B300 for frontier work
Large-scale training and reasoning inference at production density. Scales to 144 GPUs per ORV3 liquid-cooled rack.
Platforms · pricing on application
Five platforms. Priced in writing.
Three rack-scale 8-GPU server systems and two single-card accelerators. We do not publish prices because Blackwell pricing moves weekly. Send the configuration and quantity and you get a written number, valid for a stated period.
01Catalogue
Platforms and configurations.
| Platform | Configuration | Price |
|---|---|---|
| Server platforms, 8-GPU rack-scale systems | ||
| HGX B300 EnterpriseBlackwell Ultra, 8U air or 4U liquid | 8× NVIDIA B300, 288 GB HBM3e per GPU (2.3 TB total). 5th-gen NVLink 1.8 TB/s. Embedded ConnectX-8 SuperNICs up to 800 Gb/s. 2× Intel Xeon 6, up to 8 TB DDR5. | On application |
| HGX B200 EnterpriseBlackwell, 4U, 8U and 10U variants | 8× NVIDIA B200, 180 GB HBM3e per GPU (1.4 TB total). 5th-gen NVLink 1.8 TB/s. 8× ConnectX-7 at 400 Gb/s. Up to 15× faster real-time LLM inference than H100-class, per NVIDIA. | On application |
| HGX H200 EnterpriseHopper, 8U air-cooled | 8× NVIDIA H200, 141 GB HBM3e per GPU (1.1 TB total). 4.8 TB/s memory bandwidth. 4th-gen NVLink 900 GB/s. 8× ConnectX-7 at 400 Gb/s. | On application |
| GPU accelerators, single card | ||
| NVIDIA H200 NVLPCIe, memory-bound inference | 141 GB HBM3e, 4.8 TB/s. PCIe 5.0, 600 W. 2- or 4-way NVLink. | On application |
| NVIDIA RTX PRO 6000Blackwell, workstation or server edition | 96 GB GDDR7 ECC, 24,064 CUDA cores. 5th-gen Tensor Cores. PCIe 5.0, 600 W. | On application |
Pricing depends on configuration, quantity, cooling, integration scope, freight, duty and support. We quote in writing against a specific specification and quantity, with a stated validity period. Nothing on this page is an offer.
02Choosing
Which platform is right?
B200 for volume
Mature supply, validated software and the broadest choice of cooling and chassis.
H200 for price-performance
The broadest software ecosystem, for inference fleets where Blackwell-class compute is not required. Pricing is softening as Blackwell ramps.
03Scale
From one node to a multi-rack SuperCluster.
Clients deploy validated scalable units, built to Supermicro and NVIDIA reference architectures, not single boxes.
| Scalable unit | Configuration |
|---|---|
| 144-GPU rackB300, liquid-cooled | 18× 2OU B300 nodes, Quantum-X800 InfiniBand 800G, 48OU ORV3 rack, 33 kW power shelves. |
| 1,152-GPU SuperClusterB300, liquid-cooled | 8 compute racks, 3 Quantum-X800 networking racks, 2 Supermicro 1.8 MW in-row CDUs with N+1 hot-swap pumps. |
| 256-GPU scalable unitH200, air-cooled | 32× 8U nodes, Quantum-2 InfiniBand 400G NDR leaf-spine fabric, in-band and out-of-band management. |