All products 
RTX PRO 6000 for rent: 96 GB on a single card for large models and rendering
The NVIDIA RTX PRO 6000 is a professional Blackwell-generation card with 96 GB of GDDR7 memory with error correction (ECC). That is more than the 80 GB of an H100 PCIe and twice as much as an RTX 6000 Ada. The main reason to pick a GPU server with RTX PRO 6000 is to run a model on one card that used to need two, without splitting the weights across accelerators. It is still a full graphics card, too: it has RT cores for ray tracing and hardware NVENC encoders, so people rent it for rendering and video as well.
Three cards with similar names
Catalogs make it easy to mix up three different generations, and their prices differ several times over:
- Quadro RTX 6000 — Turing architecture, 24 GB. An old graphics card with too little memory and compute for current neural networks.
- RTX 6000 Ada — 48 GB, the previous generation. Close in design to the L40S.
- RTX PRO 6000 Blackwell — 96 GB of GDDR7, fifth-generation tensor cores with FP4, FP6 and FP8 formats. This is the card collected on this page.
Blackwell comes in variants as well. The Server Edition is built for racks: passive cooling, up to 600 W, with airflow from the server fans. Max-Q is a 300 W workstation version with a blower fan; it is easier to fit several in one chassis, but under sustained load it runs at lower clocks. If the plan line just says "RTX PRO 6000", ask the host which variant it is: speed depends on it, and so does MIG availability.
What fits into 96 GB
A rough estimate: in FP16 a model takes about 2 GB per billion parameters, with 8-bit quantization about 1 GB, with 4-bit 0.5–0.6 GB. On top of that you need headroom for the KV cache, which grows with context length and the number of concurrent requests.
- FP16 without quantization — models up to roughly 30–40 billion parameters with moderate context.
- 8-bit — a 70-billion-parameter model entirely on one card, with room left for the cache.
- 4-bit — models in the 100–120 billion class. FP4 support in the tensor cores speeds up this kind of inference if your engine handles the format.
- Fine-tuning — LoRA for 13–30B models in 16-bit and QLoRA for 70B. Full training of large models is not a job for this card.
- Images and video — image generation in large batches, heavy video generation models that feel cramped in 24–48 GB.
- Rendering and 3D — scenes with huge textures that do not fit in gaming cards, offline rendering, video transcoding on NVENC.
MIG, multiple cards and when an H100 is better
The Server Edition supports MIG: the card can be split into four isolated instances of about 24 GB each, every one with its own memory and compute units. That is handy when several small models or several users share one server. If a plan lists a share of the card with part of its memory, it is a vGPU or MIG slice, not a whole card: compare the memory size, not the model name.
The "×2" and "×4" plans give 192 and 384 GB in total, but the RTX PRO 6000 has no NVLink, so the cards talk over PCIe 5.0. For inference this usually does not get in the way: each card can run its own copy of the model, and splitting one model across two cards over PCIe works acceptably. For training, where cards constantly synchronize gradients, the slower link eats up the gain. If you plan multi-GPU training, look at H100 servers with NVLink.
Driver, CUDA and the image
The card has compute capability 12.0. It needs the R570 driver branch or newer and CUDA 12.8 or later. In older images with CUDA 12.4 the card either will not be detected or the framework will find no kernels for it. Once the server is delivered, check the nvidia-smi output, the CUDA version inside the container, and that PyTorch or your inference engine is built with support for this architecture.
Hourly billing or a month
Renting a server with RTX PRO 6000 in the catalog costs roughly $830 to $5,700 a month, with a typical plan around $1,800. The low end is a single cloud card, the high end is dedicated machines with four cards. Most plans are billed by the hour, which pays off for a trial run, a one-off data processing job or a render against a deadline: you pay only for the hours the machine runs. If the card will be busy more than two thirds of the month, compare with the monthly price and with the dedicated GPU servers section. Before paying, check that the host accepts your card in your currency; if you process personal data of EU residents, also check the data centre country for GDPR purposes and its latency to your users.