Supplier onboarding
From a real application to eligible serving capacity—and the checks that still have to be completed.
In this chapter
What a supplier sells#
A supplier sells authorized inference performed by infrastructure it owns or is permitted to operate. This can include owned GPU servers or cloud-hosted model deployments, subject to the relevant hosting and model terms. A consumer AI subscription or unused provider credit balance is not automatically transferable supply.
A supplier is paid for qualified demand it actually serves under the marketplace's eventual terms. Connecting a server does not guarantee requests, utilization, earnings, or profit.
The application available today#
| Field | Requirement |
|---|---|
| Wallet account | Sign in; the server assigns ownership from the session. |
| Supplier name | A meaningful operator or business name. |
| Endpoint | An HTTPS base URL without embedded credentials, query parameters, or fragments. |
| Model | The exact identifier expected by the endpoint. |
| Operations email | A usable contact for review and incidents. |
| Serving rights | Explicit confirmation that the operator may supply the inference. |
| Retention description | A clear account of prompt/response handling, not an unsupported privacy badge. |
The current application accepts one record per wallet and starts at pending_review. It persists across sign-out and restart. Duplicate submission is rejected instead of silently replacing a reviewed record. There is no self-service editing or activation guarantee yet. API credentials must not be included in the application text.
Checks required before buyer traffic#
- Confirm the operator's identity/authority as required by the business and payment model.
- Verify endpoint control and the right to serve or resell the model.
- Review the model license and any upstream contractual restrictions.
- Validate endpoint destinations against SSRF, private-address access, redirects, and DNS-rebinding risks before making probes.
- Exchange appropriately scoped credentials through a secure channel.
- Test model/API compatibility, token reporting, limits, latency, health, and failure behavior.
- Review privacy and retention commitments; distinguish contractual assurances from attested execution.
- Agree pricing, settlement, dispute, suspension, and payout conditions.
- Start with limited traffic and monitor observed performance before expanding.
An application is not activation#
Future seller controls should include available capacity, concurrency limits, drain/pause, credential rotation, actual request outcomes, and reconciled earnings. A public self-service home-GPU installer is a separate product with update, sandboxing, networking, and hardware-support requirements.