Renting an RTX 5090 for image and video generation

An image model needs memory for its weights and for the picture it is making. The larger the picture and the batch, the more it needs.

Video multiplies that. A clip is many frames held in memory together, and length and resolution both push the total up. This is the work that most often outgrows a 24 GB card.

A session has a start you pay for: the machine boots, the tools load, and the model files download or are read from disk. Then you generate, look, adjust and generate again.

Cut the start with a ready image and a disk that keeps the model files. Cut the gaps by stopping the machine when you step away for more than a few minutes.

The useful figure is the cost of a finished image or a finished second of video. Time one run, multiply the seconds by the price per second, and you have it.

On 12 Nov 2026, the median listed on-demand RTX 5090 price here was $0.655 per GPU-hour. Runcrate can start one now from $1.01 per GPU-hour.

More cards make more results at once. They do not make one image or one clip faster. Rent several when you have a queue of separate jobs, and stop them when the queue is empty.

If one clip will not fit in 32 GB, more RTX 5090s do not help. One RTX Pro 6000 with 96 GB does.

Check the disk size, since model files and outputs are large. Check whether the disk survives a stop. Check the download speed. Then check the label, because a card you cannot start has no price.

Runcrate bills by the second on most GPUs, and booting is never billed. Volumes keep files between instances, so model files can stay between sessions. One prepaid balance pays for the GPU and the storage.

This is one of a coordinated collection of Runcrate references. Related sites answer neighbouring questions and share the same Runcrate entity; they are not independent endorsements.