custom background image

GPU B300


Scale up with NVIDIA B300 GPU instances

2.3 TB of sole-tenant GPU memory and 24 TB of local NVMe storage to train, fine-tune and serve the largest models, in a reliable and transparent European cloud.

nvidia_b300

Why choose NVIDIA B300 GPUs?

The Blackwell Ultra architecture doubles the memory per GPU compared to the H200. Your largest models fit on fewer GPUs.

Massive

288 GB of HBM3e memory per GPU, twice as much as an H200

Fast

8 TB/s of memory bandwidth per GPU, 1.7× more than the H200, ensuring stable response times, even with long contexts.

Interconnected

The 8 GPUs are connected in NVLink via NVSwitch: your models have 2.3 TB of sole-tenant memory.

Sovereign

Available in our Public Cloud, delivering flexibility, transparency, and compliance with European standards.

Optimised for your most demanding AI workloads

An 8 GPU NVLink instance acts like a single giant accelerator: no splitting the model between multiple servers.

ai agentic

Agentic AI and reasoning

Run more agents and sessions in parallel: the memory freed up by the model accommodates more simultaneous conversations.

ai_models

Long-context inference

Serve context windows holding several hundred thousand tokens to many simultaneous users.

cpu_ai

Native FP4 precision

Blackwell Ultra supports the NVFP4 format, halving a model's memory footprint compared to FP8.

SPECIFICATIONS

Technical specifications

Icons/concept/Hardware/Hardware GPUCreated with Sketch.

GPU

8 × NVIDIA B300 (Blackwell Ultra architecture)

GPU memory

2,304 GB of HBM3e (288 GB per GPU), 8 TB/s of bandwidth per GPU

Ephemeral local storage

24 TB NVMe in RAID 5, up to 80,000 IOPS

Icons/concept/Hardware/Hardware GPUCreated with Sketch.

GPU connection

8 GPUs interconnected via NVLink using NVSwitch

Automation

Management via your Control Panel, API, OVHcloud CLI, etc.

Icons/concept/Cloud/Cloud PadlockCreated with Sketch.

Security and compliance

ISO27001, SOC certifications, health data hosting.

Maximise your ROI with flexible GPU infrastructure

Pricing transparency

Pay only for the resources you use, with no hidden fees. Maintain control over your costs while enjoying optimal performance.

Instant scalability

Scale up or down your GPU resources on demand, in just a few clicks. Easily adapt your capacity to AI and data workloads.

Sovereignty and compliance

Your data is hosted in a secure, transparent, and compliant European cloud, certified to meet regulations (GDPR, ISO, HDS).

Barrier-free accessibility

Open and accessible B300 GPUs: from proof of concept to production deployment, with no limitations on how many you need or the type of hardware used.

Your questions answered

What is the difference between the B300 and the H200?

The B300 has been built on the NVIDIA Blackwell Ultra architecture. It offers 288 GB of HBM3e memory per GPU (versus 141 GB), 8 TB/s of memory bandwidth (compared to 4.8 TB/s) and is compatible with FP4 precision. For the 8 GPU instance, the 8 GPUs are connected via NVLink, versus two groups of 4 for the H200.

What happens to local storage when I stop my instance?

Local storage is ephemeral: its data is erased when the instance is deleted. Use it for working datasets, intermediate checkpoints and weight cache, and keep your reference data on Object Storage or Block Storage.

Can I get a 1- or 3-year contract?

Yes: Get 30% off for 1 year and 50% off for 3 years compared to the no-commitment pricing.

Which models is the 8 GPU instance suitable for?

It is designed for training and fine-tuning very large models, and for long-context inference serving a large number of simultaneous users. With 2.3 TB of GPU memory connected via NVLink, the model does not need to be split across multiple servers.

What is the SLA for GPU instances?

The SLA guarantees 99.99% monthly availability on GPU instances. For further information, please refer to the General Terms of Service.