Fabric & isolation
The proposed design combines high-speed GPU interconnects with separately scoped storage and management networking. Customer resource boundaries are agreed in the technical schedule.
Dedicated NVIDIA B300 clusters
Planned B300 infrastructure with complete nodes, high-speed fabric, storage and operating support.
Planned hardware specifications
Planning targets are validated as a complete system. Hardware figures alone are not a performance guarantee.
- Accelerator
- NVIDIA B300 / Blackwell Ultra
- Cluster configuration
- Dedicated clusters built from complete eight-GPU nodes
- Memory
- 288 GB HBM3e per GPU · approximately 2.3 TB per node
- Fabric target
- 800 Gb/s per GPU · topology confirmed at technical validation
- Storage target
- NVMe-backed parallel file system and S3-compatible object storage
- Power & cooling
- Validated for each deployment and configuration
- Delivery status
- Pre-deployment · final density subject to OEM and system acceptance
Deployment capacity and expansion schedules are scoped through a qualified technical review. Final configurations require OEM, electrical and cooling validation before acceptance.
GPU cluster architecture
Explore the layers that turn accelerators into usable, dedicated capacity.
Illustrative cluster architectureSelect a layer
THE ACCELERATOR
NVIDIA B300
Blackwell Ultra with a planned 288 GB of HBM3e per GPU. The accelerator is the beginning of the system.
REFERENCE DESIGN · SUBJECT TO VALIDATIONFabric, storage and operations
Storage & data
Local scratch, a parallel file system and object storage serve different parts of the workload. Capacity, throughput, retention and transfer terms are specified together.
Power & cooling
GPU density, continuous and transient power, cooling capacity and redundancy must be accepted as one system before procurement and handover.
Operations & support
Planned onboarding, monitoring, incident escalation and reporting. Response targets and remedies belong in the signed service schedule.
Workload acceptance tests
Acceptance tests follow the customer’s job, not a generic performance headline.
| Test area | What is agreed |
|---|---|
| GPU & fabric | Workload profile, distributed-job behavior and interconnect validation |
| Storage | Dataset access, checkpoint throughput and agreed capacity |
| Thermal & power | Sustained load, transients and cooling behavior in the selected configuration |
| Isolation | Customer access and boundaries across compute, storage and management |
| Handover | Runbook, escalation path, reporting and acceptance evidence |
Product roadmap
Planned product phases alongside customer-led capacity expansion. Timing and availability remain subject to validation and signed agreements.
Reserved B300 clusters
Dedicated whole nodes with an agreed term and acceptance process.
On-demand nodes
Eight-GPU nodes for evaluation and burst workloads, subject to capacity.
Inference endpoints
Dedicated and serverless access for supported models.
GB300 NVL72
A potential next platform, conditional on demand, engineering and financing.