NVIDIA HGX B300 · North America

HGX B300 GPU servers for production inference.

Dedicated, single-tenant HGX B300 servers with 288 GB of HBM3e on every GPU. Serve and train large models without buying, racking or operating the hardware yourself.

GPU memory
288 GBHBM3e per GPU
Memory bandwidth
Up to 8 TB/sper GPU
Per server
8 GPUson 1.8 TB/s NVLink
Tenancy
Single tenantbare metal

Get a deployment plan and quote

Share your requirements. We'll send a deployment plan with confirmed dates and a quote.

How do you want to use them?

Pricing is quoted on your workload after a first call.

Specs

NVIDIA B300 GPU specs at a glance.

NVIDIA's published figures for one HGX B300 board, next to a typical PolarGrid B300 node.

HGX B300 board · NVIDIA published
GPUs8 NVIDIA Blackwell Ultra SXM GPUs (B300)
GPU memory288 GB HBM3e per GPU; 2.1 TB per board
Memory bandwidthUp to 8 TB/s per GPU
NVLink5th gen, 1.8 TB/s GPU-to-GPU, 14.4 TB/s total
FP4 Tensor Core144 PFLOPS sparse, 108 PFLOPS dense
FP8/FP6 Tensor Core72 PFLOPS
FP16/BF16 Tensor Core36 PFLOPS
TF32 Tensor Core18 PFLOPS
FP32 / FP64600 TFLOPS / 10 TFLOPS
Networking8 ConnectX-8 SuperNICs, up to 800 Gb/s each

Source: NVIDIA HGX platform page and HGX AI Factory reference architecture. Tensor Core figures include sparsity unless noted. CPU, system memory and storage are set by the server maker.

Typical PolarGrid B300 node
GPUs8x NVIDIA B300
CPUs2x AMD EPYC 9575F or 2x Intel Xeon 6787P
System memory2 to 3 TB DDR5
Local storage4x 3.84 TB NVMe
GPU fabric800 Gb/s InfiniBand
Public network25 GbE

Need more than one server?

We connect nodes over an 800 Gb/s InfiniBand fabric, for multi-node inference, training and fine-tuning. Shared storage can be added.

What fits

What fits on one HGX B300 server.

Weights take about 2 bytes per parameter at FP16/BF16, 1 byte at FP8 and half a byte at FP4, plus room for the KV cache, which grows with context length and concurrent requests.

  • 70B

    On a single GPU

    About 140 GB at FP16 fits on one B300, leaving the other seven GPUs for more replicas.

    ~49% of one GPU for weights

  • 671B

    Large mixture-of-experts models

    Roughly 700 GB at FP8, which leaves well over 1 TB for KV cache on one server.

    ~30% of one server for weights

  • ~1T

    Trillion-parameter models

    Fit on one server at FP8 (about 1 TB) or FP4 (about 500 GB).

    ~43% of one server at FP8 for weights

Starting estimates. Actual usage depends on your serving engine, context length and batch size. Send us your model and we'll size it

Two ways to use them

Dedicated servers or managed inference.

Both run on single-tenant B300 servers in data centers across North America.

Bare metal

Dedicated B300 servers

Bare-metal nodes with full control of your operating system, drivers and serving stack.

  • No shared hypervisor and no other customers' workloads on your hardware
  • We handle power, cooling, networking, burn-in testing and hardware replacement
  • Multi-node clusters on 800 Gb/s InfiniBand, up to a 72-node leaf
API

Managed inference on B300s

Deploy your model on PolarGrid's inference platform and call it through an API, without managing servers.

  • Open-weight models or your own checkpoints, on vLLM, SGLang, TensorRT-LLM or your own image
  • OpenAI-compatible endpoint, or your image's own API
  • Evals, canary releases and automatic rollback
Availability

B300 supply is tight. We source across several suppliers.

New B300 servers are hard to get across the market. PolarGrid contracts capacity across several independent suppliers and data centers, which can mean earlier delivery than ordering hardware directly.

PolarGrid B300 serversBuying HGX B300 servers yourself
SupplyCapacity contracted across several independent suppliers and data centersOne order, in one vendor's queue
OperationsPower, cooling, networking, burn-in and hardware replacement handledYou rack, power, cool and maintain it
Multi-node800 Gb/s InfiniBand fabric, up to a 72-node leafYou design and build the fabric
TenancySingle tenant, no shared hypervisorSingle tenant
Delivery datesConfirmed in your deployment planSet by the vendor's queue

Delivery dates depend on the number of nodes and whether they share one InfiniBand fabric, location, dedicated servers or managed inference, and contract length.

HGX B300 servers

Get a deployment plan and quote.

Share your model, ideal BoM and location. We'll send a deployment plan with confirmed dates and a quote.