Hardware

B300 GPU rental pricing: what to compare

GPUIndexes Research··2 min read
THE SHORT ANSWER

Compare full B300 SXM GPUs separately from GB300 system configurations and MIG partitions. NVIDIA specifies 288 GB of nominal memory per B300 GPU. GPUIndexes tracks single-GPU, on-demand B300 offers and publishes a provider-floor median only when at least three providers qualify.

Check which B300 configuration the quote describes

NVIDIA’s DGX B300 documentation specifies eight B300 Blackwell Ultra GPUs with 288 GB of memory each. A cloud listing can expose one GPU, a larger multi-GPU instance, or a partition of a GPU. These are different rental units and should not share one price comparison.

The GPUIndexes B300 basket uses the exact full-GPU b300-sxm catalog model and one GPU per instance. It excludes GB300 entries and B300 MIG partitions. The reference is therefore specific to that catalog configuration rather than a claim about every Blackwell system.

Understand nominal and catalog memory labels

The current source catalog uses a 262 GB label for its full B300 SXM model. NVIDIA’s published nominal specification is 288 GB. We disclose that discrepancy rather than silently presenting the source label as the manufacturer specification.

For the exact b300-sxm model, our matching rule accepts the legacy 262 GB label or the nominal 288 GB label. Downloads preserve both the catalog value and nominal value. This mapping does not apply to GB300 or to reduced-memory partitions, and it is not a promise about the memory available to a particular application. Confirm the configuration with the provider.

Compare the whole commitment

Start with instance GPU count and billing type. A low per-GPU rate inside an eight-GPU server can still carry a much larger whole-instance bill than a one-GPU offer. Storage, CPU, network charges, billing increments and minimum commitments can also change the cost.

The B300 reference page shows the current source timestamp, provider coverage and individual provider floors. It intentionally excludes spot and reserved pricing, so its reference should not be directly compared with a headline promotion or an interruptible offer without reconciling the terms.

Extra memory is not a workload benchmark

GPU memory capacity is one constraint in selecting hardware. Throughput, model precision, sequence length, batch size, software support and interconnect can all matter. A hardware specification alone cannot establish the cost per completed job.

Use the rental reference to understand the observed pricing basket, then benchmark a representative workload and inspect the provider’s final quote. This guide contains no independently measured training or inference performance claims.

Check the observed rental market

These observations update separately from the guide. Use their capture timestamps when citing a price.

B300 SXM rental reference →

No B300 SXM reference is published for the snapshot from Sep 16, 2026 at 21:57 UTC: 0 providers qualified, below the minimum of 3. Quotes and coverage can change; a missing benchmark is not a zero price.

Common questions

How much memory does a B300 GPU have?

NVIDIA specifies 288 GB of nominal GPU memory for B300. The GPUIndexes source catalog currently has a legacy 262 GB label for the full b300-sxm model; the methodology documents that mapping.

Does the B300 benchmark include GB300 or MIG partitions?

No. It includes only the exact full B300 SXM catalog model in single-GPU on-demand instances. GB300 entries and MIG partitions are excluded.

Sources & further reading

  1. NVIDIA DGX B300 hardware specification
  2. RunPod B300 model information
  3. GPUIndexes memory normalization and eligibility rules

Published by GPUIndexes, operated by Quanta Cloud LLC. Ownership, editorial policy and corrections.

Continue reading