NVIDIA Blackwell Ultra

Rent B300 servers, reserved from data centers

B300 is the Blackwell Ultra GPU, usually rented as HGX B300 servers with eight GPUs and x86 CPUs. Each GPU carries 288 GB of HBM3e, so a single eight-GPU server holds models that needed two servers on the previous generation.
Memory per GPU
288 GB HBM3e
Architecture
Blackwell Ultra
Listings
On request
From
Ask

B300 capacity listed now

No B300 capacity is listed right now. Send a request and we will source it from our owners.

Tell us what you need

What B300 is good for

  • Training and fine-tuning large models on eight-GPU nodes
  • Serving 400B+ parameter models from one server
  • FP4 and FP8 inference at high throughput

What to ask before you reserve B300 capacity

  • The exact server (HGX B300 and its CPU and RAM), and the local NVMe per node.
  • How the servers are interconnected (InfiniBand or RoCE, and at what speed per GPU) if you will train across nodes.
  • Whether the price covers storage, egress and support, or only the GPU-hours.
  • The uptime they commit to, and what you get back when they miss it.

Questions

How much does it cost to rent a B300?
Owners set their own price per GPU-hour, which depends on the number of servers, the term and the start date. Send a request with what you need and we come back with offers.
What is the minimum term for reserved B300 capacity?
Usually one month or more, set by each owner. Longer terms usually get a lower price per GPU-hour.
How much memory does a B300 have?
288 GB HBM3e per GPU (NVIDIA Blackwell Ultra).
Can I rent a single B300 server?
Yes, where the owner allows it: each listing shows how many servers it has and you can ask for one or for all of them.

Need a few hours rather than a reservation? See B300 pricing by the hour.

Tell us what you need

Sign in to send a request. It opens a conversation with the owners and the Paralon team, and you get an email for every reply.

Sign in to request

Free, with Google, GitHub or an email code. No commitment.