally intensive, but the environmental burden does not end after training. Once a model is embedded in search, office software, customer service and consumer devices, billions of inference requests can collectively outweigh the footprint of a single training run. Rapid hardware turnover adds embodied emissions from chip fabrication, construction and global supply chains, as well as a growing stream of electronic waste.
Electricity is the most visible constraint. The IEA expects data-centre demand to grow far faster than overall electricity consumption through 2030. Renewables are projected to meet nearly half of the additional generation needed globally, but natural gas and coal will also contribute. Location and timing therefore matter as much as annual electricity totals. A data centre operating in a fossil-heavy grid during peak hours can increase emissions and strain local capacity even if its owner buys enough renewable-energy certificates to match annual consumption elsewhere.
The impact is highly concentrated. A national percentage can look manageable while a cluster of data centres overwhelms a local grid, delays other connections or requires new generation and transmission. In the United States, Berkeley Lab estimates data centres used about 4.4 per cent of electricity in 2023; its 2025 update projects they could account for 9.5 to 15.3 per cent by 2030. Such uncertainty is itself a planning problem. Utilities may either be underbuild and face reliability risks or overbuild infrastructure whose costs are ultimately shared with consumers.
Water presents a subtler trade-off. Many data centres use water-based cooling because it can reduce energy consumption compared with air cooling. Water is also consumed indirectly in electricity generation and semiconductor manufacturing. A low-carbon facility can still be problematic if it withdraws water from a stressed basin or competes with households, agriculture and ecosystems during drought. Conversely, eliminating water cooling may raise electricity use and emissions. A credible strategy must optimise carbon and water together, using local conditions rather than a universal rule.
Efficiency gains can also trigger a rebound effect. If a model becomes cheaper and more efficient, organisations may use it for far more tasks, increasing total consumption. Generative features are already being added to products where their value is marginal, partly because novelty drives adoption. Efficiency per query is useful, but the atmosphere and watershed respond to total energy, emissions and water—not to improved ratios alone.
Conventional sustainability yardsticks fall short
Corporate reporting has not kept pace with AI deployment. Many organisations do not know the energy or water associated with the models they buy through cloud services. Providers disclose facility-level metrics such as power usage effectiveness, but customers often lack workload-level information. Meanwhile, “carbon neutral” claims may rely on annual renewable matching or offsets that do not reflect the emissions produced at the time and place computing occurs.
There is also a boundary problem. A company may count electricity used to run an AI service while excluding the embodied carbon of servers, network equipment and data-centre construction. It may celebrate emissions avoided by an optimisation tool without establishing a credible baseline or checking whether behaviour changed elsewhere. Benefits are frequently modelled; costs are frequently partial. The result is a comparison tilted in AI’s favour before the analysis begins.
A further risk is substitution without transformation. An AI-generated sustainability report may be produced faster, but it does not reduce emissions unless it changes investment or operations. A sophisticated emissions forecast is not the same as decarbonisation. Organisations should distinguish between AI that measures, AI that recommends and AI that automatically changes a physical system. All can be useful, but their claims of impact should differ.
A practical framework for responsible deployment
Organisations can balance innovation and environmental responsibility by treating computing resources as a managed input rather than an invisible overhead. Five disciplines are especially important.
First, apply a materiality test before deploying AI. Define the problem, the environmental or business outcome and the simplest adequate method. A rules-based system, conventional analytics or a smaller specialised model may perform as well as a large general-purpose model. High-impact applications—such as grid optimisation or industrial efficiency—should receive priority over low-value automated content and speculative features.
Second, measure the full lifecycle. Procurement teams should request workload-level electricity, location-based emissions, water consumption and hardware lifecycle data from cloud and model providers. Assess training and inference, direct and indirect water, and embodied impacts. Report absolute totals alongside intensity metrics. For sustainability use cases, compare the footprint with verified avoided emissions, water or materials over an explicit time period, and disclose uncertainty.
Third, design for computational efficiency. Use the smallest model capable of meeting the required accuracy; compress or quantise models where appropriate; reuse existing models rather than retraining; limit unnecessary prompts and outputs; cache repeated results; and route simple requests to less intensive systems. Software architects should set carbon, energy and latency budgets just as they set financial and performance budgets. Efficiency must be paired with usage controls so that cheaper computation does not automatically create unlimited demand.
Fourth, make infrastructure location- and time-aware. Prefer data centres with efficient cooling, low-carbon grids and low water stress. Shift flexible training jobs to periods when clean electricity is abundant. Pursue hourly carbon-free energy matching rather than relying only on annual certificates, while supporting additional clean generation, storage and grid capacity. Water strategies should prioritise recycled or non-potable sources where feasible, disclose basin-level risk and establish drought-response plans developed with local communities.
Fifth, govern AI and sustainability together. Environmental review should be integrated into AI approval, procurement and product design—not added after deployment. Give senior leaders a dashboard covering total electricity, location-based emissions, water, embodied carbon, avoided impact and cost. Set escalation thresholds for energy-intensive use cases. Link sustainability targets to the teams choosing models and cloud regions and independently verify major claims.
From footprint accounting to system responsibility
Better organisational practice is necessary but not sufficient. Governments and utilities need transparent information about proposed data-centre loads, realistic connection requests and who pays for new generation and grid infrastructure. Planning rules should consider cumulative impacts on electricity, water and communities, not assess each facility in isolation. Common measurement standards would help customers compare providers and reduce selective disclosure.
The technology sector must also move beyond a narrow focus on data-centre efficiency. A highly efficient facility can still drive a large absolute increase in consumption. Providers should publish comparable data on model energy, emissions and water; extend hardware life where possible; recover materials; and give customers tools to choose lower-impact models, regions and operating times. Transparency should be granular enough to guide decisions without exposing security-sensitive information.
Most importantly, the social purpose of AI should remain visible. Sustainability is not simply a race to perform more computation with fewer resources. It is the effort to improve human wellbeing within ecological limits. That requires asking who benefits from an AI system, who bears its environmental costs and whether communities have a voice in decisions about land, water and energy.
The verdict: conditional acceleration
Artificial intelligence can accelerate sustainability, but it will not do so automatically. Its greatest value lies in helping people operate complex physical systems more intelligently: balancing grids, reducing industrial waste, anticipating hazards and making infrastructure more resilient. Its greatest danger lies in scale without discipline—ubiquitous deployment, opaque supply chains, fossil-powered growth, stressed watersheds and efficiency gains swallowed by rising demand.
The paradox is therefore manageable, but only if organisations abandon the idea that digital activity is immaterial. Every model choice is also an infrastructure choice. Every AI product has an opportunity cost measured in power, water, materials and capital. The responsible path is not to stop innovation; it is to direct scarce computational resources towards outcomes that justify their footprint, measure the consequences honestly and keep reducing both the impact per task and the total impact of the portfolio.
AI’s sustainability story will ultimately be written not by what the technology can do in theory, but by what institutions reward in practice. If success is defined as maximum adoption and computing growth, AI may deepen the problem it promises to solve. If success is defined as verified net environmental and social value, AI can become an important—though never sufficient—part of the transition to a more resilient, low-carbon economy.