Become a Provider

Step-by-step guide to becoming a GPU Computing Provider on Swan 2.0 Inference Cloud

This guide walks through turning your GPU into an AI inference endpoint on Swan Chain — from starting a local model server, to installing the computing-provider agent, to earning stablecoin revenue from real inference traffic.

circle-info

Looking to consume models instead of provide? See How to Use Swan Inference.

For hardware tiers, collateral economics, revenue splits, and slashing rules, see the Provider Onboarding section of the Swan 2.0 overview. This page focuses on the hands-on setup.

0. Check prerequisites

Providers connect outbound to Swan Inference over WebSocket — no public IP, domain, or SSL setup is required. You just need a capable GPU and one of two supported OS/inference-engine stacks:

Platform
Minimum hardware
Inference engine

Linux (NVIDIA)

GPU with ≥ 8 GB VRAM (Tier C); 24 GB+ recommended (Tier A)

SGLangarrow-up-right (recommended), vLLM, or Ollama

macOS (Apple Silicon)

M1/M2/M3/M4 with ≥ 16 GB unified memory

Legacy GPUs (TESLA P4, GTX 1050 Ti, anything < 8 GB VRAM) cannot serve modern inference workloads and will not receive traffic. Full tier-to-model mapping is in Hardware Tiers.

You'll also need:

  • Go 1.22+ to build the computing-provider agent

  • Docker 24.0+ with the NVIDIA Container Toolkit (Linux only)

  • A funded wallet or credit card for collateral (step 5)

1. Start a model server

Your GPU needs an OpenAI-compatible inference server running locally. Swan Inference will route requests to it via the computing-provider agent.

Linux (NVIDIA) — SGLang

Verify it's healthy: curl http://localhost:30000/v1/models.

macOS (Apple Silicon) — Ollama

Verify it's healthy: curl http://localhost:11434/api/tags.

circle-info

The quickstart uses Qwen 2.5 7B as an example, but earnings scale with real token traffic. Browse the model catalogarrow-up-right to find in-demand models with less provider competition.

2. Install the computing-provider agent

Clone and build from source (mainnet):

Full install details including the NVIDIA Container Toolkit setup are in the computing-provider READMEarrow-up-right.

3. Run the setup wizard

The wizard creates your provider account (or logs you into an existing one), auto-discovers your running model server, and writes config.toml and models.json:

A typical run looks like this (macOS + Ollama):

Save the sk-prov-* key — it's shown once and authenticates this provider to the network.

If you already have a sk-prov-* key (for example, from the web signup at inference.swanchain.io/provider-signuparrow-up-right), pass it directly:

Config files land in ~/.swan/computing/:

  • config.toml — WebSocket URL, API key, node name

  • models.json — mapping from Swan Inference model IDs to your local endpoints

circle-info

Consumer keys (sk-swan-*) and provider keys (sk-prov-*) are different. The computing-provider agent only accepts sk-prov-* keys.

Configuration reference

The wizard writes sensible defaults, but if it failed to discover your model server, you run a non-standard port, or you want to serve multiple models, edit these files directly.

Provider config (~/.swan/computing/config.toml)

Model endpoints (~/.swan/computing/models.json)

Field
Required
Description

endpoint

Yes

URL of your local inference server (SGLang, vLLM, Ollama)

gpu_memory

Yes

GPU VRAM used by this model in MB

category

Yes

Model type: text-generation, image, embedding, audio

local_model

No

Local model name if different from the key (e.g., Ollama's qwen2.5:7b)

api_key

No

API key if your model server requires authentication

The keys in models.json must match valid Swan Inference model IDs. Run computing-provider models catalog or check the model catalogarrow-up-right for the full list.

The agent watches models.json and hot-reloads on change — no restart needed. You can also force a reload:

4. Start the provider and pass benchmarks

Run the agent:

Then check your status:

You'll move through these stages automatically:

Stage
What happens
Typical duration

Connect

Agent opens a WebSocket to Swan Inference, registers your models, and auto-runs math / code / latency benchmarks

Instant

Collateral

Deposit via Stripe or on-chain SWAN (step 5)

Instant

Approval

Admin reviews your benchmark results and collateral

< 24 hours

Active

Traffic starts flowing — you earn per-request revenue

Ongoing

Provider activation stages shown in the dashboard
The My Provider tab visualizes the activation flow: Start → Connect → Deposit Collateral → Approved → Active & Earning.

5. Deposit collateral

Once approved, deposit collateral to unlock full traffic routing. Two options:

Method
Currency
Processing
Refund

Stripe

Credit/debit card (USD)

Instant

7-day waiting period, back to original card

On-chain

SWAN tokens on Swan Mainnet

Requires SwanETH gas

7-day waiting period, back to your wallet

Provider collateral deposit panel
Provider dashboard's Collateral Deposit panel — verify your wallet, then pay the required amount via Stripe or on-chain crypto.

Collateral amounts scale with hardware tier and earning multiplier. See Computing Provider Collateral for the full table.

6. Monitor earnings and uptime

The Provider dashboard at inference.swanchain.io/dashboardarrow-up-right shows live earnings, request counts, and benchmark history.

Provider earnings dashboard
Earnings dashboard with live request volume, per-model breakdown, and payout history.

For a local view, the agent ships its own web dashboard:

Set where payouts go:

circle-info

New Provider Grace Period: For the first 7 days after activation, uptime and success-rate deprioritization are waived. Use this window to stabilize your setup before full routing weight kicks in.

Switching or adding models

Edit ~/.swan/computing/models.json — the agent watches this file and hot-reloads without restarting. Start additional model servers on different ports and add them all to the JSON. Full walkthrough with multi-GPU pinning is in the computing-provider READMEarrow-up-right.

Troubleshooting

Symptom
Fix

invalid provider API key

Verify key starts with sk-prov- and check ApiKey in ~/.swan/computing/config.toml

WebSocket connection failed

Confirm outbound port 443 is open; URL must be wss:// not http://

Provider online but no requests

Model name mismatch — --served-model-name must exactly match the key in models.json and a model ID in the catalogarrow-up-right

could not select device driver "nvidia"

Install the NVIDIA Container Toolkit; see computing-provider READMEarrow-up-right

Stuck in pending

Provider needs collateral + passing benchmark + hardware check. Run computing-provider inference status to see which condition is missing

Full troubleshooting catalog: computing-provider README — FAQarrow-up-right.

Next steps

Questions? Reach the team on Discordarrow-up-right or open an issue on the computing-provider repoarrow-up-right.

Last updated