Where Data Tells the Story
© Voronoi 2026. All rights reserved.
Here’s how much it costs to run a 1GW AI Data Center facility per year.
Epoch.ai's annualised cost breakdown for a 1 GW AI data centre puts the true all-in cost at $8.5 billion per year: $903M in OpEx and $7,608M in CapEx.
Inside CapEx, facility costs run $1,387M, and network infrastructure $1,167M. Servers alone: $5,021M. For every $1 hyperscalers spend running a 1 GW data centre, $8 goes into the chips.
The power story is the visible bottleneck, so it gets the coverage. The document says the real dependency is compute concentration, not power concentration, and it's more fragile than it looks: the entire cost structure assumes GPUs last 5 years. Shorten that to 3, and the annual cost jumps to $12B. Stretch it to 7, and it falls to $7B.
Hyperscalers aren't betting on power. They're betting on depreciation schedules for hardware whose obsolescence curve they don't control.