An AI data center site has to deliver more power to less floor space than any facility before it. A single NVIDIA GB200 NVL72 rack draws about 120 kW and needs liquid cooling. What changes for site selection is the power: AI tenants need firm service, a fast ramp and room to grow, often from tens of MW to hundreds. Invariant matches AI buyers to sites with documented power, from energized 10 MW inference sites to a 1.5 GW campus.
Before you start
- A GB200 NVL72 rack draws about 120 kW. 1,000 GPUs need roughly 2 MW at the meter including cooling.
- AI tenants need firm power. Curtailable mining rates do not work for most AI contracts.
- Plan for liquid cooling and heat rejection outdoors; land per MW falls less than rack density suggests.
- Invariant lists AI-ready sites from 10 MW inference to 1.5 GW training campuses.
Power per GPU
At about 120 kW per 72-GPU rack, each GPU draws roughly 1.7 kW at the rack. Add cooling and power losses at a PUE of 1.3 and 1,000 GPUs need about 2.2 MW at the meter. A 100,000 GPU cluster needs over 200 MW. Training and inference differ: inference fits in 5 to 50 MW sites near users, training wants hundreds of MW in one place.eeNews EuropeUptime Institute
Firm power and redundancy
Many energized sites on the market were built for bitcoin mining on interruptible rates. AI tenants usually need firm service with backup. Read the power contract before pricing a conversion, and budget for generators, UPS and a second feed where the tenant requires them.
Cooling and land
Liquid cooling shrinks the building per MW but still rejects heat outside through dry coolers or towers. Substations and cooling yards scale with MW, so an AI campus still needs roughly 0.7 to 2 MW per gross acre.
Sites available through Invariant now
A selection from Invariant’s current US inventory that fits this search. Specs are as marketed by the owner and are checked against utility documents before any introduction.
| Region | Power | Land | Timing |
|---|---|---|---|
| Arizona | 1.5 GW Phase 1 | 15,000+ acres | Grid 2027 |
| Pennsylvania | 350 MW, signed utility service agreement | Former plant campus, lease | 50 MW end 2026, 300 MW end 2027 |
| Ohio | 300 MW approved | 300+ acres | 1H 2027 |
| Indiana | 200 MW, load study done | ~80 acres | Substation planned |
| Texas metro | 44 MW, +250 MW | ~50 acres, 1M sq ft | Now |
| Oregon | 10 MW hydro, to 100 MW+ | Lease + land option | Now |
| US (3 facilities) | Tier III data centers, 3.5 to 10 MW | Built | Now |
Exact locations, the utility documents behind each MW figure and pricing are shared on a call. We also hold sites that are not listed publicly; owners ask for discretion and a site that is shopped everywhere loses value.
Browse data center sites by size, state and type
Questions that come up.
How much power does an AI data center need?
About 2.2 MW at the meter per 1,000 Blackwell-class GPUs at a PUE of 1.3. Inference sites run 5 to 50 MW; training campuses run hundreds of MW to gigawatts.Uptime Institute
Can a bitcoin mining site become an AI data center?
Often, if the power is firm or can be made firm. The utility service and land carry over; most buildings and electrical systems get rebuilt for AI.
How do I get matching sites from Invariant?
Book a call or call (628) 220-8972 with your MW, states and energize-by date. We send matching sites, including off-market ones.
Sources and references.
Site specifications are as marketed by the seller and are checked against utility documents before any introduction. Public sources checked on . The location, equipment, and operating plan determine which requirements apply to a project.