Core Harbor
1,024- or 1,152-GPU SuperPODs*
Carrier hotels and peering exchanges
WHAT IT IS FORLarge distributed training, fine-tuning, batch inference, and model hosting
THE PLATFORM
Buy dedicated metal, managed platforms, networking, and delivery through one portfolio. Train in Core Harbor, serve a region through Middle Harbor, and run near-real-time inference at the access edge, all on a network HarborOne owns.
GPU infrastructure across all three Harbor tiers
B300 SuperPODs, dedicated GPUs, GPU instances, bare metal, GPU partitioning where supported, Reserve or On-Demand
HarborOne Fabric, the network
Private Layer 2, hyperscaler public cloud on-ramps, IP transit, peering, and HarborOne-owned IP routing. Qualifying private cloud connections to AWS and Azure can reduce egress rates by up to 77%, before connection and service charges.
The edge
CDN delivery of cacheable content from ISP-embedded sites
The brain
Private inference endpoints, OpenAI-compatible API, model serving, RAG, fine-tuning, guardrails, and token metering
ARCHITECTURE
1,024- or 1,152-GPU SuperPODs*
Carrier hotels and peering exchanges
WHAT IT IS FORLarge distributed training, fine-tuning, batch inference, and model hosting
Regional GPU capacity
Tier 2 and Tier 3 metros
WHAT IT IS FORRegional inference and burst capacity, closer to the people and devices that use it
Compact GPUs
ISP PoPs, central offices, headends, small data centers, and qualified buildings
WHAT IT IS FORNear-real-time inference, with an ≈5 ms network-path target at selected sites*
SuperPODs are available in 1,024- and 1,152-GPU configurations. Site power profiles depend on the selected systems and facility design.
≈5 ms is a network-path target for selected sites, not total inference latency. End-to-end latency depends on the access network, routing, model size, serving hardware, queueing, and response length.
B300 SuperPODs, dedicated GPUs, GPU instances, bare metal, and GPU partitioning where supported, on Reserve or On-Demand terms
Kubernetes workloads, Slurm-compatible batch jobs, virtual machines, and containers
PostgreSQL, S3-compatible object storage, vector databases, cache, and dashboards
Virtual networks, identity and policy controls, DDoS mitigation, WAF, IPS, and DLP by deployment
Exact service availability varies by Harbor tier, region, hardware, and customer configuration.
| Workload | HarborOne location | Why it fits |
|---|---|---|
| Large model training | Core Harbor | B300 SuperPOD with InfiniBand and high-performance storage |
| Fine-tuning and batch | Core or Middle Harbor | Put compute near the data, the capacity, and the operating requirements |
| Regional inference | Middle Harbor | Serve metro users from a nearby GPU pool |
| Near-real-time inference | Edge Harbor | Qualified edge GPU sites with an ≈5 ms network-path target* |
≈5 ms is a network-path target for selected sites, not total inference latency. End-to-end latency depends on the access network, routing, model size, serving hardware, queueing, and response length.
We map your workload to a tier, design the path, and come back with a plan and a price.