Dedicated NVIDIA GPU clusters, reserved for your workload.
NVIDIA HGX B200 and B300 systems for training, fine-tuning, and inference. Single-tenant, connected by a non-blocking fabric, and tested before you log in.
Two HGX platforms. Eight GPUs per system.
Choose by workload. B200 covers most training and inference. B300 adds memory and network bandwidth for long-context and memory-bound work.
| HGX B300 | HGX B200 | |
|---|---|---|
| Architecture | NVIDIA Blackwell Ultra | NVIDIA Blackwell |
| GPUs per system | 8 | 8 |
| GPU memory per system | 2,304 GB HBM3e | 1,440 GB HBM3e |
| GPU interconnect | NVLink, 1.8 TB/s per GPU | NVLink, 1.8 TB/s per GPU |
| Scale-out network | 800 Gb/s per GPU | 400 Gb/s per GPU |
| Suited to | Long-context and reasoning models; Memory-bound inference; Large mixture-of-experts training | Pre-training and continued training; Fine-tuning at scale; Batch and online inference |
Figures are NVIDIA's published HGX platform specifications. CPU, system memory, local storage, and NIC models are confirmed per deployment.
Capacity on terms you can budget.
Reserved clustersMulti-month and multi-year terms
A fixed allocation of HGX systems reserved for your organization, from one system to multiple racks. Rates are locked for the term.
- Dedicated systems and fabric partition
- Capacity planned with you before installation
- Expansion options written into the agreement
Dedicated inferenceMonthly or reserved
Single-tenant serving capacity sized to your traffic, with consistent latency because no other customer shares the hardware.
- Sized from one system upward
- Bring your own serving stack, or use ours
- Private connectivity to your applications
Short-term capacityPlanned after general availability
Shorter commitments for experiments, evaluations, and bursts, offered from capacity not held under reservation.
- Availability depends on reserved demand
- Same hardware and network as reserved clusters
- Simple upgrade path to a reservation
Everything between the GPU and your job.
We operate the layers you would rather not, and leave the rest to you.
Orchestration
Bare metal with root access, managed Kubernetes, or managed Slurm. You choose per cluster.
Drivers and libraries
NVIDIA driver, CUDA, NCCL, and container toolkit, versioned and updated on a schedule you approve.
Storage
High-throughput parallel file system for training data and checkpoints, plus S3-compatible object storage.
Network
Non-blocking scale-out fabric between systems, a separate storage network, and private links to your environment.
Observability
Per-GPU telemetry, fabric counters, and node health, exportable to your own monitoring tools.
Access
SSO-backed access to management interfaces, per-user audit logs, and network-level isolation between customers.
- Your workloadsTraining, fine-tuning, inference
- OrchestrationBare metal, Kubernetes, or Slurm
- Fabric and storageNVLink, scale-out network, parallel file system
- ComputeNVIDIA HGX B200 and B300
- FacilityPower, cooling, network, security
No cluster is handed over untested.
Each system runs a fixed acceptance suite before handover. You get the full results, so performance questions start from data.
1. Burn-in
Sustained load on every GPU, memory, and power supply to surface early failures before you see them.
2. Fabric validation
Collective-communication benchmarks across the full cluster to confirm bandwidth on every link.
3. Storage validation
Read and write throughput tests from every node against the target for your cluster.
4. Handover report
Results for every test, per system, delivered with credentials and runbooks at handover.
GPU cloud questions
Can we bring our own software stack?
Yes. Bare-metal clusters come with root access. If you would rather not manage the base layer, we run Kubernetes or Slurm and you schedule jobs on top.
Which interconnect will you use between systems?
Each deployment uses a non-blocking scale-out fabric sized to the per-GPU network bandwidth of the platform. The fabric technology is confirmed with design partners as part of the reference architecture.
What happens when hardware fails?
Failed components are replaced by our on-site team. Reserved clusters include spare capacity terms in the agreement so a failed system does not reduce what you are paying for.
Can we visit the facility?
Design partners and customers can arrange escorted visits once the facility is operational.