Deploying AI servers across the United States could add between 24 and 44 million tonnes of carbon dioxide equivalent annually by 2030, alongside a water footprint exceeding 700 million cubic metres a year. This came from bottom-up modelling by Xiao et al., 2025 in Nature Sustainability, led from Cornell, KTH Royal Institute of Technology and Politecnico di Milano. They found that best industry practice cuts only 73% of residual emissions, leaving the sector short of its own 2030 net-zero pledges without heavy reliance on offsets and water credits of uncertain durability.
The dominant policy narrative treats this as a capacity problem, not a structural one. The International Energy Agency's Energy and AI (2025) report projects that renewables will supply most new data-centre electricity through 2035, with fossil fuels covering as little as 15% of the increment. Gui & Dai, 2026 complicate this optimism with a game-theoretic model of the "power couple" framing itself. When the market rewards capability roughly as fast as scaling raises its energy cost, developers push toward larger models regardless of whether the marginal megawatt-hour is clean, so renewable expansion relaxes the scaling constraint rather than limit fossil generation. The authors call this an adaptation trap, as climate damages rise, so does the value of AI-enabled adaptation, strengthening the incentive to keep scaling on whatever power is available.
The divergence is not neutral. Muller, 2026, in an NBER working paper from Carnegie Mellon, translates the emissions and air pollution from US data centres into a dollar figure, finding damages of roughly 25 billion USD a year, a cost currently borne by surrounding communities rather than by the sector itself. Read against Gui and Dai's logic, this is not a lag to be closed by better siting; it is a symptom of the same mechanism, since developers facing no penalty for the marginal tonne of carbon have little reason to slow their scaling.
It is important however to also consider opposite tendencies, such as those documented by Ding et al., 2024 at Lawrence Berkeley National Laboratory, whose Nature Communications study found AI-driven building management cuts commercial energy and carbon use at scale by double digits, without new construction. The asymmetry is the point. AI's capacity to save energy in buildings, grids and industrial processes is real, but it operates at the margin of an existing stock, while its own demand for power compounds continuously as models scale. Whether AI becomes a net accelerant or a net drag on the energy transition depends less on the technology than on whether policy makes the cost of the marginal tonne of carbon binding on the sector that creates it.