NVIDIA DGX B300
DGX B300 is NVIDIA's factory-built Blackwell Ultra appliance: eight B300 GPUs with 2.3 TB of HBM3e, dual Xeon 6776P hosts, eight ConnectX-8 NICs at 800 Gb/s each, and a single NVIDIA support contract in a 10U air-cooled chassis. It replaces DGX B200 at the top of the DGX line for inference and reasoning workloads, trading B200's stronger FP64 for 50% more GPU memory and about 1.5x the dense FP4 throughput. It is also the first DGX deployable in MGX racks, with front-facing I/O and a 54 VDC busbar variant alongside the classic 200-240 V PSU version. Same buyer logic as every DGX: pay the premium for NVIDIA-validated integration and support, or build the equivalent HGX B300 OEM box and keep BOM control.
- DGX B300 is air-cooled at 14.5 kW in a 10U chassis. That is a legacy-colo-killer, not a liquid-cooling problem: no coolant loop needed, but very few air-cooled racks deliver 14.5 kW to one node, let alone several, and 1,500 CFM per node makes hot-aisle containment mandatory.
- The inlet ceiling dropped to 30 C (vs 35 C on DGX B200). Facilities running warm-aisle setpoints for efficiency need to re-check their supply-air temperature before swapping B300 nodes into B200 slots.
- 144 PFLOPS FP4 is a marketing figure; dense per-GPU NVFP4 is 15 PFLOPS (120 PFLOPS per node). Size clusters on dense numbers and your own model benchmarks, not the headline.
- The 2.3 TB of HBM3e is the real purchase driver: a single node now holds a ~600B-parameter model in FP8 with KV-cache headroom. If your working set fits in DGX B200's 1.44 TB, the B300 premium buys you little for today's workload.
- ConnectX-8 at 800 Gb/s means Quantum-X800 InfiniBand or Spectrum-X800 Ethernet switch fabric to exploit it; pairing with a 400 Gb/s fabric halves the scale-out bandwidth you paid for.
- FP64 regressed versus B200. Teams with HPC-adjacent workloads (CFD, simulation) mixed into their AI cluster should keep B200 or H200 nodes for those jobs rather than assuming newer is faster everywhere.
How much does a DGX B300 cost?
NVIDIA does not publish a list price. Reported baseline pricing is $300,000 to $350,000 as of Q1 2026, with fully configured reseller quotes running $400,000 to $500,000 depending on support term and region. Quotes go through NVIDIA partners.
Is the DGX B300 air-cooled or liquid-cooled?
Air-cooled. NVIDIA's user guide specifies front-to-back airflow of roughly 1,500 CFM and a 14.5 kW maximum draw in a 10U chassis, with a 10 C to 30 C inlet window. No liquid loop is required, but the per-rack power density usually forces high-density colo space anyway.
DGX B300 vs DGX B200: which one?
DGX B300 carries 2.3 TB of GPU memory (vs 1.44 TB), about 1.5x the dense FP4 throughput and 2x the attention throughput, plus 800 Gb/s ConnectX-8 networking. DGX B200 keeps a significant FP64 advantage and a wider 35 C inlet window. Reasoning/inference at scale favors B300; mixed HPC workloads can still favor B200.
DGX B300 vs HGX B300: what is the difference?
Same 8-GPU Blackwell Ultra platform either way. DGX B300 is NVIDIA's fixed-configuration appliance with NVIDIA-direct support; HGX B300 is the baseboard OEMs (Supermicro, Dell, HPE, Lenovo, Aivres) build their own servers around, including liquid-cooled and denser variants NVIDIA does not sell as DGX.