Cloud GPU for self-hosted AI agents
Price the model-serving workload separately from the agent’s tools and orchestration. Renting a GPU is useful only when you need to host a model rather than call an API.
In our latest check: AWS, Cerebrium, Koyeb, Modal, OVHcloud, Runcrate and RunPod rent all four picks, so one account covers the move from the L4 to the H100: on Runcrate that is $0.58 to $2.85 an hour on-demand.
L4
Compare L4 only if its memory and software support meet your workload requirements. The table supplies the published capacity and current price; it does not claim this card runs your application faster.
L40S
Compare L40S only if its memory and software support meet your workload requirements. The table supplies the published capacity and current price; it does not claim this card runs your application faster.
A100
Compare A100 only if its memory and software support meet your workload requirements. The table supplies the published capacity and current price; it does not claim this card runs your application faster.
H100
Compare H100 only if its memory and software support meet your workload requirements. The table supplies the published capacity and current price; it does not claim this card runs your application faster.
Calculate compute cost per completed task
Enter the output and runtime from your own workload estimate or measurement. No throughput is assumed. Choose hardware only after checking your application requirements.
Compute per completed task = node hourly rate × runtime ÷ completed output. Rates checked 21 November 2026. Include startup, checkpoint and idle time in billed runtime. Storage, egress and other fees are additional.
Worth knowing
- Use the model memory calculator for the exact model and context length.
- Budget GPU uptime as well as active token generation; a process waiting on an external tool may leave a rented GPU idle.
- Compare the expected token bill with the full rental bill before choosing self-hosting.
FAQ
No. These are published hardware and rental-price comparisons. Supply your own memory, runtime and latency requirements before choosing a configuration.
Providers counts every provider listing an offer that meets the pick's memory, in any billing class. 7 days is the average price move of the same on-demand configurations listed a week ago and today, not the change in the price shown. Prices render from today's verified snapshot, not from when this guide was written. Full table on the homepage; break-even math in the calculator.