Data Center Energy Consumption

In the AI Story

Data center energy consumption divides into computational load (servers performing useful work) and overhead (cooling, power conversion, lighting, physical security). The power usage effectiveness (PUE) ratio measures total facility energy divided by IT equipment energy—a perfectly efficient data center would have PUE of 1.0, meaning zero overhead. Modern hyperscale facilities achieve PUE of 1.1-1.2, meaning 10-20% overhead, a dramatic improvement from the industry average of 2.0-2.5 in the early 2000s. But PUE improvements address overhead, not the computational base load. As AI workloads shift the computational intensity upward—more floating-point operations per transaction, more memory bandwidth, higher GPU utilization—the base load grows faster than PUE improvements reduce overhead. Total facility energy consumption increases even as efficiency metrics improve.

The geographic distribution of data center energy demand reflects the economic geography of technology: northern Virginia's 'Data Center Alley' hosts over 300 facilities consuming roughly 25% of the state's electricity. Central Oregon attracts data centers through cheap hydroelectric power. Iceland, Ireland, and Singapore compete for facilities using renewable energy availability, favorable climates (reducing cooling costs), or connectivity to submarine cables as attractions. But geography that favors data centers often creates local resource conflicts—Oregon communities protested Google's water consumption, Arizona residents challenged Microsoft's aquifer draw, Irish grid operators warned that data center growth threatened electricity reliability for residential and industrial users. The concentration reflects rational corporate site selection; it produces localized environmental and political stress.

The carbon intensity of data center operations depends entirely on the energy mix of the serving grid. A facility in Quebec drawing hydroelectric power has near-zero operational carbon emissions. An identical facility in West Virginia drawing coal-fired electricity has carbon intensity hundreds of times higher. Corporate renewable energy purchases—power purchase agreements committing to add renewable capacity equivalent to consumption—reduce grid-average carbon intensity but do not guarantee the specific electrons consumed are renewable. The accounting distinction matters: a company can be 'carbon neutral' on an annual net basis while drawing fossil electricity every hour its data centers operate, because renewable generation and consumption occur at different times and places. The gap between marketed commitments and thermodynamic reality is where greenwashing lives.

Projections of future data center energy demand carry unavoidable uncertainty because they depend on the interaction of three variables moving at different speeds: computational demand (growing rapidly), efficiency improvements (significant but slower), and grid capacity additions (slowest of all). The IEA's 2026 projection of 1,000+ TWh assumes current growth rates moderate; if they don't, the figure could reach 1,200-1,500 TWh. If efficiency improvements accelerate through better chips, better algorithms, or better cooling, the figure could be lower. The range matters because 1,000 TWh represents roughly 3% of projected global electricity demand, while 1,500 TWh approaches 4.5%—the difference between a large new demand category and one of the largest sectoral demands in the global energy system, comparable to all global aviation or all steel production.

Origin

Data center energy tracking began in earnest in the mid-2000s when facilities transitioned from small server rooms to hyperscale campuses. Jonathan Koomey's influential studies (2007-2011) established the methodology for estimating aggregate consumption from bottom-up facility counts and top-down electricity statistics. Eric Masanet's 2020 recalibration in Science corrected earlier overestimates, finding that efficiency improvements had kept total growth below projections—a finding that provided temporary reassurance before AI workloads triggered renewed acceleration.

Smil's engagement with data center energy appears across his writing on electricity systems and his 2026 Bankinter presentation specifying that U.S. grids need approximately 50 GW of new capacity by 2030. The figure synthesizes utility planning documents, corporate announcements, and his career-long documentation of how long it actually takes to add generating capacity. His framework insists that data center energy is not a niche technical concern but a major new variable in national energy policy, requiring the same infrastructure planning rigor as electrification of transport or industrial decarbonization.

Key Ideas

Doubling trajectory. Global data center electricity consumption projected to more than double from 460 TWh (2022) to >1,000 TWh (2026), with AI workloads as primary driver—unprecedented four-year acceleration in a mature industrial sector.

Grid share concentration. Data centers exceed 10% of electricity supply in six U.S. states and 25% in Virginia; concentration creates local grid stress, price impacts, and resource allocation conflicts invisible in national aggregates.

Efficiency-demand race. PUE improvements and chip efficiency gains are real but insufficient to offset computational demand growth; total consumption increases even as consumption per operation decreases—textbook Jevons rebound.

Carbon accounting complexity. Renewable energy purchases create accounting appearance of carbon neutrality while facilities continue drawing from fossil-heavy grids; the gap between annual net accounting and hourly operational reality obscures true climate impact.

Fifty-gigawatt requirement. Smil's specific estimate that U.S. grids need ~50 GW of new capacity by 2030 for AI growth—equivalent to fifty million-person cities—represents construction challenge without recent peacetime precedent.