Serverless GPU vs Dedicated GPU: Which Should You Use?

Serverless and dedicated GPU work draw on available wallet funds through different lifecycles. Understanding those lifecycles helps prevent duplicate requests, abandoned machines and incorrect assumptions about temporary reservations. For model and application fit, use the architecture worksheet.

Dedicated compute: authorize a machine, then track runtime

Select a current authenticated offering and request a quote. The response separates selectedCustomerRate from authorizedMaxCustomerRate. Review both before launching the same configuration with its approved ceiling. A changed ceiling can produce 409 and requires a new review. An article's hourly example cannot substitute for that quote.

An accepted launch returns 202 and an instance identifier. Follow instance status until ready, then use the displayed SSH details. The GPU count describes devices within one machine; it does not create a distributed cluster. Keep the instance identifier with job records so you can match later usage to the actual workload. See instance lifecycle and termination.

Closing SSH, ending a notebook cell or stopping the application is not the platform termination operation. Save outputs to a verified destination, request termination through supported controls and confirm the resulting state. Do not assume that a stopped process removes runtime charges or that instance-local files survive termination.

Serverless: authorize a request, then settle delivered work

A supported inference request reserves an estimated amount before dispatch. Chat authorization incorporates message character counts and the requested output ceiling; these are estimate inputs, not measured token counts. Settlement uses reported usage when available and can fall back to the estimate when usage is absent. The response for normalized chat/image includes decimal billing values.

A failure before delivery attempts hold release. Delivered work whose settlement encounters an error enters reconciliation instead of being treated as free. Consequently a timeout cannot prove that generation failed, and a reservation is not itself a final charge. Keep one logical operation key and avoid submitting concurrent duplicates.

Compare the right accounting evidence

GET /wallet exposes balance, reserved and available values as decimal strings; available is balance less reservations, clamped to zero. A healthy nominal balance may therefore be insufficient for another authorization. The held field is an alias of reserved, not an additional amount to subtract.

GET /usage reports GPU-instance usage windows, not a combined serverless export. Use the correct dashboard records and generation response evidence for inference. Reconcile a common UTC window and currency before comparing paths; do not add wallet debits to the usage charges that already explain those debits. The wallet troubleshooting guide walks through discrepancies.

Budget for completed useful work

Dedicated cost depends on metered runtime and the selected configuration; serverless cost depends on modality and the model's pricing units. Compare total settled cost divided by accepted useful results at the required quality. Include failed attempts, idle runtime and operational effort in the analysis without inventing a universal request price or utilization break-even. Read pricing and authorization before committing resources.