What it is. Blackwell is the name of Nvidia's current-generation GPU architecture, announced in March 2024 and now shipping in volume. It succeeds the Hopper architecture (the H100, H200) that defined the first wave of generative-AI training. Blackwell-class products include the B100 and B200 GPUs, the GB200 Grace-Blackwell superchip (one Grace CPU plus two B200s on a board), and the GB200 NVL72 rack, which links 72 B200 GPUs into one coherent compute domain over Nvidia's NVLink fabric. The mid-cycle refresh, called Blackwell Ultra or Blackwell 300, pushes memory bandwidth and HBM capacity further.
Why it matters. Two reasons. First, the per-GPU jump in training throughput versus Hopper is roughly 2.5x for FP8 workloads and meaningfully more for the new FP4 path that Blackwell introduces. That changes the unit economics of frontier-model training. Second, the NVL72 rack-scale design lets a single domain hold model weights that previously required slow cross-rack networking, which is what makes trillion-plus parameter models tractable to train in reasonable time.
Who buys it. Almost all of it ships to a handful of buyers: Microsoft, Meta, Google, Amazon, Oracle, CoreWeave, and the sovereign-AI projects in the UAE and Saudi Arabia. Together they account for the overwhelming majority of Nvidia's data-center revenue, which is itself the overwhelming majority of total revenue. Blackwell-class systems were over 70% of Nvidia's data-center business as of the Q1 print covered in our earnings coverage.
What's next. Nvidia's stated roadmap has a one-year cadence: Rubin in 2026, Feynman after that. Each generation is expected to keep the same NVL rack pattern but increase per-rack compute and memory. The capex implication for hyperscalers is that the GPU refresh cycle just got shorter, which is part of what's driving the elevated capex coverage on this site.
Bottom Line
Blackwell is the chip your AI capex is buying right now. When you read "Nvidia data center revenue," "GB200 backlog," or "AI training cluster," picture a rack of these.