RTX 4090 or RTX 5090: which to rent

The RTX 4090 is an Ada Lovelace card with 24 GB of memory. The RTX 5090 is its Blackwell successor with 32 GB. Both are consumer cards, and most providers rent them one at a time.

The eight extra gigabytes are the practical difference. They decide whether a larger model loads at all, how large a batch you can run, and how much room is left for long prompts or large images.

On 12 Nov 2026, the median listed on-demand RTX 4090 price here was $0.435 per GPU-hour and the lowest was $0.33. On the same day the median listed RTX 5090 price was $0.655 per GPU-hour and the lowest was $0.36.

Runcrate can start an RTX 4090 now from $1.01 per GPU-hour. Runcrate can start an RTX 5090 now from $0.72 per GPU-hour.

Stay on the RTX 4090 when the model and its working memory fit in 24 GB with room to spare, and when you count the work in images or tokens per dollar. Image generation at common sizes, small language models and adapter fine-tuning of small models all sit here.

Stay on it too when the card waits on you. If most of the session is you reading results and changing settings, a faster card finishes sooner and then idles at a higher price.

Move to the RTX 5090 when you are trimming a job to fit: lowering the batch, shortening the context or dropping the resolution to stay inside 24 GB. Time spent working around memory is paid time.

Move too when the job is long and the card is busy throughout. A newer card that finishes sooner can cost less in total at a higher hourly figure. Measure it. Run the same short job on both and divide each bill by the work done.

If one card must hold more than 32 GB, the RTX Pro 6000 has 96 GB. It is listed for more per hour than either consumer card and for less than the datacenter GPUs.

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.