A new review article in Nature Reviews Clean Technology examines the growing environmental footprint of AI data centers and outlines strategies to reduce their carbon emissions, water consumption, and grid impact. The authors project data center power demand could increase by roughly 650 terawatt-hours per year between 2023 and 2028, driven by AI workloads. They find that embodied emissions from chip fabrication and construction account for over half of total emissions at large AI facilities, and that inference tasks can contribute 40 to 60 percent of a model's lifetime CO2 equivalent emissions. The review covers several mitigation approaches. Using recycled or older hardware can cut embodied emissions by 10 to 20 percent. Grid integrated data centers that shift training to periods of high renewable output can reduce carbon intensity by about 10 percent in grids with strong renewable penetration. The authors also highlight trade offs between water use and carbon emissions, noting that efforts to cut water consumption can sometimes increase CO2 output. The piece is a solid overview for anyone tracking how AI growth interacts with grid decarbonization and resource constraints.
