GPUs with 96 GB VRAM and more
Cards with at least 96 GB of video memory start at €2.12 an hour on VRAMfarm (RTX PRO 6000). 96 GB holds a language model of about 145B parameters in 4-bit or 43B in FP16. For images and video: any open image or video model, several of them in memory at once.
96 GB on the RTX PRO 6000 Blackwell is the most memory in a single workstation card: 70B with very long context, big video models and heavy 3D scenes.
Cards with 96 GB or more, cheapest first
| GPU | Type | Memory | TFLOPS FP32 | Where | Price per hour | ≈ per month | Action |
|---|---|---|---|---|---|---|---|
| RTX PRO 6000 | workstation | 96 GB | ≈125 | +8 | €2.12/h | €1,526 per month | Rent |
| H200 | data centre | 141 GB | ≈67 | +2 | €5.43/h | €3,910 per month | Rent |
| B200 | data centre | 192 GB | ≈80 | €16.12/h | €11,606 per month | Rent |
To launch, your balance must cover one hour of the machine. You are billed by the minute for actual use.
What runs on 96 GB
Language models
Up to about 145B parameters in 4-bit (AWQ, GPTQ, GGUF Q4) or 43B in FP16, with room for a normal context. Long context and many parallel users need more. The VRAM calculator works it out for your model.
Images and video
Any open image or video model, several of them in memory at once.
Training
QLoRA of 70B, LoRA of 32B with a 16-bit base, full fine-tuning of 8B with an 8-bit optimiser.
Typical jobs at this size
- 70B models with long context
- Full fine-tuning of 7–8B models
- Big-batch inference with vLLM
Frequently asked questions
What is the cheapest GPU with 96 GB or more?
The NVIDIA RTX PRO 6000 Blackwell, from €2.12 an hour on VRAMfarm, billed by the minute.
Which LLM can I run on 96 GB?
About 145B parameters in 4-bit or 43B in FP16, with room for a normal context. For example: Qwen3 14B, Qwen3 32B, Llama 3.3 70B.
Can I combine the memory of several GPUs?
For language models, yes: vLLM and llama.cpp split a model across cards, so 2 × 24 GB can run what needs about 40 GB. For image and video models memory does not add up, one card has to hold the model.
How do I know 96 GB is enough?
Add up the weights (2 GB per billion parameters in FP16, 0.6 GB in 4-bit), the context cache and 10–15% for buffers, or use our VRAM calculator. If the model does not fit, it either fails with “CUDA out of memory” or runs much slower with offloading.
Rent 96 GB of VRAM
Top up from €10, pick a card from the list and start in a few minutes.