Docs / Connect Verda

Connect a Verda account

Verda (formerly DataCrunch) is a Finnish GPU cloud with A100, H100, H200 and RTX Pro 6000 capacity, spot pricing at half the list rate and three EU locations (FIN-01, FIN-02, FIN-03, all in Finland). The price index lists Verda rates.

What you need

1. A Verda project with GPU quota in the location you want.
2. OAuth client credentials: in the Verda console open Credentials in the sidebar, then Cloud API credentials, and click Create. Copy the client id and the client secret; the console shows the secret once.

Link it

In the dashboard, Connect a cloud, choose Verda, paste the client id and secret and give the account a name. LLM Hangar verifies the credentials live by exchanging an access token; nothing is stored if the exchange fails. The credentials are stored envelope-encrypted, the same as every other provider. LLM Hangar refreshes the short-lived access tokens automatically.

curl -X POST https://app.llmhangar.com/v1/providers/verda/link \
  -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  -d '{"name": "Verda FIN-03", "client_id": "...", "client_secret": "..."}'

What LLM Hangar creates in your project

• Instances are ordered by hostname (llmd-<deployment id>) from a docker-enabled CUDA image, so a retried create adopts an instance that already exists instead of ordering a second one.
• Startup scripts are created per deployment start and deleted with it. They carry the deployment's own keys, which is why they are never shared between deployments. Verda allows 100 per project; the dashboard warns at 50.
• Stages are NVMe block volumes in the deployment's location (0.20 USD per GB-month at list), attached to the instance, formatted once by the first boot and kept when it stops. A stopped deployment bills the stage volume only; a resume orders a fresh instance and reattaches it.
• Spot instances are ordered with the volumes kept on reclaim; recovery works as described in Spot instances.
• Teardown deletes the instance, its OS volume and its scripts, and purges the volume from Verda's trash (a deleted volume otherwise sits there for 96 hours and counts toward the storage quota), then verifies the project lists none of them before the deployment is marked destroyed.
• Capacity depends on the location. If your first choice is unavailable, LLM Hangar tries the fallback configurations you approved.