A checklist before your first GPU rental

Before a first GPU rental, size the memory, pick the card, price a month, read the terms, ask about storage and egress, and set a way to stop paying.

INFRATOAPP · PUBLISHED 26 OCTOBER 2026 · 5 MIN READ · 3 SOURCES · PRICES LAST CHECKED 20H AGO

The order matters more than the list

Do it in this order: work out the memory your job needs, choose the smallest card that holds it, price the hours you will really use, read the provider's terms, ask about storage and data transfer, and only then launch, with a way to stop the bill already set up. Each step narrows the next one, and skipping the first is how people rent an 80 GB card for a job that fits in 24 GB.

None of the steps takes long. Together they take an afternoon, and they turn first-invoice surprises such as the wrong card, the forgotten machine and the storage bill nobody expected into decisions made on purpose.

When renting for a team, do the steps once and write them down as the team's own routine. The second rental then starts from the notes of the first, and nobody has to rediscover which provider rounds up to the hour or charges for a stopped disk.

Step 1: size the memory before the card

Model weights need roughly the parameter count times the bytes per parameter: 2 bytes in 16-bit, 1 byte in 8-bit, half a byte in 4-bit. A 7 billion parameter model is about 14 GB of weights in 16-bit; a 32 billion parameter model is about 64 GB in 16-bit and about 16 GB in 4-bit.

That is the weights only. The key-value cache and activations take more on top, grow with context length and batch size, and cannot be reduced to one figure. Training needs far more again for gradients and optimiser state. Leave room, and run a short test to see the real peak.

Write the figure down with its assumptions: precision, context length, batch size, and the job: serving or training. When you change any of those later, the memory figure changes with it, and so might the card.

Step 2: pick the card, then compare it

With a memory figure, the cards sort themselves into tiers: 24 GB cards such as the L4, A10 and RTX 4090; 48 GB cards such as the L40S and A40; 80 GB on the H100 and the A100 80 GB part; then 141 GB on the H200 and 192 GB on the B200 and MI300X. Pick the smallest tier with headroom.

Cards with a listed on-demand rate: 51, each with its own page. For a 48 GB job, for example, the L40S page lists every seller (cheapest on record: $0.38). Write down the exact card name including form factor, because the same family name can mean different parts.

Do not let the cheapest rate pick the card for you. A 24 GB card at a low rate that cannot hold the model is no saving, because the job does not run at all. Choose the card by memory first and only then compare its sellers.

Step 3 and 4: price the month, then read the terms

Multiply the rate by the hours you expect, not by 730 unless the machine will never stop. The rental cost calculator does this per seller. Where the result is close to what a reserved term would cost, you have a decision to make later, but not on the first rental.

Then read the terms page you would normally skip. Find the billing increment, any minimum charge, whether a credit card hold is placed, the cancellation rule, and whether the rate you saw is on-demand or something else. How to read a price page, earlier in this series, goes through what to look for.

Also check how you will pay. Some providers want a card on file, some sell credits in advance, and some apply a spending limit to new accounts. A limit you did not know about can stop a run halfway, which costs more than the limit saved.

Step 5: ask about the extras

Persistent storage bills while the instance is stopped, and data moved out may be charged by the gigabyte. Ask for both prices, and estimate the size of your dataset, your checkpoints and what you will download. The post on egress and storage has the full list of questions.

Check also where the data has to live. When a dataset has to stay in one country, the region decides which rate applies and possibly which providers you can use at all.

Note the answers next to the rate in the same sheet. When you look at a second provider later, the blanks in your notes show which questions you forgot to ask the first time.

Step 6: set up the way to stop paying

Before you launch, decide how the instance will stop. Set a spending alert or budget cap if the provider has one. Put a shutdown at the end of your training script. Make a note in the team channel of what is running and who owns it.

Then launch something small first: one card, one short test, the full pipeline from data load to saved output. Look at the first invoice line before you scale up. A mistake found on one card for an hour costs very little; the same mistake on a full node for a weekend does not.

Read that first invoice line by line, even if it is small. Check the hours billed against the hours you think the instance ran, and look for any storage, address or transfer line you did not expect. Where something does not match, ask now, while the sums are small enough to settle quickly.

What we could not answer

⚑The memory arithmetic covers weights only; the cache and activation figures depend on your workload and cannot be given here.
⚑This site holds no data on billing increments, holds or cancellation terms, so step 4 is a list of questions, not answers.

What this rests on

infratoapp, “A checklist before your first GPU rental”, read 27 October 2026. cards with a listed on-demand rate: 51. https://www.infratoapp.com/blog/a-checklist-before-your-first-gpu-rental

Questions people ask about this

What should I check before renting a GPU for the first time?

Check how much memory your job needs, then choose the smallest card that holds it with headroom. Price the hours you will really use, read the billing increment and cancellation terms, ask about storage and egress charges, and set a budget alert or automatic shutdown before launching.

How do I know which GPU my model fits on?

Multiply the parameter count by the bytes per parameter: two in 16-bit, one in 8-bit, half in 4-bit. That gives the weights. Add room for the key-value cache and activations, which grow with context and batch size and need testing, then pick a card with more memory than the total.

Should I start with one GPU or a full node?

Start with one card and a short test of the whole pipeline. It shows whether the data loads, the model fits and the output saves, and it shows you what the first invoice line looks like. Scale to a full node only once the small run works.

These render as FAQPage structured data · answers are filled from the same live figures