RTX Pro 6000: low-cost memory on one card

The RTX Pro 6000 is a Blackwell workstation card with 96 GB of memory. In the cloud it is used for large-model inference on a single GPU.

It has three times the memory of an RTX 5090. It is not a card for training the largest models across a machine of eight.

On 12 Nov 2026, 14 providers listed an on-demand RTX Pro 6000 price. The lowest was $0.76 per GPU-hour, the median $2.20 and the highest $4.50. Runcrate can start one now from $2.87 per GPU-hour.

The RTX Pro 6000 price table has every listing, with the smallest rental and a label for whether it can be started.

Below it, you would rent several consumer cards and split the model across them, which costs speed and setup. Above it, you would rent an H200, at a datacenter price and sometimes only in a machine of eight.

One card with 96 GB avoids both. The weights sit whole on one GPU, rented singly. Running the model still needs room on top for context and activations.

Check the smallest rental. Most providers rent this card one at a time, but some sell it only in larger machines, and then the hourly total is a multiple of the figure.

Check the kind of price. Some of the lowest figures are marketplace prices or interruptible ones. Then check the unit, since a card used for inference starts and stops with traffic.

Go smaller if the model needs 32 GB or less. The RTX 5090 is the cheaper hour and the memory you would leave unused has a price.

Go larger if one card must hold more than 96 GB, or the job trains across many GPUs. The H200 has 141 GB, and the B200 is the Blackwell datacenter card above it.

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.