GPU Rental Cost: Read the Live Ask First

What is a GPU rental cost?

A GPU rental cost is the ask on one listed machine, multiplied by the seconds that machine stays up. The ask is set on the marketplace at the moment you accept it. The price page for that model is a starting figure for the class, and it can differ from the machine you are about to launch.

Two numbers get mixed together in screenshots. One is the class: how much memory the card has, and which jobs that memory can hold. The other is the offer: the machine that is free right now, and the rate that will be counted per second. Pay the offer. Use the class to decide whether that offer can finish the job.

This page does not freeze either number. Open the live GPU list when you are ready to rent. If the list and an article disagree, the list is the current ask.

What does the rental include?

The rental is time on a specific card. It is not a year-long reservation and it is not a seat on a shared notebook. You choose a memory class, fund a balance of at least $100, and boot from a template such as vLLM, a LoRA setup, or ComfyUI. The GPU charge runs while that machine runs.

USDT or USDC can be deposited on Solana, Ethereum, or BNB Chain. USDT can also be deposited on TRON. SOL and ETH can fund the same balance. The balance pays the rental. Public GPU and template pages can be read before you sign in. The deposit address, the balance, and the rental itself belong to the account.

The steps that turn the account into a running machine are in the how-to linked below. This note stays on the bill: which number is real, which card class the job needs, and when the clock stops.

Which memory class fits the job?

Sort by memory before you sort by the hourly figure. The job fails when the weights or the context do not fit, and a retry on the same card repeats that failure. Move up a memory class when the process runs out of VRAM. A cheaper card in the same class does not add memory.

The table names the class, a card that represents it, and the job that usually fits. Read the matching price page for what that memory is for. Then confirm a live offer. The table is a filter. It is not a quote.

When is 24 GB enough?

A 24 GB card such as an RTX 4090 is the usual fit for a short LoRA run or a ComfyUI batch. It is the wrong place for a 70B-class chat model. If the first request on that card runs out of memory, stop the machine and choose a larger class. Do not keep launching the same 24 GB offer.

When does the job need 80 GB?

An H100 SXM, with 80 GB, is the step up for a 70B-class model or a longer vLLM context. SXM and PCIe cards that share the H100 name do not share bandwidth. Read the H100 SXM page for the class, then confirm a live offer before you treat two H100s as the same rental.

When do you move to 141 GB?

Move to an H200, with 141 GB, when the H100 run is out of memory rather than merely slow. More memory is the fix. A cheaper 24 GB card is not. If the H200 offer is the one listed, that offer is the bill, even when an older H100 screenshot looks lower.

Why do the price page and the marketplace disagree?

The pricing table lines up VRAM, compute, and a starting figure so you can narrow an H200, an H100, an A100, or a 4090. The marketplace lists machines you can rent now. The same model name in both places does not mean the two numbers match. Supply moves. A class can have a starting figure on the price page and a different ask on the only machine that is free.

Use the table to choose a class. Use the marketplace for the ask you will actually pay. Write that ask down with the GPU name. Do not copy an hourly figure into a runbook and reuse it tomorrow. Tomorrow's offer is a new rental.

When does per-second billing stop?

Billing is per second, and the GPU clock stops when the machine stops. An idle server that is still running is still a rental. Shut the pod off when the job ends. A stopped pod does not keep the hourly ask.

A Docker pod can be terminated when the job is done. A root KVM machine can pause GPU billing and leave a standby storage charge, so check which kind you launched before you assume the account is idle. Per-second billing does not make a forgotten running instance free. The longer note on per-second billing is linked below.

What sits outside the hourly ask?

The hourly ask is the GPU clock on a running machine. It is not the deposit, and it is not a fee for reading the GPU list. The deposit has a $100 minimum and sits in the account until a machine is running. A transfer below that minimum, or a transfer on the wrong token or chain, never becomes rental time.

Stopping the machine ends the GPU clock. It does not rewrite the ask you already consumed. Seconds already counted stay on the bill. A relaunch tomorrow is a new offer, so the note you saved is a record of the class and the template, not a coupon for the old rate.

Templates such as vLLM, LoRA, and ComfyUI change how fast you reach a first result. They do not change the ask. Installing CUDA on an empty disk spends the same clock while you wait. Boot the template, run one request, and stop the machine if the memory class was wrong.

Which mistakes make the bill larger?

Five patterns cost more than a slightly higher hourly ask. Picking the lowest number on the page and ignoring VRAM. Keeping a 70B model on 24 GB and relaunching it. Leaving the pod up after the last request. Pasting yesterday's hourly figure into a runbook. Sending a deposit on the wrong token or chain, then waiting on a transfer that cannot credit.

The minimum deposit is $100. Wait for that chain to confirm before you treat the transfer as lost. If it still does not show, check the token, the chain, the amount, and the address on the page. Then send the transaction hash, chain name, amount, deposit address, and account email to support@novapod.io.

What do you check before you deposit?

Open the live GPU list and note the class you need. Read one price page, such as H100 SXM, H200, or RTX 4090, for what that memory is for. Then open the marketplace and accept an offer that is actually listed. Boot a template instead of installing CUDA on an empty disk.

Write down the GPU name, the ask you accepted, and the template. That note is what you compare if you relaunch tomorrow. If the new offer is higher or lower, the new offer is the one you pay. The note from yesterday explains the class. It does not set the price.