GPU Cloud Pricing Explained: What Determines Hourly Cost?

A GPU price describes an offering and a billing unit. It does not by itself describe model throughput, total project cost or future availability. Compare complete configurations and understand the distinction between a displayed rate, an authorization ceiling and settled usage before launching compute.

Identify the exact item being priced

Record GPU model, device count, memory variant, CPU, system RAM, disk, region, tier, capacity class and supported environment. Two rows with the same accelerator name may differ in the resources that feed it. A multi-GPU offering is one machine containing several devices, not several independently managed machines or a promised cluster interconnect.

The public availability API groups customer-visible configurations and prices for comparison. It does not expose the deployment offering ID. Use authenticated GET /gpus with gpus:read to obtain that ID for quoting. The authenticated list is capped at 100 offerings and has no cursor pagination; a missing row is not proof that every possible configuration is unavailable.

Separate selection from authorization

A quote request to POST /instances uses intent: "quote" and the selected configuration. It returns selectedCustomerRate, authorizedMaxCustomerRate, currency and capacity-related fields. Review the selected rate and ceiling separately. A quote does not launch a machine or reserve future inventory.

For launch, send the reviewed configuration and ceiling with intent: "launch", an appropriate Idempotency-Key and instances:write. Omitted intent defaults to launch, so an integration should set it explicitly. If the current ceiling no longer matches, a 409 requires a fresh quote and review rather than silently increasing the authorized amount. See quote and launch fields.

Estimate runtime cost without inventing a bill

The planning relationship is rate multiplied by billable duration, summed over the relevant resources and rate periods. Use consistent units and currency. Do not assume a billing increment, a minimum charge or separate storage/network fees from an article: verify what the selected offer and current billing terms actually include.

Keep estimates separate from observed customerCharge records. Money is returned as decimal text; ordinary binary floating-point accumulation can introduce reconciliation differences. Compare the ledger and usage over a common time window, and do not add reservations to final charges as though both were expenses.

Why lower hourly cost can lose

A smaller machine may require more retries, CPU preprocessing or reduced batches. A larger one can waste money when idle. Evaluate settled cost per accepted job under a fixed quality and latency target, including load time and idle runtime in the operational analysis. No speedup follows merely from an accelerator's name.

GPU process exit and SSH disconnection do not invoke platform termination. Export required artifacts, terminate using supported controls and confirm status. Use the lifecycle guide and the optimization worksheet. Current rates belong in pricing and the live offering review, not in an undated static comparison table.