BUILD
Dense GPU compute, placed where the work happens
Harbor Compute provides GPUs for training and hosting inference. Core Harbor supports 1,024- and 1,152-GPU SuperPOD configurations and can scale beyond those sizes. Middle Harbor provides regional capacity, while Edge Harbor hosts inference closer to your users. HarborOne operates the hardware across all three tiers.
Core Harbor scales dense GPU compute
Core sites sit next to carrier hotels, peering exchanges, and major optical routes, with SuperPODs built for distributed training, fine-tuning, batch inference, and model hosting.
| Scope | What runs there |
|---|---|
| Within each system | NVLink and NVSwitch connect GPUs inside an HGX B300 system |
| Across the cluster | InfiniBand fabric links GPU systems for distributed training |
| Storage and management | High-throughput, in-band services, and separate out-of-band management |
| Site infrastructure | Ciena-class core routers, DWDM transport, diverse fiber, and high-density power and cooling |
MIDDLE HARBOR
Middle Harbor regionalizes AI compute
Regional GPU capacity in Tier 2 and Tier 3 metros keeps inference, endpoints, and burst capacity close to your users, with no backhaul to a distant core.
Edge Harbor runs inference near users
Compact accelerators run models at the network edge: ISP PoPs, central offices, headends, and small data centers.
| Scope | What it means |
|---|---|
| Proximity target | ≈5 ms network-path target for selected deployments near users |
| Accelerator options | RTX PRO 6000-class systems, AMD or Intel GPUs, and purpose-built accelerators such as Positron, selected to fit each site's power envelope and workload |
| Site qualification | Not every site in the footprint runs GPUs. Edge Harbor means an access site with accelerators installed. |
How you run it
| Capability | What you get |
|---|---|
| Deployment models | Bare metal, virtual machines, containers, and Kubernetes |
| Batch and HPC | Slurm-compatible jobs: submit the sbatch and srun scripts your team already has |
| GPU sharing | GPU partitioning where the hardware supports it |
| Commercial terms | Reserve for a program, or On-Demand for a single run |
| Data services next to the GPUs | PostgreSQL, S3-compatible object storage, vector databases, and cache |
Specs
| Spec | Value |
|---|---|
| Core Harbor GPU | NVIDIA HGX B300 systems |
| Core Harbor scale | 1,024 or 1,152 GPUs per SuperPOD* |
| In-cluster fabric | InfiniBand |
| In-system fabric | NVLink and NVSwitch |
| Edge accelerators | RTX PRO 6000-class, AMD, Intel, and purpose-built accelerators such as Positron |
| Deployment models | Bare metal · VMs · Containers · Kubernetes |
| Batch | Slurm-compatible |
| Commercial terms | Reserve · On-Demand |
| Node identity | TPM attestation |
Frequently asked questions
What GPUs does Core Harbor run?
NVIDIA HGX B300 systems, with NVLink and NVSwitch inside each system and an InfiniBand fabric across the cluster.
Does HarborOne own the hardware?
Yes. HarborOne operates its own GPU fleet rather than reselling another provider's capacity.
Can I run my existing Slurm scripts?
Yes. Harbor Compute accepts Slurm-compatible batch jobs, so the sbatch and srun scripts you already have will run.
Can I choose where my workload runs?
Yes. Pick the tier and the region. Harbor Compute places the workload there and HarborOne Fabric carries the traffic.
What is the difference between a tier and a region?
A tier describes how close the compute sits to your users: Core, Middle, or Edge. A region describes the jurisdiction it runs under.
Tell us what you want to train
Send the model size, the data volume, and the deadline. We come back with the tier, the cluster, and a price.
