To rent a GPU for AI without overpaying, budget the median, not the floor, and match VRAM to the model. Vast.ai lists an RTX 4090 median of $0.49/hr and a floor of $0.14/hr, about 3.5× apart. Stop the machine when the job ends. A forgotten instance costs more than stepping up a card class.
01. What does it cost to rent a GPU for AI if you ignore the floor?
Budget $0.49/hr for a 4090, not $0.14.
The gap is about 3.5×. Vast.ai lists that median and that floor side by side. The floor is one machine. A week of work does not fit on one lucky listing. Type the median into the budget before you open a host page.
That 3.5× spread is a 4090 fact. It is not a law for other chips. This page does not reprint their rows. When you rent a GPU for AI today, the planning number is the 4090 median, and the floor is the trap beside it.
Chasing the floor burns the evening when that one host is gone. Starting from the median skips the detour. Then you pick the card.
A forgotten instance still costs more than card class on a short job. One median hour is easy to explain. The same hour, left running through an afternoon you did not book, is the overpay. Price the hours you will run, and price the stop in the same line.
A 3090 at a $0.17/hr median is the 24GB alternative.
A 3090 and a 4090 share that 24GB ceiling. The higher 4090 rate buys throughput, not a larger model. Stay with the 3090 when the queue is short and the weights already fit. Move up when a slower card would sit rendering all afternoon and the lower hour stops being lower.
An H100 is a different order.
An H100 SXM holds 80GB. It is a training card, not a faster 4090 for a batch that already fits in 24GB. An A100 is the other datacenter card in this class, at 40GB or 80GB, and this page prints no A100 dollar price. Do not carry the 4090's 3.5× gap onto that row. When the weights do not fit, read the 4090 and H100 rental medians and take the training-card figure from there. Keep image and short-video jobs off the H100 filter.
02. How do you match the card to the job before you type a card number?
Write the VRAM number down before you sort by price.
24GB covers image work and 7B–8B inference. 32GB is the step when video is already tight. 80GB is the step when the weights themselves do not fit. Write the number. Sort by price after.
A 3090 and a 4090 are both 24GB. A 5090 is 32GB. An H100 SXM is 80GB. An RTX PRO 6000 is 96GB. Two 4090s are not one 48GB pool. The 4090 has no NVLink, so plan capacity per card unless the software splits the job.
Headroom is part of the fit. Preview buffers and a video graph can push a weight that looks safe on paper over the edge. Long training is a different problem, because it also needs a disk. For the job today, one card at the right ceiling is the whole choice.
A cheap hour that runs out of memory is a wasted hour.
The rerun is the expensive part. You paid the hour and the file is not there. A lower sticker that runs out of memory is a second bill for the same frame. Count the pipeline, not only the weight. If this workflow already fails on 24GB, the next step is 32GB.
A cluster is the wrong default.
One matched card first. Rent gpu cluster is the phrase for a later failure: one card with the right VRAM still cannot hold the run. That sentence is the whole guide. Do not open a cluster filter while you are still choosing between 24GB, 32GB, and 80GB.
03. How do you stop the first hour disappearing into setup?
A cold shell plus a large weight download is unpaid work at a paid rate.
The meter starts when the instance starts. Weights are often on another network, and a cross-border link bills that download like a frame. When you rent GPU online, a bare shell is the ordinary first-hour leak. Refuse it. Do not pay a live rate for a progress bar.
Prefer an image that already holds the tool you came to run.
ComfyUI-class environments are the pattern. Confirm the image on the listing before the meter starts. How to rent a GPU for image generation is that check. A CUDA title does not mean your graph is installed. Rent a GPU for ComfyUI only when the image names it, then run one small frame before the batch.
Task rental skips the shell.
You send the job and take files back. That pattern is what SpareHive is building.
SpareHive is a waitlist for task-based image and video jobs: a preinstalled environment, a prompt or job in, files back, no SSH, and no multi-gigabyte weight download over a slow link. Per-second billing and a $5 minimum top-up are the plan. The marketplace is not taking rental jobs yet. Join the waitlist.
