Gridvara maps the physical footprint of the AI buildout — every tracked data center and GPU cluster, sized by power draw, from megawatt research clusters to gigawatt frontier campuses.
Gridvara is an interactive atlas of AI computing infrastructure. Each point on the 3D globe is a physical facility — who owns it, who uses it, how many chips it runs, how much power it draws, and what it cost to build. The largest frontier campuses have full profiles with construction timelines, hardware breakdowns, and Epoch AI's projections of where they're headed.
The data comes from Epoch AI's open Frontier Data Centers and GPU Clusters databases (CC-BY 4.0), built from satellite imagery, construction permits, and public disclosures. Epoch estimates this covers roughly 15–20% of global AI compute — the real buildout is larger still.
Each line is one frontier campus's IT power over time. Solid where Epoch AI has observed capacity; dashed where the line runs into projections of announced expansion.
The steepness is the story: capacity that took the cloud industry a decade to assemble is now added by single sites in quarters.
Tracked IT power grouped by the company that owns or is the primary user of each facility. "Others" collects every operator below the frontier tier — hundreds of smaller research and cloud clusters.
The largest five operators account for 77% of all IT power Gridvara tracks.Computed from Epoch AI open data
Dated events from campus construction timelines — status changes and step increases in power. Each links to the full profile.
Six headline figures from Epoch AI's Trends in Artificial Intelligence dashboard.
The total computing power of the world's stock of AI chips is growing 3.4× per year — doubling roughly every 7 months.
Learn more ↗The largest known AI data center — xAI's Colossus 2 in Memphis — packs computing capacity equivalent to 1.1 million NVIDIA H100 chips.
View its profile →Gigawatt-scale AI data centers — the size of a mid-sized city's power demand — are being built in about two years.
Learn more ↗A typical AI data center with 1 gigawatt of IT power costs roughly $38 billion in up-front capital expenditure.
Learn more ↗Compute used to train frontier language models has grown 5× per year since 2020 — a ~10,000× increase across the period.
Learn more ↗AI chip performance per dollar has improved by 37% per year — doubling every 2.2 years across two decades of accelerators.
Learn more ↗Ranked by current IT power. Switch country to see where each grid is carrying the load.