> ## Documentation Index
> Fetch the complete documentation index at: https://docs.omnia-voice.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Dedicated endpoints

> Private GPU capacity for a specific model, billed per GPU-hour.

A dedicated endpoint is a **private GPU deployment** of a model, reserved for your
workspace. Unlike shared per-token inference, you get isolated, guaranteed
capacity, billed per GPU-hour with per-minute granularity, only while it runs.

## When to use dedicated

<CardGroup cols={2}>
  <Card title="Guaranteed throughput" icon="gauge-high">
    Reserved GPUs mean predictable latency and no contention with other tenants.
  </Card>

  <Card title="Isolation" icon="shield">
    Your traffic runs on capacity dedicated to your workspace.
  </Card>

  <Card title="Serve fine-tuned models" icon="wand-magic-sparkles">
    Deploy a model you [fine-tuned](/fine-tuning/overview) and call it by its
    routing key.
  </Card>

  <Card title="Sustained high volume" icon="chart-line">
    At high, steady request rates a dedicated GPU can be more economical than
    per-token pricing.
  </Card>
</CardGroup>

## How billing works

Dedicated endpoints are billed **per GPU-hour**, metered per-minute, while the
endpoint is running, **not** per token. The rate depends on the GPU type, GPU
count, and number of running replicas. Individual inference requests to the
endpoint carry no per-token charge; you've already paid for the GPU time.

The GPU price is Omnia's own configured rate (which includes the platform's
markup). The available GPU types, regions, and flavors come from the
[templates endpoint](/dedicated/deploy), so you always deploy against
combinations that actually exist.

<Warning>
  Deploying requires enough prepaid balance to cover a minimum runway (by default
  **one hour** of GPU time). If your wallet can't cover it, deployment is refused
  and the response tells you the **exact shortfall** to add.
</Warning>

<Note>
  A running endpoint whose wallet becomes exhausted is **automatically stopped** to
  prevent further charges. Top up the wallet and start it again when you're ready.
</Note>

See [Billing](/concepts/billing) for how the wallet is charged across shared,
dedicated, and fine-tuning usage.

## Autoscaling

An endpoint autoscales between the `minReplicas` and `maxReplicas` you set. It
scales up under load and back down when idle, and cost scales with the number of
running replicas. Set `minReplicas` to keep warm capacity always available, and
`maxReplicas` to cap both throughput and spend.

## Lifecycle

<Steps>
  <Step title="Deploy">
    Pick a base model, GPU type, count, region, and replica range. See
    [Deploying an endpoint](/dedicated/deploy).
  </Step>

  <Step title="Use">
    Route to the endpoint by passing `model = dedicated/<routing-key>` on the
    normal inference API. It autoscales within the range you set.
  </Step>

  <Step title="Manage">
    Scale, disable, or delete it any time. See
    [Managing endpoints](/dedicated/manage).
  </Step>
</Steps>

## Managing via API or dashboard

You deploy and manage dedicated endpoints from the dashboard, or programmatically
against the management API at `https://gateway.omnia-voice.com/v1` using the same
workspace API key you use for inference. Creating, updating, and deleting
endpoints are **admin-gated** operations. See
[Deploying an endpoint](/dedicated/deploy) and
[Managing endpoints](/dedicated/manage).

<Note>
  If you need a GPU count above 8 per replica, or a region that isn't listed in the
  templates, the dashboard shows **"contact sales"**; those deployments are
  arranged directly.
</Note>