04. Which billing habits create the overpay?
The increment matters less than a machine left running.
Confirm on the listing whether it bills per second, per minute, or per hour. The gpu rental billing increment is a line on that offer, not a site-wide rule. Per-second billing trims a job you actually stop. Per-hour billing can round a short render up. Neither one saves a machine left on through the night. Read the unit. Then write the stop into the job.
Stop the instance as part of the job, not as a chore for later.
"Later" is the bill. Stop in the same sitting as the copy. When the files are on your side, the instance ends. A note for the morning is how a median hour becomes a night you never ordered. Last frame, copy, stop, then close the laptop. If you cannot stay for the copy, do not start.
A small balance is a useful fuse.
SpareHive's planned minimum is a $5 top-up, and it is not live. On a marketplace you use today, a small balance limits a forgotten night. Fund this job. Do not park a large balance under a machine you might leave on.
05. When is a cheaper interruptible listing the wrong saving?
Interruptible is cheaper because the host can take the card back.
Use it when the job restarts cleanly. Interruptible GPU rental trades a lower sticker for a reclaim. The host can take the card. Your run stops. You keep what you already stored. This page does not invent a discount percentage. Listings move. The test is the resume, not the size of the markdown.
A render that cannot resume does not belong on an interruptible card.
The second full run erases the discount. A reclaim that returns you to frame zero bills the job twice. Use a reclaimable card only when a checkpoint you store can continue the run. If you cannot name that file, pay for a listing the host does not pull.
Long training with checkpoints is a server problem.
A multi-hour train that needs a disk is not a billing trick. When the work has to survive the instance, rent a GPU server for a training run. This page is for a job you finish, copy, and shut off.
Ten minutes before you start, score the offer on five lines.
| Check | Pass | Fail |
|---|---|---|
| VRAM | Model fits, with headroom for the pipeline | Price sorted first, memory second |
| Price you type into the budget | 4090 median $0.49/hr | 4090 floor $0.14/hr |
| Setup | Tooling already on the image, or a task submit | First hour spent pulling weights |
| Stop rule | Instance stops when the file is downloaded | "I'll check it in the morning" |
| Job length | Under a few hours, restartable | Multi-hour train that needs a disk |
06. What do you check in the ten minutes before you start?
VRAM, median price, disk, and the stop button.
Four checks. The model fits, with headroom. The budget shows the 4090 median of $0.49/hr, not the $0.14/hr floor. The disk can hold the outputs until you copy them. You have found the stop control and you will use it before you leave. Fail one and you do not start.
Ten minutes of reading is cheaper than the first hour of a bad template. Open the listing and read those four fields before the image boots. A disk smaller than the batch is a fail even when VRAM is right. Find that out now, with the meter off.
Download path for the outputs.
Ephemeral machines wipe. Take the files with you. Copy, then stop. A stop that lands before the copy deletes the only good frames and sends you back to a paid rerun. If the offer says the volume dies with the instance, believe it. The stop button is not an archive.
That is the preflight. Median in the budget. Tooling already on the image, or a task instead of a shell. Files copied to your machine, then the stop. A long train that needs a disk belongs on the server guide.
FAQ
Is the cheapest listing the right way to rent a GPU for AI?
No. On Vast.ai the 4090 floor is $0.14/hr and the median is $0.49/hr. Plan on the median, then pick the smallest card that holds the model. The floor is one listing, not a monthly plan.
Do I need SSH?
No. Not for a finished image or a short video. A task flow takes a job and returns files. You need a shell when you are debugging a train, and that belongs with a server you can store to.
How do people overpay in the first hour?
They pay a live rate while weights download, then leave the instance up. Stop the machine when the outputs are on your side. Prefer an image that already has the tool installed.
Are interruptible GPUs safe for training?
Only if the run restarts from a checkpoint you store. If a takeover means starting over, the lower sticker price costs more. Keep a non-resumable train on a listing the host cannot pull.
Is SpareHive open for rentals?
No. It is a waitlist for task-based image and video jobs. Per-second billing and a $5 top-up are the plan, not a live checkout.