AI Servers to Draw 258 Terawatts by 2027, Strangling Grid Capacity

Global data center electricity consumption is projected to surge 26% this year to 565 TWh, with AI-optimized servers expected to consume more power than all conventional data center hardware combined by 2027, according to Gartner forecasts cited by Tom's Hardware.

The increase is significant: AI servers consumed approximately 95 TWh in 2025; Gartner projects this to rise to 175 TWh in 2026—an 84% year-over-year increase—and further to 258 TWh in 2027. This would make AI-optimized hardware responsible for 31% of total data center power this year, up from around 20% in 2025. Meanwhile, conventional server power demand is nearly stagnant, increasing only 1.2% to about 195 TWh this year and remaining near 200 TWh through 2027. Cooling follows a similar trend: denser AI racks will push cooling electricity up 22.6% to 195 TWh in 2026, nearly equaling the power draw of all traditional servers globally. By 2030, AI-optimized hardware alone is forecast to account for roughly half of all data center power consumption.

Infrastructure teams now face power availability as the primary constraint on AI scaling, with regional grid health directly impacting hyperscaler service level agreements. The United States alone accounts for 204 TWh of the global total—36%—with dedicated AI data centers consuming about 68 TWh of that. However, grid capacity is not keeping pace; Gartner notes that over 75 data center projects valued at $130 billion were halted in early 2026 due to power and water access disputes. In Virginia, data center load has driven utility rates high enough that one county asked government employees to conserve electricity.

Architects are adapting by bypassing the grid. Some operators are deploying on-site gas generators to avoid interconnection queues, trading carbon targets and fuel logistics for time-to-power. Hyperscalers are also signing substantial offtake deals—Meta secured 6 GW of nuclear capacity for upcoming facilities, and one developer is repurposing retired U.S. Navy reactors for a Tennessee AI site. However, these projects are years away; recommissioned plants and the earliest small modular reactors are not expected online until 2028 or later. Until then, gas and liquid fuel remain the only near-term options for stranded compute, introducing emissions compliance and supply chain risk into platform operations.

The Gartner forecast includes supply shortages, geopolitical disruption, and cancelled projects, yet still projects total data center consumption exceeding 1,200 TWh by 2030—a level that could outstrip available grid supply. Worldwide power demand is projected to hit 132 GW this year, up from 104 GW in 2025. Analyst Linglan Wang described data center power security as the new battleground for scaling and protecting margins. This leaves platform leads with a shrinking window to secure interconnection agreements, redesign facilities for liquid cooling, and model rack-level power draw before the queue lengthens further.