NVIDIA B300 vs GB300 Servers: What Blackwell Ultra Changes in 2026

NVIDIA's Blackwell Ultra generation, sold as B300 in HGX servers and as GB300 in Grace-based rack systems, is the Blackwell part most large buyers are being quoted in late 2026. The short version: it carries 50% more GPU memory than B200, faster low-precision inference and twice the network bandwidth per GPU, at higher power. HGX B300 suits enterprises that want an 8-GPU server in a standard rack. GB300 NVL72 is a 72-GPU, liquid-cooled rack for teams serving or training very large models at scale. This guide sets out what actually changed, what each format demands from your facility, and who should buy which.

What does Blackwell Ultra change compared with B200?

NVIDIA's own technical breakdown lists the differences clearly (NVIDIA Technical Blog). Blackwell Ultra raises maximum HBM3E capacity from 192GB to 288GB per GPU while memory bandwidth stays at 8 TB/s. Dense NVFP4 throughput goes from 10 to 15 petaFLOPS, and attention-layer acceleration roughly doubles. Maximum power rises from 1,200W to 1,400W. What does not change is just as important: FP8 throughput and NVLink bandwidth (1,800 GB/s per GPU) are the same as B200.

• Memory: up to 288GB HBM3E per GPU versus 192GB, which lets larger models or longer contexts fit without splitting across more GPUs.
• Inference compute: 15 petaFLOPS dense NVFP4 versus 10, a gain that applies only if your serving stack uses FP4.
• Networking: NVIDIA's HGX comparison lists 1.6 TB/s of networking bandwidth for HGX B300 against 0.8 TB/s for HGX B200 (NVIDIA HGX), through ConnectX-8 SuperNICs at 800 Gb/s per GPU.
• Unchanged: FP8 throughput, NVLink bandwidth and HBM bandwidth, so FP8 training jobs see less benefit than FP4 inference.

One trade-off matters for HPC buyers. NVIDIA's GB300 NVL72 page lists 100 TFLOPS of FP64 for the whole rack (NVIDIA GB300 NVL72), while the GB200 NVL72 page lists 2,880 TFLOPS of FP64 Tensor Core performance (NVIDIA GB200 NVL72). If double-precision simulation is a meaningful part of your workload, B200, GB200 or H200 may be the better fit.

What is the difference between HGX B300 and GB300 NVL72?

HGX B300 is an 8-GPU baseboard that server makers build into x86 systems with Intel Xeon or AMD EPYC host processors. The eight GPUs share an NVLink domain, and you scale beyond one server over InfiniBand or Ethernet. NVIDIA lists 2.1 TB of total GPU memory for HGX B300 against 1.4 TB for HGX B200, and 108 petaFLOPS of dense FP4 against 72 (NVIDIA HGX). Dell's PowerEdge XE9780 lists the HGX B300 option as eight 270GB, 1,100W SXM6 GPUs, so per-GPU memory and power in HGX servers can sit below the 288GB and 1,400W chip maximums (Dell).

GB300 NVL72 is a complete rack: 72 Blackwell Ultra GPUs and 36 Grace CPUs joined by a single fifth-generation NVLink domain of 130 TB/s, with 20 TB of GPU memory and 37 TB of fast memory in total (NVIDIA GB300 NVL72). Because all 72 GPUs act as one NVLink domain, very large mixture-of-experts models can be served without crossing the slower scale-out network. You buy it as a rack-scale system, not as individual servers. Our explainer on GB200 NVL72 rack-scale design covers the architecture, which GB300 keeps.

How much power and cooling does a B300 server need?

Plan for considerably more than an H100 or H200 server. The air-cooled Dell XE9780 ships with twelve 3,200W Titanium power supplies (Dell), which tells you the class of feed and PDU capacity each chassis needs. Dell also announced direct liquid-cooled XE9780L and XE9785L models for denser deployments (Dell Technologies). In practice, many air-cooled halls can host one or two HGX B300 servers per rack before they run out of power or airflow.

GB300 NVL72 has no air-cooled option. Lenovo's product guide rates the rack at 135kW TDP with peaks up to 155kW, fed by six or eight 33kW power shelves, with roughly 90% of heat removed by liquid and facility inlet water accepted at up to 45°C (Lenovo Press). That means coolant distribution units, facility water loops, reinforced floors and new power distribution before the rack arrives. Our guides to liquid cooling for Blackwell and 1,000W-plus GPU servers go into the facility work.

Who should buy HGX B300, and who should buy GB300 NVL72?

• HGX B300 fits enterprises and research groups serving or fine-tuning large models in an existing data center, especially where the extra memory lets a model fit on one 8-GPU node instead of two.
• GB300 NVL72 fits AI labs, cloud providers and very large enterprises serving trillion-parameter or large mixture-of-experts models, with liquid cooling already built or funded.
• B200 remains sensible when FP8 training is the main job, because FP8 and NVLink performance are the same and B200 servers draw less power. See our HGX B200 buyer's guide.
• H200 or PCIe GPUs still make sense for moderate inference, FP64 work, or racks that cannot take 10kW-plus servers.

Is B300 available in 2026?

Yes, though large allocations go first to hyperscalers. Microsoft Azure brought a production GB300 NVL72 cluster of more than 4,600 GPUs online in October 2025 (NVIDIA Blog), and AWS made P6-B300 instances generally available in November 2025 (AWS), adding regions through 2026, most recently Jakarta in September (AWS). NVIDIA reported record data center revenue of $89.0 billion for the quarter ended July 26, 2026 (NVIDIA), and demand still runs ahead of supply, so expect allocation-based lead times and short quote validity rather than shelf stock.

Should you buy B300 now or wait for Vera Rubin?

NVIDIA says Vera Rubin is ramping into full production and partners began shipping Rubin racks in September 2026, but early volume is going to the largest cloud customers, and Rubin racks need even more power and liquid cooling. If you need capacity in the next two or three quarters, Blackwell Ultra is the realistic choice. Our guide to the Vera Rubin release timeline sets out the decision in detail, and GPU shortage strategies covers how to secure supply.

What should be on your B300 request for quote?

• The workload and precision: FP4 inference, FP8 training or FP64 simulation changes whether B300 is worth its premium over B200.
• Air or liquid cooling, available power per rack, and floor loading at the target site.
• Host CPU, system memory and boot and data storage, since DDR5 and NVMe costs have risen sharply in 2026.
• Scale-out network: 400G or 800G InfiniBand or Ethernet, switches and optics, which have their own lead times.
• Warranty and support tier, and whether NVIDIA AI Enterprise software licensing is included.

Nexus Compute supplies new GPU servers and components through authorized distribution, with the full manufacturer warranty, and checks every configuration against your power, cooling and platform before dispatch. Send us your workload and site constraints and we will quote B200 or B300 options within 48 business hours, with purchase orders and net terms available for qualified organizations.