AI’s environmental footprint is no longer just a carbon story. A United Nations University report published June 3 warns that AI-driven data center growth also carries water and land costs, complicating how governments, cloud buyers, and infrastructure operators evaluate sustainability claims.
Behind every AI workload is a data center, a power source, and a cooling system. As AI moves deeper into enterprise software, those infrastructure choices are becoming part of the cost, risk, and vendor-disclosure questions buyers have to weigh.
The environmental costs of AI infrastructure
The United Nations University report frames AI as a physical infrastructure system, not just a software layer. Data centers, chips, cooling systems, electricity grids, land, water, and mineral supply chains all shape the environmental cost of training and running AI systems.
Power demand is the clearest near-term pressure point. The International Energy Agency projects electricity generation for data centers will rise from 460 terawatt-hours in 2024 to more than 1,000 terawatt-hours in 2030. That pressure is already showing up in AI data center power demand, where utilities and cloud buyers are weighing whether local grids can support new campuses.
Carbon is only one part of the footprint. Lower-carbon electricity can still carry water or land costs, and a procurement claim focused on “renewable” or “net-zero” electricity may not show where a workload runs, how a site is cooled, or what pressure it adds to the local grid.
The report, as summarized by the Associated Press, estimated that global data centers used 448 trillion watt-hours of electricity in 2025. It also says AI accounts for about 20% of data center energy use and could reach 40% by 2030. AI infrastructure is expanding beyond chat tools. Recent AI factories built around robotics and simulation show how enterprise AI can depend on data centers, chips, sensors, networks, and industrial systems working together.
AI demand adds new vendor disclosure questions
Inference — the everyday use of deployed AI systems — can become a recurring infrastructure cost. AI embedded in support tools, search, coding assistants, analytics platforms, and security workflows creates ongoing demand for compute, power, and cooling. Recent agentic and multimodal AI rollouts show how workloads can move from text prompts into actions, images, video, and background tasks.
A Google-authored paper reported a 33-fold reduction in energy use and a 44-fold reduction in carbon footprint for the median Gemini Apps text prompt between May 2024 and May 2025. Outside experts have argued that per-prompt estimates can still leave out broader system impacts, including indirect water use and location-based emissions.
That leaves a disclosure gap for procurement teams. Facility-level reporting on energy use, water consumption, emissions, and land impacts is not yet standardized across AI infrastructure providers. As a result, a “100% renewable” claim may address electricity procurement without answering separate questions about local water use, land impact, or grid strain.
The UNU report calls for governments to include AI infrastructure in energy, water, and land-use planning while requiring more standardized environmental reporting. For enterprise buyers, the same issue is becoming part of vendor evaluation: where workloads run, how facilities are powered and cooled, and what environmental data providers disclose.
Until that reporting improves, cloud buyers may be able to compare performance and price more easily than the energy, water, and land costs behind those workloads.
Also read: As AI infrastructure costs climb, OpenAI’s profitability challenge shows why compute demand is becoming a business-model problem as much as a technical one.