AI Data Center Requirements for GPU Clusters

Whether the site can keep the equipment running.

What it covers

GPU racks draw 40-80 kW while most data centers run 10-30 kW, and the Uptime Institute found in 2025 that 67% of existing data centers cannot support modern GPU densities. Site evaluation checks that the building can keep the GPUs running at the uptime the contract requires.

What you should check for

PowerkW per rack, committedThe signed order covers the systems' draw. Two DGX B300s need 29 kW; a GB300 NVL72 rack up to 140 kW. See power and cooling readiness.
Energized, with a date in writingThe colo's power date, re-confirmed each quarter.
Backup power and UPSGeneration and fuel on site, UPS separate from generation, transformers set back with emergency cutoffs: the same questions an insurer asks.
CoolingCooling matched to densityPast roughly 40 kW per rack, air cooling stops working and direct-to-chip liquid cooling becomes mandatory.
Cost of the cooling choiceAn air-cooled site at a PUE of 1.8 draws 18 MW for 10 MW of servers; liquid cooling at 1.15 draws 11.5 MW. At $0.07/kWh the gap is about $4M a year.
Resilience and accessTier against the SLAUptime Tier I and II allow 23-29 hours of downtime a year, Tier III about 1.6; check the registry.
Recovery accessA named facility contact, hands on site, control of the cage, and a plan to decommission or re-contract the cluster in place.
Fire and securitySuppression in the data hall, monitored alarms, cabinet locks, cameras and perimeter: the facility fact sheet the property insurer will want anyway.

Sources: NVIDIA, GIGABYTE, American Compute.

The market is splitting by power density

Some facilities can only run older, air-cooled equipment like Hopper; some can run Blackwell; a limited handful can adopt Rubin, whose 2027 racks are designed for as much as 600 kW. An evaluation places the site in one of those groups and checks it against the GPUs being financed.

Time is the other constraint. A new data center takes 18-36 months, a substation 2-4 years, and power delivery in Northern Virginia about 7 years, up from 12-18 months. A site that is not energized at signing is a schedule risk, not a site.

Many sites lack what customer contracts require, especially retrofitted crypto mines and modular builds. Those can work; they need the checks below done on paper and on site.

Red flags

Power "expected" at signingOver a third of the 110 data center projects due in 2025 missed their dates; power is the bottleneck as of early 2026.
Retrofitted mine without redundancyMining tolerates outages; a GPU offtake contract with an uptime SLA does not.
Air-cooled hall for liquid-cooled GPUsThe cluster cannot run at full power until a retrofit nobody has priced.

Discuss a transaction

Send us the equipment list and the loan or lease you have in mind. We will come back with an indication and the questions we would need answered.

Prefer email? hello@amcompute.com