AI DELIVERY
Your model, closer to the people using it
HarborOne's AI delivery network serves your models behind a private, OpenAI-compatible API. Each request is routed to a capable site, and returns the response over a network we own.
The request journey
Model and site selection
Inference on a nearby GPU
Generated response returning
Training is solved. Serving isn't
Harbor ADN runs your model close to your users: Core Harbor for scale, Middle Harbor for regional demand, Edge Harbor for the fastest response.
No round trip to a distant region.
What it does
| Capability | What you get |
|---|---|
| Private inference API | An OpenAI-compatible endpoint, so code your developers already wrote points at HarborOne |
| RAG and fine-tuning | Retrieval and tuning against your own data, next to the GPUs that serve it |
| Guardrails | Policy applied inline on the gateway |
| Token metering | Every call metered, with usage visible per model |
| Placement by tier | The request runs at the tier that fits the workload, from Core Harbor to the access edge |
NEAR-REAL-TIME WORK
Proximity changes the product
Some workloads tolerate a distant region. These do not.
Agricultural technology
Field vision, crop monitoring, and equipment assistance
Gaming
Interactive characters, personalization, and local event analysis
Video and streaming
Real-time content analysis and responsive media workflows
Robotics
Perception and task assistance near the operating site
Smart cities
Traffic, infrastructure, and public-space event response
Telehealth
Low-latency workflow support for remote care teams
Use Edge Harbor for workloads that benefit from network proximity, and keep safety-critical control and clinical decisions inside your own validated workflow.
What ≈5 ms means, and what it does not
The ≈5 ms figure is a network-path target for selected Edge Harbor deployments near users. It is not a total model-response time. End-to-end latency depends on the access network, routing, model size, serving hardware, queueing, and response length. We will tell you what your workload can expect on your route before you sign anything.
Specs
| Spec | Value |
|---|---|
| API | Private inference endpoint, OpenAI-compatible |
| Capabilities | Model serving · RAG · fine-tuning · guardrails · token metering |
| Placement | Core Harbor · Middle Harbor · Edge Harbor |
| Edge proximity | ≈5 ms network-path target at selected sites* |
| Edge accelerators | RTX PRO 6000-class, AMD, Intel, and purpose-built accelerators such as Positron |
≈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.
Frequently asked questions
What is an AI delivery network?
A network that runs models in production and routes each request to a site capable of serving it, rather than sending every request to one distant region.
Can my existing OpenAI API code run on HarborOne?
Yes. Harbor ADN serves models behind an OpenAI-compatible API.
Where does inference run?
At the Harbor tier that fits the workload. Large batch work runs in Core Harbor, regional serving in Middle Harbor, and near-real-time work at qualified Edge Harbor sites.
Is ≈5 ms the response time I will see?
No. It is a network-path target at selected sites. Your end-to-end time depends on your access network, model, and response length. Ask us and we will work it through for your route.
Bring us the workload with the tightest loop
Tell us what has to respond, and how fast. We come back with the tier, the site, and what your route can do.
