Rent a GPU by matching the model to memory first, then launching the offer that is listed now. A short LoRA run or a ComfyUI batch often fits a 24 GB card. A 70B-class chat model, or a long vLLM context, usually needs an H100 or H200.
Which card do you rent?
The GPU list shows what is offered. Each price page shows that card's starting ask and what the memory is for. Pick the class before you compare hourly numbers. The cost note linked below is the longer explanation of why a cheaper card can be the more expensive rental.
Where is the price you will pay?
Open the marketplace after you know the card. The offer there is the one you can launch. The hourly number is the rate billed per second while the pod is running. Stopping the machine stops that clock. Do not copy an hourly figure out of an old article and treat it as today's price.
How do you pay for the rental?
Rent a GPU with crypto once the account exists. Deposit USDT or USDC on Solana, Ethereum, or BNB Chain, USDT on TRON, or SOL or ETH. That balance pays the rental. Per-second crypto billing is explained in the note linked below. Public GPU and template tools can be read without a key. Balance, rentals, and the deposit address need the account.
What do you launch?
Launch a template instead of installing CUDA by hand. vLLM, LoRA, and ComfyUI each have a playbook: one smoke test, then more traffic or a longer run. If the first request runs out of memory, move up a VRAM class. Do not keep a 70B model on a 24 GB card.
What do you record after it starts?
Write down the GPU name, the ask you accepted, and the template. That note is what you use if you relaunch tomorrow. Shut the pod off when the job is done. An idle server is still billed per second.
- Rent a GPU from the live list
- H100 SXM starting ask
- RTX 4090 starting ask
- Launch templates
- Marketplace offers
- What a GPU rental costs
- How per-second crypto billing works