Blackwell Ultra overtakes GB200: what B300 changes compared with B200

On 19 November 2025 NVIDIA reported results for the third quarter of its fiscal 2026 and said that Blackwell Ultra had become its leading architecture across all customer categories. Futurum’s review of the results recorded that GB300 shipments had crossed GB200 and accounted for roughly two-thirds of Blackwell revenue. Data centre revenue for the quarter was $51.2 billion, and Jensen Huang described cloud GPUs as sold out. NVIDIA also said that Rubin remained on track to ramp in the second half of its fiscal 2027.

What changed from B200

NVIDIA’s HGX specification page gives a direct comparison of the eight-GPU baseboards. HGX B300 carries 2.1 TB of GPU memory against 1.4 TB on HGX B200. Dense FP4 Tensor Core throughput rises from 72 to 108 PFLOPS, while the sparse FP4 figure of 144 PFLOPS and the FP8/FP6 figure of 72 PFLOPS (sparse) are the same on both boards. NVLink is unchanged: 1.8 TB/s GPU-to-GPU and 14.4 TB/s in total. Networking bandwidth per board doubles, from 0.8 TB/s to 1.6 TB/s.

The practical reading is that B300 is a memory and inference upgrade rather than a new interconnect generation. A team serving large models, where the size of the model and its context decide how many GPUs are needed, gets about half as much memory again per node and more dense low-precision throughput. A team training at FP8 sees little difference on the specification sheet, and the doubled network bandwidth matters mainly for clusters that scale out across many nodes.

Through 2026 the transition continued. NVIDIA attributed data centre revenue of $89.0 billion in the second quarter of fiscal 2027, reported in late August 2026, to the ramp of Blackwell Ultra. TrendForce, in a release dated 28 August 2026, expects GB300 rack-scale systems to be NVIDIA’s main shipment driver in 2026, with demand holding through the first half of 2027. It also expects the five largest North American cloud providers to take about 60% of procurement in 2026, down from about 70% in 2025, as demand from other buyers grows.

Pricing pressure has moved the other way. Futurum noted that NVIDIA’s procurement commitments rose to $279 billion, mainly tied to memory, and that customers had been told of AI server price increases above 15% in some cases.

What it means for buyers

For new eight-GPU servers the default specification is now HGX B300, and B200 systems are increasingly a matter of available stock rather than new allocation. When comparing quotations, check the exact board (HGX B300 or HGX B200), the GPU memory per node, and the NIC configuration, since B300 systems are designed around twice the network bandwidth and an under-specified fabric wastes it. Ask for a written lead time tied to the specific OEM build, confirm whether the price is fixed or indexed to memory costs, and verify that NVIDIA AI Enterprise or other software licences are quoted separately. InTechnologies can source HGX B300 and B200 platforms from OEM channels on request.