Tool · Site planner

What site does your cluster need?

Pick the GPUs and how many. See the power, cooling, rack density and space it takes, and what kind of building can host it.

GPU platform
Nodes 64 nodes · 512 GPUs

8 GPUs per node. Planning more than 5,000 GPUs? Start a build →

Cooling
Facility
Facility power1.4 MW1.0 MW IT × PUE 1.40
Rack density29 kWHigh density
Racks3732 GPU · 5 network & storage
White space1,110 sq ftIncl. aisles
Heat to reject291 tons1.0 MW all to air
Power bill$1.0M / yr12,571 MWh at max draw

Site class

Colocation hall or wholesale suite

A dedicated room or suite, usually leased by the MW.

  • Committed MW in a single hall
  • N+1 or 2N UPS and generators sized to the load
  • Liquid-cooling loop in place if the racks need it

Typical lead time 1 to 6 months where capacity exists

Cooling and density · 29 kW per rack

Needs hot or cold aisle containment, and often rear-door heat exchangers. Many older halls cap well below this.

Sized for max draw, because training runs close to it and a site usually outlives its first workload. Average draw, and so the power bill, runs lower: training typically averages 75–85% of max, inference 55–65%.

Planning estimates from public spec sheets, at each system’s maximum rated power, plus 12% for networking, storage and management. Real draw is often lower; real sites add redundancy on top. Use this to frame the conversation with a site or an engineer, not as a design.

Need a site, or the hardware to fill one?Start a build