This started as a prompt asking how the power draw of a modern AI data center compares to a typical one. That question turned into a rabbit hole: pulling the numbers, building the visualizations below, and editing all of it into this post.
The United States spent three decades building roughly four thousand data centers. The AI buildout is about to match that scale with a handful of sites. Scroll to see the ratio.
The field
There are a little over 4,000 data centers in the US today, with a combined capacity north of 50 gigawatts. Most of them are unremarkable by AI standards: enterprise colocation halls, cloud regions, telecom facilities, drawing a few megawatts apiece. Call it 5 MW as a rough median. That's the dot in the visualization above, an illustrative approximation rather than a facility-by-facility count.
That fleet took the better part of thirty years to build. It's the accumulated infrastructure of the entire internet economy: banking, retail, streaming, enterprise software, government systems, all of it.
One campus
A single frontier AI campus now draws as much power as roughly 200 of those facilities combined. OpenAI's Stargate site in Abilene, Texas, built with Oracle, SoftBank, and infrastructure partner Crusoe, is the clearest example: about 0.3 GW operating today, on a path to 1.2 GW as more of its buildings come online. xAI's Colossus 2 in Memphis is moving even faster, already drawing close to 1 GW roughly a year after breaking ground.
Both look more like industrial power projects than conventional data centers. Several of the newest sites pair with on-site natural gas turbines to skip interconnection queues, and most are committing to closed-loop liquid cooling to blunt concerns about water use.
The land alone tells the story. Once its eight buildings are finished in mid-2026, Stargate's Abilene campus will cover about 875 acres, close to the footprint of Central Park. Meta's Hyperion campus in Holly Ridge, Louisiana, is already several times larger.
A traditional server rack draws about 10 kW. The NVIDIA GB200 NVL72 racks filling these campuses draw about 120 kW in the same footprint, over ten times as much. Air cooling can't keep up at that density, which is why every one of these sites is built around direct liquid cooling from the start.
Five in 2026
By Epoch AI's count, five campuses are set to cross the 1 GW threshold this year alone:
- Anthropic-Amazon, New Carlisle, Indiana: January 2026
- xAI Colossus 2, Memphis, Tennessee: February 2026
- Microsoft, Fayetteville, Georgia: March 2026
- Meta Prometheus, Ohio: May 2026
- OpenAI Stargate, Abilene, Texas: July 2026
Five sites, five gigawatts, roughly a thousand equivalent "typical" data centers worth of load, about a quarter of everything the country built before them, added in a single year.
A gigawatt is easiest to picture next to a power plant: one large reactor unit produces about that much. The five 2026 campuses match five reactors' worth of new demand. The entire existing US fleet, built up over thirty years, comes in around fifty.
What it strains
The interconnect matters more than the GPU count. EPRI's most recent modeling puts data centers at 4 to 5 percent of US electricity demand today, headed toward 9 to 17 percent by 2030 on current trajectories. Utilities that used to plan load growth in single-digit percentages a decade ago are now fielding requests for gigawatt-scale single customers, sited wherever land, water rights, and transmission capacity happen to line up. Texas, and increasingly the Ohio Valley and the Southeast, are where the map keeps pointing.
One caveat
The dot count and per-facility wattage here are deliberately approximate, a way to make an abstract capacity number physically legible rather than a facility-by-facility audit. The five 2026 campuses, their dates, and the footprint and rack figures are real and sourced below. The "4,000 dots" backdrop is a stylized read of the broader US fleet, not a registry.
Sources
- Epoch AI, "OpenAI Stargate: where the US sites stand"
- The Data Center Engineer, "Five AI data centers to reach 1 GW power capacity in 2026"
- Epoch AI, "The largest AI data center campuses"
- EPRI, "Powering Intelligence 2026"
- Yahoo Finance, "Stargate's First Data Center Site is Size of Central Park"
- Fortune, "Meta's Hyperion AI data center will sprawl to four times the size of Manhattan's Central Park"
- NVIDIA, "GB200 NVL72"
- Sunbird DCIM, "Is Your Data Center Ready for the NVIDIA GB200 NVL72?"