NVIDIA doesn't do incremental. The GB300 — the successor to the H100 and B100 — is the latest proof of that philosophy, delivering architectural changes that go beyond simply adding more cores or memory bandwidth.
What's New in Blackwell Ultra
The headline number is 2.5× training throughput, but that figure understates the architectural shift. The GB300 introduces a new "micro-transformer" execution unit that handles attention layers natively in silicon, rather than decomposing them into general matrix operations. This alone accounts for roughly 40% of the throughput gain on transformer-based models.
Memory has also been rethought. The GB300 uses a stacked HBM4 configuration with 288GB of on-chip memory — enough to hold a 70B parameter model entirely in VRAM with room for a large batch. For inference at scale, this eliminates the weight-offloading overhead that currently limits throughput on A100/H100 clusters.
What It Costs
A single GB300 SXM module is expected to land around $45,000. Cloud instances will follow the usual premium, but hyperscalers are already in queue — AWS, Azure, and Google Cloud all confirmed GB300 availability in H2 2026.
The Competition
AMD's MI350X and Google's TPU v5p are the main competitors, and both are strong products. But NVIDIA's software moat — CUDA, cuDNN, and the entire training ecosystem built around it — remains the decisive factor for most AI teams. Switching costs are high; performance deltas are, for now, insufficient to justify a full migration.
