Domain names illustration

NVIDIA RTX PRO 6000 Blackwell GPU instances


NVIDIA RTX Pro 6000 Blackwell

Inference, RAG, agents or fine-tuning: with the RTX PRO 6000 Blackwell, your models and their context fit on a single card. Get started with 1 or 2 GPUs, then scale up to 4 or 8 as your projects grow, all in a reliable and transparent European cloud.

The memory you need when you're shifting up a gear

Enough memory for your models, latest-generation architecture and a range that grows with you: everything you need to put your AI projects into production.
  • 96 GB vRAM

    Your models and their context fit on a single card, with room left for your users.
  • Blackwell generation

    The NVIDIA Blackwell architecture supports the FP4 format, so you can fit more models into a given memory.
  • 1 to 8 GPUs

    Start on 1 GPU, scale to 2 or 4 in the same range as your projects take off.
  • Keep control

    Hosted on our European Public Cloud, with transparent pricing and no hidden fees.

More room for your models, and for your users

  • flexible OVHcloud

    Longer contexts

    The memory left free by the model can hold longer documents and more simultaneous conversations.
  • AI Notebooks

    More user-friendly fine-tuning

    Fine-tune your models with larger batches, without juggling your available memory.
  • AI models

    Multiple models, side by side

    On a 2 or 4 GPU instance, run one model per GPU: an assistant, an embeddings model, a vision model…
  • App Script

    Your usual tools

    Recent versions of PyTorch, vLLM and NVIDIA NGC images support Blackwell: you start without having to relearn anything.

SPECIFICATIONS

Technical specifications

  • GPU

    GPU

    From 1 to 8 NVIDIA RTX PRO 6000 Blackwell GPUs
  • GPU

    GPU memory

    96 GB of ultra-fast GDDR7 per GPU
  • Architecture

    NVIDIA Blackwell, 5th generation Tensor Cores with FP4 support
  • Automation

    Managed via your control panel, API, OVHcloud CLI, etc.
  • Cloud Shield

    Security and compliance

    ISO27001, SOC certifications, health data hosting…

From prototype to production, the RTX PRO 6000 Blackwell supports AI projects that need memory without targeting the largest models on the market.

  • adjustable OVHcloud

    Serve your LLMs

    Get your open-source models into production, with enough headroom to absorb user spikes.
  • innovative OVHcloud

    Connect your documents

    Connect your models to your internal document databases with RAG: long contexts, sourced answers.
  • global OVHcloud

    Deploy agents

    Run agents that chain tools and reasoning, with the memory required to track long sessions.
  • adjustable OVHcloud

    Specialise your models

    Fine-tune a model on your business data, on 1 to 8 GPUs, by the hour or with a commitment.

Ready to get started?

Create an account and launch your services in minutes

Benefit from US$200 free on your first Public Cloud project

Documentation

  • Choosing a GPU for large language model inference

    Assess the memory needed for loading the model, context size, numerical precision, and the expected volume of requests. This planning phase helps you compare available resources without limiting your decision to theoretical power alone. You can also incorporate requirements related to data, preparatory processing, and the scalability of your service.
    Find out more
  • Create and launch a first Public Cloud instance

    Prepare your Public Cloud project, select a region, and choose an instance suited to your workload. The guide explains the key creation and startup steps to verify access to the virtual machine. You can then install your software environment, organise your data, and validate your calculation chain with an already available Cloud GPU resource.
    Find out more
  • Deploy a GPU application on Managed Kubernetes Service

    Configure a GPU-accelerated application in a Managed Kubernetes Service cluster. The guide walks you through its setup via your management interface, as well as its deployment with Helm. This approach helps you prepare container coordination, GPU resource allocation and the application lifecycle before it goes into a production environment.
    Find out more

Frequently asked questions about GPU instances

Are the GPUs of the 2 and 4 GPU instances connected via NVLink?

No. The RTX PRO 6000 Blackwell does not have NVLink: its GPUs communicate via the PCIe bus. This configuration is suitable for serving several models in parallel, one per GPU, or models split across a few GPUs. For a very large model split across several GPUs, choose the 8 GPU B300 instance.

Which NVIDIA RTX PRO Blackwell configurations will be reservable?

Reservations will exclusively concern NVIDIA RTX PRO 6000 Blackwell configurations bringing together 8 GPUs. Two commitment periods are available: 1 or 3 years. You will only be able to subscribe to them when these new instances are launched.

For which projects should you choose the RTX PRO 6000 Blackwell?

For inference, RAG, agents and fine-tuning of models up to approximately 40 billion parameters. Its 96 GB per GPU accommodates the model and its context on a single card, without moving to a higher-range GPU.

How is it different from the L40S?

The RTX PRO 6000 Blackwell offers 96 GB of GDDR7 memory per GPU, twice as much as the 48 GB of the L40S. It is based on the NVIDIA Blackwell architecture, which supports the FP4 format. The L40S remains an excellent choice for more compact models and multimodal GenAI.

When should you choose an H100, an H200 or a B300 instead?

To train large-scale models or serve larger models, our H100, H200 and B300 instances offer very high-bandwidth HBM memory. The 8 GPU B300 instance also connects its GPUs via NVLink for larger models.

What level of availability does a GPU instance guarantee?

The SLA guarantees 99.99% monthly availability on GPU instances. This value serves as a benchmark for developing your service continuity and organising the components associated with your architecture.