ChatGPT GPU Rental: Quotes, Hourly Minimums and Reuse

An agent should check whether a task fits, review its total billing conditions and ask for spending approval before renting a GPU. CloudGPU provides an anonymous Agent self-check, a machine-readable contract and MCP setup instructions. Live capabilities and quotes take precedence over examples in an article.

Start with tools and owner approval

Use a client that supports the CloudGPU MCP server or the documented remote MCP connection. Whether a particular ChatGPT account exposes an appropriate connector depends on that client’s available features and permissions; a normal chat cannot install the tools by following a payment link.

The published account flow is connect_account → the owner signs in with Google and reviews the access limits → finish_connect. If funding is needed, request_top_up returns a link for the owner and wait_for_funds checks credited balance. The Agent does not receive a payment card or permission to approve its own spending.

You can give the Agent this request:

Check CloudGPU’s live capabilities and whether my task fits. List my existing instances before considering a new rental. Show the GPU, billing increment, minimum charge, startup reservation, expected setup time and output plan. Wait for my approval before creating anything or changing limits.

Do not paste API keys, passwords, device codes or signed download links into public examples. The connect flow delivers the approved credential through its intended client mechanism.

Separate the billing increment from the startup reservation

An hourly displayed price is a rate. The server quote also specifies how runtime is rounded and the balance required to start.

Quote condition	What a short session means
Minute billing	Runtime rounds up to whole minutes under the quoted minimum.
Started-hour billing	Each begun hour is charged in full, including less than an hour. Unused charged time is not refunded on voluntary stop.
Startup reservation	Balance held at startup; this is separate from the final billing increment. Minute-source unused prepaid minutes are returned on destruction under that source’s terms.

For an illustrative hourly rate P and an eleven-minute billed-runtime session, a minute source gives 11 × P / 60, while a started-hour source gives P, before any separate disk or storage charges. This is arithmetic, not a measured complete-task price. Model downloads, environment setup, retries and output transfer can extend the runtime.

CloudGPU’s quote is source-specific. A whole-hour source requires explicit acceptance of its billing terms. Agents using Jobs must follow the actual Job and rental quote requirements; selecting a GPU name is not acceptance of a different billing increment. Review current pricing and the live quote before committing funds.

Reuse an existing instance when it is compatible

The MCP Agent instructions distinguish two workflows:

• For an existing running rental, check ownership, readiness, available memory/disk and its remaining approved limits. Several approved commands can run sequentially on that same instance through run_command.
• Independent run_job and prepared rerun calls provision separately. They do not silently attach a new task to an existing rental.

Reuse can avoid another model download or startup when the existing environment fits. It does not make an idle GPU free. Keeping a started-hour rental across the next billing boundary adds another full-hour charge. Do not extend the owner’s limits or keep a GPU idle simply because part of an hour was prepaid.

A saved Workspace version is another kind of reuse: files can be applied to an eligible instance in a new private directory. Applying files does not install all dependencies or execute a project, and the version must still be available and within its retention and budget.

Choose the task path and output plan

Use the matching guide before choosing a GPU yourself:

• ComfyUI image workflows
• AI video production
• Open-source model inference
• LoRA and QLoRA fine-tuning

Template availability is not proof that every model and workflow fits every card. Check current task-fit blockers and published task evidence. Acceptance examples are limited samples, not a production success-rate guarantee.

Save or export the requested outputs before destroying a rental. A GPU’s local disk is not permanent storage. Automatic output collection must be supported and explicitly enabled by the live contract; Workspace capacity, budget and expiry remain separate. Finally, verify the result and release status instead of assuming a successful command means the GPU was released.

For the longer account, Job and Workspace walkthrough, see GPU Jobs for AI Agents.