Training a single large AI model can consume as much electricity as hundreds of U.S. homes use in a year
Training and running AI models demands massive amounts of electricity, specialized hardware, and water-intensive cooling systems housed in sprawling data center facilities.
AI systems are energy-intensive because the mathematical operations that power them — particularly the matrix multiplications used in neural networks — must be performed billions or trillions of times, at speed, on specialized hardware that draws significant electrical power.
The process begins with training, in which a model learns patterns from enormous datasets. Training requires running computations across thousands of processors simultaneously, often for weeks. The hardware of choice is the graphics processing unit, or GPU, originally designed for rendering video game visuals but well-suited to the parallel calculations AI demands. A single large-scale training run can consume hundreds of thousands of kilowatt-hours of electricity — comparable to the annual consumption of hundreds of U.S. households. Purpose-built chips, including custom accelerators from companies like Google $GOOGL and Amazon $AMZN, are increasingly used alongside GPUs to improve efficiency. And chipmakers such as Qualcomm $QCOM are pushing further into data center AI silicon to meet that demand.
After training, inference — the stage at which a deployed model generates a response to a user query — adds a separate, continuous energy load. Because millions of queries arrive every day across services such as chatbots, search, and content generation tools, inference at scale accumulates into a substantial draw on the grid.
All of this computation must happen inside data centers: large facilities housing racks of servers connected by high-speed networking. Data centers require not only electricity to run the processors but also cooling systems to remove the heat those processors generate. Cooling can account for roughly 30 to 40% of a facility's total energy use. Water is commonly used in cooling loops, making data centers significant consumers of both electricity and fresh water.
The physical scale of the infrastructure required is considerable. A modern AI-oriented data center may occupy hundreds of thousands of square feet and draw tens or hundreds of megawatts of power — enough to strain local utility grids. Lawmakerss have begun scrutinizing how that grid demand is allocated between technology companies and ordinary ratepayers, and some utilities have responded by creating dedicated rate structures: Oregon's main utility, for example, introduced a higher electricity rate specifically for data centers while reducing rates for residential, commercial, and industrial customers. Data centers are also increasingly exposed to physical risk, as flooding, wildfires, and extreme heat threaten facilities whose siting decisions were made before climate hazards intensified.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
