01 / Reserved GPU capacity
Your next cluster. Entirely yours.
Single-tenant B300 bare metal for the workloads that need their own infrastructure. Reserve a complete cluster with a deployment plan built around your site and schedule.
The application
Room for the whole workload.
Bare-metal control. A dedicated fabric. Capacity planned at the cluster level.
- 01
High-volume inference
Run predictable inference services on dedicated hardware with capacity reserved for your workload.
- 02
Training and fine-tuning
Keep control of the software stack and single-tenant fabric for training and fine-tuning runs.
- 03
Whole-pod offtake
Plan multi-year whole-pod capacity for a neocloud or GPU marketplace.
Reference specification
Pod 32. In detail.
The reference specification is the authority on component counts and configuration. The renders show an illustrative design study.
Explore the technology| Compute | 32 Supermicro HGX B300 nodes |
|---|---|
| GPUs | 256 NVIDIA B300 |
| Racks | 4 |
| Operating load | 645 kW |
| System memory | 128 TB |
| Storage | 983 TB raw NVMe per module |
| Fabric | Rail-optimized 400G Ethernet |
| Fabric throughput | Configurable 3.2 or 6.4 Tbps per node |
| Cooling | Direct-to-chip liquid, redundant CDUs |
| Site water | No site water requirement |
| Tenancy | Single-tenant bare metal |
Engineering and operations
Measured before the handoff.
Factory, site, integration.
A 24-stage test program follows the module from factory build through site commissioning and integration.
Witnessed acceptance.
Named gates F-1 through F-7 carry witnessable evidence into handoff, including measured NCCL fabric-performance pass bars.
A site-specific schedule.
Targeted to be operational within roughly three months of pre-payment, confirmed site by site. Factory build and burn-in run in parallel with site preparation.
Start a conversation
A little context. A useful call.
Three short questions, then choose a time. No project names or sensitive details needed.
Question 1 of 3
Your role
Only broad answer categories are recorded. Keep sensitive project details for the conversation.
Book without the questions Open Cal.com in a new tabFurther detail
Before we talk.
Who owns and operates the hardware?
Pacific builds, deploys, and operates each module. You get single-tenant bare-metal access to the cluster. Operation is unstaffed on site, run through remote operations with contracted regional field service.
Which network configurations are available?
Rail-optimized 400G Ethernet. Configurable 3.2 or 6.4 Tbps per node. The configuration is selected with our engineering team during the deployment plan for your workload.
How fast can capacity be live?
Targeted to be operational within roughly three months of pre-payment, confirmed site by site. Site work is limited to a prepared pad and power feed and runs in parallel with the factory build.
What evidence comes with a deployment?
Every module ships with witnessed acceptance evidence from a 24-stage factory, site, and integration test program, including measured NCCL fabric-performance pass bars at gates F-1 through F-7.
What does it cost?
Commercial terms are discussed privately on a qualified capacity call. Pacific does not publish pricing.