Rent a GPU for Open-Source LLM Inference
Choose a card for this task
Check current availability, your quote and its billing minimum before paying. Sources may bill by the minute or by each started hour; an hourly source charges a full hour even if you stop earlier.
You choose the GPU. Model downloads and setup consume rented time. Preserve selected results explicitly; Workspace has separate eligibility, expiry and storage costs.
Choose a GPU yourself, review the current quote, then start the instance. The Ollama deployment path opens configuration; it does not rent a GPU or charge your wallet by itself. Availability and supported environments depend on the selected GPU and supplier.
Prepare the model
Use a model you are licensed to run. Record its repository, pinned revision, quantization and expected memory use. Weights, KV cache and runtime overhead must fit the selected GPU. Download public model weights directly onto the rented instance rather than uploading copies to Workspace unnecessarily. Download and setup time can consume billed runtime.
Know the billing rule before starting
The server quote determines the billing period. Minute sources round runtime up to a minute. Started-hour sources charge a full hour even if you use less than an hour; review and accept that rule before starting. The startup balance reservation and the minimum charge are different values. Optional disk and Workspace storage can have separate costs.
A quoted hourly rate is not a measured cost for a complete model task. Loading time, context length, concurrency and generation length change the total. Check the quote and your actual usage rather than assuming a fixed per-request GPU cost.
Connect your application or Agent
Ollama offers a model download API and an OpenAI-compatible interface after a model is loaded. Confirm the actual instance endpoint and authentication requirements in the console before connecting. Custom inference runtimes require their own setup and compatibility checks. Managed model APIs are a separate API gateway option.
Agents can inspect CloudGPU capabilities and use the MCP documentation. Renting still requires an authorized account, sufficient balance and acceptance of the quoted billing terms.
Save results and stop billing
Export outputs before destroying the instance. Local instance disk is not permanent storage. If you use Workspace, save a version explicitly and check its quota, budget and retention deadline. Confirm the GPU is released in your instance record. Read payment methods and live prices before funding.