Five platforms. Priced the honest way.
Three rack-scale 8-GPU server systems and two single-card accelerators. We do not publish prices, because a published Blackwell price is out of date within the week and is a lie by the time you act on it. Send us the configuration and the quantity and you get a real number, in writing, valid for a stated period.
Platforms and configurations
| Platform | Configuration | Price |
|---|---|---|
| Server platforms · 8-GPU rack-scale systems | ||
| HGX B300 EnterpriseBlackwell Ultra · 8U air or 4U liquid | 8× NVIDIA HGX B300 SXM6 · 288 GB HBM3e per GPU (2.3 TB total) 5th-gen NVLink 1.8 TB/s via NVSwitch Embedded ConnectX-8 SuperNICs to 800 Gb/s 2× Intel Xeon 6 · up to 8 TB DDR5 · E1.S NVMe | POAPer system |
| HGX B200 EnterpriseBlackwell · 4U / 8U / 10U variants | 8× NVIDIA HGX B200 SXM6 · 180 GB HBM3e per GPU (1.4 TB total) 5th-gen NVLink 1.8 TB/s 8× ConnectX-7 SuperNICs at 400 Gb/s Up to 15× faster real-time LLM inference vs H100-class | POAPer system |
| HGX H200 EnterpriseHopper refresh · 8U air-cooled | 8× NVIDIA H200 SXM5 · 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 · 1:1 GPU-to-NIC Serves 70B–180B parameter models on a single node | POAPer system |
| GPU accelerators · single-card | ||
| NVIDIA H200PCIe NVL · memory-bound inference | 141 GB HBM3e · 4.8 TB/s bandwidth PCIe 5.0 · 600 W · 2/4-way NVLink | POAPer card |
| NVIDIA RTX PRO 6000Blackwell · Workstation / Server Edition | 96 GB GDDR7 ECC · 24,064 CUDA cores 5th-gen Tensor Cores · PCIe 5.0 · 600 W | POAPer card |
Pricing on application. NVIDIA Blackwell and Hopper channel pricing moves on a weekly basis, and the number that matters to you depends on configuration, quantity, cooling architecture, integration scope, freight, duty and support wrapper. We quote in writing, against a specific specification and a specific quantity, with a stated validity period. Nothing on this page is an offer.
Which platform is actually right?
Frontier training and reasoning inference
Trillion-parameter LLM training with full FP4 Transformer Engine acceleration. Large-scale reasoning workloads at production density. Sovereign and hyperscale AI factory reference builds. Scales to 144 GPUs per ORV3 rack with 1.8 MW in-row CDUs.
The production volume tier
Where mature supply, validated software stacks and proven economics matter more than peak specification. Broadest cooling and form-factor choice in the portfolio. This is the platform behind most large commercial GPU-as-a-service fleets being built right now.
Cost-optimised scale
Best price-performance where Blackwell-class compute is not required. Broadest installed base and ISV coverage. Long-context inference with extended KV cache for RAG and agentic workloads. Note that H200 pricing is compressing as Blackwell ramps.
From one node to 1,152 GPUs.
Individual systems are the building block, not the product. What clients actually deploy is a scalable unit: a validated, cabled, burnt-in cluster with a compute fabric, a storage fabric, management networking and a cooling architecture that the receiving facility can actually support.
We build to the Supermicro and NVIDIA reference architectures — 32-node H100/H200 scalable units at 256 GPUs; B300 ORV3 racks at 144 GPUs each, scaling to eight compute racks, three Quantum-X800 networking racks and two 1.8 MW CDUs for 1,152 GPUs in a single SuperCluster.
And then it needs somewhere to live →| Scalable unit | Configuration |
|---|---|
| 144-GPU rackB300 · liquid | 18× 2-OU liquid-cooled B300 nodes · 36× Intel Xeon · Quantum-X800 InfiniBand 800G XDR · 10× 33 kW power shelves (165 kW, n+2) · 48-OU ORV3 rack |
| 1,152-GPU SuperClusterB300 · liquid | 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 unitH100 / H200 · air | 32× 8U air-cooled nodes · 8× leaf + 4× spine Quantum-2 InfiniBand 400G NDR · in-band and out-of-band management · 9× 48U racks · 34× 208 V 60 A 3-phase PDU |
Platforms and pricing, answered.
What does an NVIDIA HGX B200 or B300 system cost?
Pricing is on application. That is not evasion — it is the only defensible position. Blackwell and Hopper channel pricing moves on a weekly basis, and any figure we published here would be wrong by the time you acted on it. The real number also depends on configuration, quantity, cooling architecture, integration scope, freight, duty and the support wrapper. Send us the specification and the quantity and we return a written quotation with a stated validity period.
What is the difference between HGX B300 and HGX B200?
B300 is Blackwell Ultra: 288 GB of HBM3e per GPU, ConnectX-8 SuperNICs embedded on the board at up to 800 Gb/s, available 8U air-cooled or 4U liquid-cooled. B200 is the volume Blackwell tier: 180 GB HBM3e per GPU, ConnectX-7 at 400 Gb/s, and the broadest choice of chassis. B300 is for frontier training and large reasoning inference; B200 is the production workhorse.
Should I still be buying H200?
Sometimes, yes. H200 has the most mature software ecosystem and the broadest ISV coverage, and it serves 70B-180B parameter models on a single node. For inference-heavy fleets where price-performance beats peak FLOPs, it remains the right economic answer. H200 pricing is also softening as Blackwell ramps, which cuts both ways: better entry cost, but do not lock into a long contract without an escalator clause.
Can you supply liquid-cooled configurations?
Yes. The B300 4U direct-to-chip liquid variant and the ORV3 rack-scale build with Supermicro 1.8 MW in-row CDUs are both available. Liquid cooling changes the facility requirement significantly, which is why we assess the hosting facility at the same time as the hardware.
How quickly can you quote?
For a defined configuration and quantity, normally within two business days. If the requirement is still being shaped, tell us the workload and the budget envelope and we will come back with the platform options that actually fit.
Send us the requirement. We come back with options.
Node count, GPU class, kW per rack, target market, in-service date. That is enough for us to open the conversation with the right suppliers and the right facilities. Under NDA from first contact.