The Current Generation, And The Next
NVIDIA Blackwell Ultra, Hopper, and the Rubin generation, delivered and operated as productive cluster capacity.
Radiant is an NVIDIA Cloud Partner, building AI Factory clusters to the NVIDIA Cloud Partner reference architecture. Every cluster is designed across compute, networking, and storage to operate as a validated, high-performance system rather than a collection of integrated parts.
Blackwell Ultra - GB300 NVL72
Rack-scale, fully liquid-cooled. For frontier inference, long-context reasoning, and high-throughput serving. Deployable across Radiant bare metal, GPU instances, Kubernetes, and Slurm.
72 Blackwell Ultra GPUs, 36 Grace CPUs per rack
288 GB HBM3e per GPU
ConnectX-8 SuperNIC, 800 Gb/s per GPU
Quantum-X800 InfiniBand or Spectrum-X Ethernet
FP4 precision
Hopper - H200
Production-mature for enterprise AI, fine-tuning, inference, simulation, and HPC. Deployable across Radiant bare metal, GPU instances, Kubernetes, and Slurm.
141 GB HBM3e
4.8TB/s
NVLink
Rubin - Vera Rubin NVL72
The next generation, for reasoning-heavy, high-token-volume workloads at rack scale. Rubin-ready across bare metal, GPU instances, Kubernetes, and Slurm.
HBM4
13 TB/s
NVLink 6
One fleet across every service
The same physical GPU capacity is consumed as dedicated clusters, tenant-partitioned instances, Kubernetes pools, or Slurm queues. Allocation is operated from one layer, so capacity shifts between consumption models without rebuilding the cluster underneath.
Delivered and kept productive
GPUs are onboarded, tested, provisioned, monitored, and maintained through FlightDeck, with node health tracked in real time across regions.
