Pre-Book NVIDIA B300 GPU

Blackwell Ultra on our AI platform, built according to the NVIDIA reference architecture. Available from mid-January 2027. Pre-book now to reserve your place in the launch queue.

Launching mid-January 2027

How pre-booking works

The B300 is not available to launch yet. Pre-booking registers your interest and holds your place in the queue, so you hear from us before the public launch.

  1. Step 1: Join the launch queue

    Tell us how many B300 GPUs you need, for how long, and what you plan to run. It takes about a minute and no payment is taken on this form.

  2. Step 2: Get pricing and a configuration

    Our team contacts you ahead of launch with B300 pricing, a configuration sized for your workload and the capacity we can allocate.

  3. Step 3: Go live at launch

    When the B300 launches in mid-January 2027, pre-booked requests are served first, in the order they joined the queue.

High-Capacity Memory

288 GB HBM3e per GPU

2.3 TB GPU memory per 8-GPU HGX node

8 GPUs per node

Blackwell Ultra Architecture

Blackwell Ultra GPU architecture

NVLink 5 fifth-generation NVLink and NVSwitch in the node

NVIDIA RA built according to the reference architecture

Networking and Platform

800 Gb/s per GPU

ConnectX-8 SuperNICs

Day 0 Slurm, NGC containers, Dynamo / vLLM / SGLang endpoints, PFS + object storage

B300 Powers the Next Wave of AI

From training frontier models to serving reasoning models in production. Built for the most demanding AI workloads.

Built for trillion-parameter-scale models

Train and fine-tune frontier-scale models on NVIDIA Blackwell Ultra, from a single node to a multi-node cluster.

Trillion
parameter-scale models

Reasoning and agentic AI inference

Reasoning models generate far more tokens per answer. Blackwell Ultra is built to serve them fast, at production scale.

Up to 11x
faster inference vs previous generation

Process 256K+ token contexts

Handle long conversations, entire codebases and large documents in a single context.

256K+
tokens per context

Scale from one node to a cluster

Pre-book a single node or a full cluster, and grow as your training and inference workloads do.

1,024
GPUs in one pre-booking

Pre-booking open

Reserve your B300 before launch.

Join the B300 launch queue. Tell us what you need and we will come back with pricing, a configuration and your allocation ahead of the mid-January 2027 launch.

  • Early access: pre-booked requests are served first at launch.
  • Ready on day 0: Slurm, NGC containers, Dynamo / vLLM / SGLang endpoints, PFS and object storage.
  • B300 pricing shared with you before the public launch.
  • No payment is taken on this form.

0/1000 characters

We use this to answer your question and to reply about it. Nothing else.

B300 pre-booking questions

We expect to launch the B300 in mid-January 2027. Everyone in the pre-booking queue is told as soon as the date is confirmed.

Pre-booking registers your interest in the B300 and holds your place in the launch queue. Our team contacts you before launch with pricing, a configuration and the capacity we can allocate to you.

No payment is taken when you submit the pre-booking form. Commercial terms are agreed with our sales team once pricing is shared with you.

Each HGX node has 8 B300 GPUs with 288 GB of HBM3e per GPU, 2.3 TB of GPU memory per node. GPUs in a node are linked by fifth-generation NVLink and NVSwitch, and each GPU has 800 Gb/s of networking through ConnectX-8 SuperNICs. You can pre-book from one node up to 128 nodes.

B300 pricing has not been published yet. Pre-booked customers receive it ahead of the public launch.

The NVIDIA B200 is available today on E2E Cloud and is a strong choice for large-scale training and inference. You can run on the B200 now and still pre-book the B300 for launch.

Launching mid-January 2027

Be First in Line for the NVIDIA B300

Join the launch queue today and hear from us before the public launch.