Data center facilities designed for high-density GPUs.
Blackwell systems draw more power and give off more heat per rack than the servers most data centers were built for. We are designing the power, cooling, and network around them from the start.
Design targets.
These are design targets, not measured results. The rack figure derives from NVIDIA's published maximum system power for HGX B200.
About 60 kW per GPU rack
Each GPU rack is sized for four HGX systems.
Two power paths to every rack
The two paths are independent of each other.
Four networks
Compute, storage, management, and edge traffic each have their own network.
Liquid and air cooling
Direct-to-chip liquid cooling where the system supports it, and air handling for the rest.
Power and cooling built for sustained load.
Training jobs run GPUs near full power for weeks. The electrical and mechanical design assumes that from day one.
Generation on site, surplus to the grid
On-site natural gas generation sized above our peak demand, two independent power paths to every rack, and UPS, with surplus electricity sold to the local grid.
Cooling sized for Blackwell
Direct-to-chip liquid cooling for the GPU systems that support it, and air handling for the rest, sized for sustained full load rather than average load.
Maintainable while running
Every power and cooling component can be serviced or replaced without taking customer systems offline.
Instrumented end to end
Power draw, inlet temperature, and coolant flow are measured per rack and exposed to customers with their cluster telemetry.
Four networks, each with one job.
Keeping traffic types on separate networks keeps performance predictable and limits the reach of any fault.
Compute fabric
Non-blocking GPU-to-GPU network between systems, sized to the per-GPU bandwidth of the platform.
Storage network
A separate high-throughput network to the parallel file system and object storage, so data loading never competes with collective operations.
Management network
Out-of-band access to every system for provisioning, monitoring, and recovery, isolated from customer traffic.
Edge and private links
Internet transit from multiple carriers, plus cross-connects and VPN into customer networks.
Layered access from the fence line to the rack.
Physical controls are designed to support the audit requirements on our compliance roadmap.
- Perimeter fencing, vehicle barriers, and monitored entry points
- Badge and biometric access at each security layer
- Mantrap entry to the data hall
- Locked cages and cabinets for customer equipment
- Video coverage of entrances, aisles, and loading areas, with retained footage
- Escorted visitor access, logged and reviewed
Where we build.
We look for places with grid capacity to spare, water we will not compete with residents for, and local leaders who want the jobs and tax base on terms they can see in public.
Before we buy land, we hold public meetings. Before we build, we sign a community agreement that carries our commitments on power, water, sound, and local hiring.
How we build sitesWe will publish efficiency figures once they are measured.
Power usage effectiveness and water use depend on how a facility actually runs. We will publish both once the facility is operating and has a full season of data.