Skip to main content

Managed deployments

InferCrane Cloud is the MVP path for customers who want InferCrane to operate a model deployment. It is deliberately time-boxed and prepaid. BYOC remains a separate path in which the customer pays their cloud directly. The control plane must not queue provider work unless the tenant’s available wallet credit covers the complete authorization hold. A merely positive balance is insufficient: available credit must be greater than or equal to the current hold.

Customer flow

  1. Choose a model and serving goal.
  2. Select InferCrane Cloud · prepaid and a 1, 6, or 24 hour runtime window.
  3. InferCrane checks a current exact provider price and launch evidence.
  4. Review the hourly rate, authorization hold, available credit, and any shortfall.
  5. Add credit when needed, then approve the deployment.
  6. InferCrane atomically reserves credit and queues one durable deployment operation.
  7. The runtime clock starts only when the stable endpoint is activated.
  8. At the runtime limit, InferCrane withdraws the route, deletes provider resources, settles measured runtime, and releases unused held credit.
The UI never labels a hold as spend or an estimate as an invoice. Failed launch evidence blocks the billable action. A failed pre-activation operation enters the same durable cleanup path; the hold is released only after no live provider replica remains.

What you pay

The current MVP supports one elastic RunPod GPU per managed deployment. The GPU is created with the control plane’s own RunPod credential, not a customer credential, and the product you buy is InferCrane Cloud. The server sets the hourly rate from a current, complete provider price observation. A browser request cannot set or change the rate. For example, at a rate of $0.70/hour, a one-hour deployment holds $0.816667: one hour of runtime plus the ten-minute cleanup allowance. The hold is not the final charge. The charge never exceeds the hold, and unused credit returns to the available wallet after provider deletion succeeds. A deployment that never activates is not charged. The quote returns the hourly rate, the hold, the runtime limit, the cleanup allowance, the GPU and its placement, and the time until which the quote is valid. Provider credentials and provider resource identifiers stay server-side.

Why the rate differs from the GPU price catalog

The GPU price catalog lists each provider’s public price for a GPU. It is the reference for BYOC, where you pay the provider directly and provider charges never pass through InferCrane. The InferCrane Cloud rate is a different price: what InferCrane charges to provision, serve, and delete the deployment for you. Expect it to be higher than the provider’s listed price for the same GPU. The deployment review shows the exact rate and hold before anything is charged.

Atomicity and recovery

Wallet reservation, deployment intent, and lifecycle operation are one database transaction. Idempotency keys cannot create duplicate holds. A deployment name has at most one open hold; a name deployed again after deletion gets a new hold, and the settled one stays as history. Deletion settlement and wallet release are also one transaction and are safe to retry. The background expiry worker finds:
  • activated reservations past their time limit; and
  • terminal failed or cancelled launches that never activated and have no live replica.
It queues the ordinary durable delete workflow rather than mutating wallet state directly. Provider uncertainty remains visible and credit stays reserved until cleanup is proven.

Operator enablement

Managed deployments are off by default even when a RunPod key exists. Enable them only after wallet funding, current GPU pricing, provider deletion, and the complete browser journey have been qualified:
The entrypoint reads the owner-restricted file and exports the key only inside the control-plane process. Stripe test mode should remain enabled until checkout, signed webhook replay, wallet credit, deployment hold, settlement, and refund behavior pass together. Fly.io hosts the CPU control plane and stable public gateway. It does not supply GPUs. Customer wallet credit pays for managed GPU exposure; BYOC provider charges never pass through InferCrane.

MVP limits

This is not an unlimited production cluster contract. The initial self-serve product supports one GPU, one replica, elastic serving, and a maximum 24-hour window. Longer-running or multi-GPU deployments remain BYOC or sales-assisted until reconciliation, capacity commitments, and support SLOs are qualified.