RTX 5090 or an older RTX 4090

The RTX 4090 is the previous consumer flagship, an Ada Lovelace card with 24 GB. The RTX 5090 is the Blackwell card that followed it, with 32 GB.

A third more memory is the difference you can plan around. It decides whether a larger model loads, how big a batch runs, and how long a prompt or a video clip can be.

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

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

It pays when memory was the limit. If you were cutting the batch, the context or the image size to stay inside 24 GB, the RTX 5090 removes the workaround.

It pays too when the card is busy for the whole session. A card that finishes sooner is on for less time. Run the same job on both, and divide each bill by the work done.

An older card is right when the job fits in 24 GB with room to spare and the card spends much of the session waiting for you. Idle minutes cost less on the cheaper hour.

Check how the older listing is sold. Many RTX 4090 figures are marketplace prices for one owner’s machine. Compare provider-set, on-demand prices on both sides.

Runcrate bills by the second on most GPUs, and booting is never billed, so a side-by-side test costs minutes on each card. Its list has 26 GPU models from $0.23/hr.

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.