Google Reports 37% Rise in Electricity Use Driven by AI Buildout

Google released its 11th annual Environmental Report on June 30, 2026, covering 2025 performance. The company reported a 37% year-over-year increase in electricity demand, its largest annual load growth in history, bringing cumulative growth in electricity use to more than 250% since 2019. Google said it matched 100% of its electricity consumption with renewable energy purchases for the ninth consecutive year and reduced operational (Scope 1 and 2, market-based) emissions by 2% year-over-year, while signing agreements for more than 12 gigawatts of net-new clean energy in 2025 alone - including restarting Iowa's 600 MW Duane Arnold nuclear plant with NextEra Energy, a 3 GW hydropower framework with Brookfield, and a fusion power-purchase agreement with Commonwealth Fusion Systems. Supply-chain (Scope 3) emissions grew 25% year-over-year, however, with data center construction alone contributing roughly 2.3 million tons of CO2-equivalent; Google attributes this largely to semiconductor suppliers operating on grids in Taiwan, Japan, Vietnam, and India that have limited access to clean power. Google's report states plainly: "the path to achieving our climate ambitions will not be linear - given our AI infrastructure buildout is currently accelerating faster than the grid is decarbonizing."

Google also disclosed operational efficiency figures that matter for anyone benchmarking AI infrastructure: a fleet-wide average power usage effectiveness (PUE) of 1.09 for 2025, and a claim that the energy footprint of a median Gemini text prompt fell 33-fold over the past 12 months even as total consumption kept rising. The company also said it integrated 1 gigawatt of demand-response capacity into long-term US utility contracts, letting it shift or curtail machine-learning workloads during periods of grid stress - a concrete example of workload scheduling being used as an emissions and grid-reliability lever, not just a cost one.

The operational-versus-supply-chain emissions split is the most transferable lesson here: a company can decouple its direct (Scope 1/2) footprint from compute growth through large renewable-energy procurement, but indirect (Scope 3) emissions embedded in chip and data center supply chains can still rise sharply. Teams reporting on the sustainability of their own AI workloads should track both categories separately rather than citing only operational-emissions improvements, and should treat demand-response and workload-shifting capability as a legitimate infrastructure investment, not just a compliance checkbox.

Whether other hyperscalers publish comparable Scope 1/2/3 splits and PUE figures for 2025, how fast semiconductor-supply-chain grids in Asia decarbonize relative to Google's own procurement pace, and whether demand-response commitments like Google's 1 GW US program scale enough to materially reduce peak grid strain from AI training runs.