---
title: "Migrating from legacy inference endpoints to serverless API"
slug: "ai-inference-legacy-endpoints-retirement"
source: "https://app.cloudpe.com/help/ai-inference-legacy-endpoints-retirement"
updated: "2026-10-08T05:09:58.485Z"
---

# Migrating from legacy inference endpoints to serverless API

## Overview

The legacy GPU-VM based Inference Endpoints feature (accessed via `/dashboard/inference`, using `cpe-inf-` keys) is officially deprecated and retired. Creation of new dedicated GPU endpoints is frozen, and the deployment wizard is hidden in the dashboard navigation.

Sunset date: to be announced. Existing dedicated endpoints continue operating and accruing hourly VM billing until deleted by the user or decommissioned at final sunset. All customers are encouraged to migrate workloads to the serverless CloudPe AI Inference API hosted at `https://inferapi.cloudpe.com/v1`.

## Before you start

- Deprecation status: Dedicated GPU endpoints created under the legacy architecture cannot be scaled or reconfigured.
- Account eligibility: Your organization must be KYC-verified or funded via a direct paid wallet top-up (promotional credits and sign-up bonuses do not qualify). Eligibility is checked when minting API keys or playground session tokens.
- Permissions: Generating serverless inference keys requires the `ai:keys` permission. Deleting legacy endpoints requires `vms:delete`.
- Documentation: Review serverless connection samples in [AI inference quickstart](/help/ai-inference-quickstart).

## Steps

### Step 1: Create a serverless inference API key

1. In the sidebar, open **API Keys** under **AI**.
2. Select **Create key**.
3. Provide a key name (for example, `migrated-service-key`).
4. Optionally configure RPM, TPM, concurrency, or monthly spend caps.
5. Select **Create** and copy the secret key token starting with `cpk_`.
6. Select **Done** to close the dialog.

### Step 2: Update application configuration

Update your client configuration settings:

1. Replace the legacy dedicated base URL (`<dedicated_endpoint_ip>:8000/v1`) with the unified serverless endpoint:
   ```text
   https://inferapi.cloudpe.com/v1
   ```
2. Replace your legacy `cpe-inf-` key with your new `cpk_` API key in the `Authorization: Bearer <API_KEY>` header.
3. Update model identifiers to supported serverless model slugs (for example, `llama-3-1-8b`, `gemma-3-27b`, or `qwen3-32b`). See [Supported AI models and token pricing](/help/ai-inference-models-pricing).

### Step 3: Verify serverless responses

Run a test completion request against the serverless endpoint:

```bash
curl -X POST https://inferapi.cloudpe.com/v1/chat/completions \
  -H "Authorization: Bearer $CLOUDPE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llama-3-1-8b",
    "messages": [{"role": "user", "content": "Ping"}]
  }'
```

### Step 4: Decommission legacy dedicated endpoints

Once traffic is successfully rerouted to the serverless endpoint, delete your legacy GPU VM endpoints:

1. Open **Legacy Inference** under **PLATFORM** (shown only while you still have endpoints) in the sidebar.
2. Select the legacy endpoint from the list.
3. Select delete and confirm. Decommissioning stops continuous hourly compute billing for the associated GPU VM.

## API

Serverless endpoint replaces legacy per-VM floating IPs:

| Architecture | Base URL | Auth prefix | Billing model |
|---|---|---|---|
| Legacy dedicated | `<floating_ip>:8000/v1` | `cpe-inf-` | Continuous hourly GPU VM |
| Serverless API | `https://inferapi.cloudpe.com/v1` | Inference key (`inference:invoke`) | Per-token consumed |

## Limits & billing

- Pay per token: Serverless inference eliminates the need to pay for idle GPU VM compute time. You are billed strictly for input and output tokens consumed by your requests.
- No infrastructure maintenance: Serverless removes VM lifecycle management, disk resizing, floating IP allocations, and custom vLLM parameter tuning.
- Availability: Serverless models run in Zone B with elastic replicas (`min_replicas=0`, `max_replicas=1`). Cold starts and single-replica limits apply; see [AI inference reliability and architecture](/help/ai-inference-reliability).

## Troubleshooting

| Message | What it means | What to do |
|---|---|---|
| `key not found` | The key identifier queried does not exist or was deleted. | Verify the key identifier in the API keys table. |
| `Permission denied: 'ai:keys' required` | Your account lacks permissions to manage inference keys. | Request the `ai:keys` permission from an organization administrator. |

## FAQ

**Can I deploy custom HuggingFace weights on the serverless API?**
The serverless API serves curated, high-performance open weights models. If your workload requires bespoke proprietary fine-tuned weights, provision a dedicated GPU instance. See [Deploying GPU virtual machines](/help/gpu-vm-deployment).

**What happens to my legacy endpoints if I do not migrate immediately?**
Sunset date: to be announced. Legacy endpoints remain online during the transition window, continuing hourly VM billing until deleted by the user or retired at sunset.

**Do serverless API keys expire?**
Serverless keys remain active indefinitely until explicitly revoked by an administrator.

## Related

- [Deploying AI inference endpoints](/help/ai-inference-endpoints)
- [AI inference quickstart](/help/ai-inference-quickstart)
- [Creating and managing AI API keys](/help/ai-api-keys)
- [Tracking AI usage and billing](/help/ai-usage-and-billing)