LLM and foundation-model training
288 GB per GPU and 1.8 TB/s NVLink keep large models and their training state on the GPUs; the 1.6 Tb/s fabric scales across up to 32 servers.
LLM training →Specifications
HGX B300 node specifications
Each node is an 8-GPU NVIDIA HGX B300 system built by Supermicro. Nodes are connected on a 1.6 Tb/s fabric and delivered as bare metal: your operating system, your drivers, your orchestration.
| Specification | NVIDIA HGX B300 node |
|---|---|
| System | 8-GPU HGX B300 system, built by Supermicro |
| GPUs per node | 8 × B300 SXM (Blackwell Ultra) |
| GPU memory | 288 GB HBM3e per GPU · 2.3 TB per node |
| Compute (FP4) | 120 PFLOPS per node |
| GPU interconnect | 1.8 TB/s NVLink, GPU to GPU |
| Node fabric | 1.6 Tb/s, ConnectX-8 ready |
| Power | 3,000 W redundant, Titanium-grade power supplies |
| Storage | Encrypted NVMe on every node (AES-256 at rest, TLS 1.3 in transit) |
| Tenancy | Single-tenant: dedicated compute, storage and network |
Specifications as published by NVIDIA and Sapience AI for the configuration deployed. System memory per server, storage and networking options are confirmed in your quote.
Cluster sizes
HGX B300 is sold by the cluster, not by the GPU. Aggregate figures below are the per-node specification multiplied by the node count; delivered performance depends on the workload.
| Cluster | GPUs | HBM3e memory | FP4 compute |
|---|---|---|---|
| 8 servers | 64 × B300 | 18.4 TB | 960 PFLOPS |
| 24 servers | 192 × B300 | 55.3 TB | 2,880 PFLOPS |
| 32 servers | 256 × B300 | 73.7 TB | 3,840 PFLOPS |
What it is for
Built for the jobs that need the most memory and bandwidth.
HGX B300 is the node for work that has to keep a large model, its optimiser state and its activations close to the GPUs, across several servers.
Large fine-tunes
Full-parameter fine-tuning of large open-weight models, where weights, gradients and optimiser state together run to terabytes.
Fine-tuning →Inference for the largest models
Models too large for a 96 GB GPU, served with the highest throughput from 2.3 TB of GPU memory per server.
Inference →HGX B300 or RTX PRO 6000?
How the two Blackwell nodes compare on memory, interconnect, terms and availability.
Read the comparison →Questions
HGX B300 questions.
Anything else: sales@sapienceai.eu or +421 233 329 562.
What is the minimum HGX B300 reservation?
One cluster of 8 servers (64 GPUs) on a multi-year term. Clusters come in 8-, 24- and 32-server sizes and are sold whole.
When can an HGX B300 cluster be delivered?
Reservations are open now. First deliveries are scheduled from December 2026; the delivery date for your cluster is confirmed in the contract.
How is HGX B300 priced?
Per GPU-hour on a reserved term, quoted in writing. Rates are in US dollars and exclude VAT; the quote confirms rate, term, delivery date and availability.
Where does the cluster run, and who operates it?
In a colocation data centre in Bratislava, Slovakia, on systems operated by Sapience with 24/7 monitoring by our own engineers. Customer data is processed and stored in the EU.
Can I run Kubernetes or SLURM on it?
Nodes are delivered as bare metal today, so you can install the scheduler and stack you already use. Managed Kubernetes and SLURM are in development on the same infrastructure.
Reserve HGX B300 capacity.
Tell us the cluster size, term and start month. The quote confirms rate, availability and the delivery date in writing.