Infrastructure

Data centers designed for 130 kW racks, not retrofitted for them

Every GREENAIBOX facility was built from the slab up for liquid-cooled GPU systems, sited next to renewable generation and connected by a private backbone. Here is what sits underneath your cluster.

Talk to our team Sustainability
NVIDIA DGX SuperPOD built from DGX GB200 rack-scale systems Data hall Image: NVIDIA
130 kW
Per-rack power density with direct-to-chip liquid cooling
1.10
Power usage effectiveness across GPU halls
130 MW
Contracted renewable capacity across three live regions
2N
Power distribution to the rack, with N+1 cooling and network, concurrently maintainable
Regions

Where your GPUs live

Three live regions on two continents, each colocated with its power source. Choose by latency, data residency or hardware generation.

us-west-2

Columbia Basin, Washington

Available
  • Power sourceHydroelectric
  • Capacity60 MW
  • GPUsB200, H200, H100, L40S
  • Latency5 ms to Seattle, 20 ms to San Francisco
ca-east-1

Quebec, Canada

Available
  • Power sourceHydroelectric
  • Capacity40 MW
  • GPUsB200, H200, H100, A100
  • Latency9 ms to Montreal, 18 ms to New York
eu-north-1

Northern Sweden

Available
  • Power sourceHydro and wind
  • Capacity30 MW
  • GPUsB200, H100, GH200
  • Latency12 ms to Stockholm, 30 ms to Frankfurt
us-central-1

Texas Panhandle

Q2 2027
  • Power sourceWind and solar with storage
  • Capacity100 MW planned
  • GPUsB300, GB300 NVL72
  • Latency7 ms to Dallas
Facility design

Built for the next three GPU generations

Liquid cooling everywhere

Direct-to-chip cold plates on every GPU, CPU and NVSwitch, with facility water loops at 32 to 40 C. Rear-door heat exchangers handle the remainder. No chillers most of the year.

Power you can plan around

Dual utility feeds, 2N distribution to the rack, UPS with 15 minutes of ride-through and generators with 72 hours of fuel. Long-term PPAs fix the price of energy for the life of your contract.

Physical security

Perimeter fencing, mantraps, biometric access, 24/7 staffed security and CCTV with 90-day retention. Private cages for custom cluster customers with their own access lists.

Non-blocking fabrics

NVIDIA Quantum-2 and Quantum-X800 InfiniBand in rail-optimized fat trees, sized so any cluster of up to 4,096 GPUs is non-blocking end to end.

Private backbone and peering

Diverse private wavelengths between regions, 100 Gb/s+ peering at major exchanges, and dedicated interconnects to AWS, Azure, Google Cloud and your own sites.

Operations that notice first

Every GPU, link, pump and breaker is instrumented. Anomaly detection flags a degrading HBM stack or a fabric link before it fails a job.

Inside the facility

Halls built around the rack, not the other way round

Row of NVIDIA DGX H100 systems in a DGX SuperPOD Containment row Image: NVIDIA
HGX H100-class systems in a SuperPOD row.
Illustration of a coolant distribution unit with supply and return pipes feeding a row of racks Cooling
Row-level coolant distribution unit. Warm-water loop, no evaporative cooling.
Illustration of a liquid-cooled rack with eight HGX nodes, two InfiniBand switches and a coolant distribution unit HGX rack
One rack: 8 HGX nodes, 64 GPUs, top-of-rack InfiniBand and an in-rack CDU.
Colocation

Your hardware, our halls

Custom cluster customers can deploy in a private cage in any region and get the same liquid cooling, renewable power and backbone we use for our own fleet, at our power purchase rate.

Custom Clusters
Colocation at a glance
  • Rack densityUp to 130 kW liquid, 40 kW air
  • Power pricingPass-through PPA rate
  • Cross-connectsInfiniBand, 400G Ethernet, dark fiber
  • Remote hands24/7, included in operations
  • Minimum1 rack, 12-month term

See it for yourself

We host customer site visits in every region. Bring your infrastructure team and kick the tires.

Request a site visit Trust and security