Sourced evidence
Azure ND GB200 v6 and ND GB300 v6 VM series: documented NVIDIA GPU configurations for general Azure customers, not just showcase superclusters
Beyond the headline Fairwater superclusters, NVIDIA GPU access on Azure is also offered as documented, generally available VM SKUs. Azure ND GB200 v6 (GA, per Microsoft Learn/Azure Blog): each VM has 2 NVIDIA Grace CPUs + 4 Blackwell GPUs (192GB HBM3e each) linked via 5th-gen NVLink (1.8TB/s per GPU), with 4x400Gb/s NVIDIA Quantum-2 CX7 InfiniBand per VM; 18 VMs form a 72-GPU NVL72 rack delivering up to 1.4 exaFLOPS FP4. Azure ND GB300 v6 (GA as of an Aug 2026 Microsoft Community Hub post): each VM has 2 Grace CPUs + 4 Blackwell Ultra GPUs (288GB HBM3e each, 1.5x FP4 compute vs GB200); a 72-GPU NVL72 rack delivers up to 1.44 exaFLOPS FP4, 800Gb/s per-GPU cross-rack bandwidth via NVIDIA Quantum-X800 InfiniBand, 130TB/s intra-rack NVLink bandwidth, and NVIDIA SHARP in-network compute for collective operations. These are the specific, customer-facing Azure services through which NVIDIA GPU generations reach Azure users, distinct from the Fairwater/OpenAI-dedicated superclusters. RE-VERIFIED against current Microsoft Learn pages (ND GB200-v6, last updated 2026-07-28; ND GB300-v6, last updated 2026-08-12) as part of a Sept 2026 technical-arc audit. Confirmed identical host specs for BOTH series: 128 vCPUs (Grace), 4 local NVMe disks totaling 16 TiB, and 16 remote-storage disks delivering 80,000 IOPS / 1,200 MBps -- storage/IOPS capacity is unchanged between GB200-v6 and GB300-v6 generations; only GPU (Blackwell vs Blackwell Ultra), memory (192GiB vs 288GiB HBM3e per GPU; 900GiB vs 864GiB system LPDDR5X), and per-GPU InfiniBand bandwidth (400Gb/s vs 800Gb/s) differ. GB200-v6 rack: 1.4 exaFLOPS FP4, 13.5TB shared HBM, 130TB/s intra-rack NVLink, 28.8Tb/s scale-out (4x400Gb/s Quantum-2 CX7 InfiniBand per VM x 18 VMs). GB300-v6 rack: 1.44 exaFLOPS FP4 (~1.5x GB200 compute), ~37TB fast memory (~20TB HBM + ~17TB Grace CPU memory), 130TB/s intra-rack NVLink, ~1.1M tokens/sec/rack LLM inference throughput (Microsoft states this is ~27% higher than ND GB200 v6), 800Gb/s per-GPU InfiniBand (2x GB200's per-GPU bandwidth). This confirms a genuine architectural evolution (compute/memory/networking scale up) on a stable storage/compute-host chassis design across the v6 generation.
- Source
- Microsoft Learn (Azure Virtual Machines docs) / Microsoft Azure Blog / Microsoft Tech Community
- Type
- Company filing or release
- Published
- October 9, 2025
- Confidence
- 82/100
- Filed under
- NVIDIA
Supports these links
Cite this
InsightNodes, "Azure ND GB200 v6 and ND GB300 v6 VM series: documented NVIDIA GPU configurations for general Azure customers, not just showcase superclusters" (NVIDIA). Source: Microsoft Learn (Azure Virtual Machines docs) / Microsoft Azure Blog / Microsoft Tech Community, October 9, 2025, https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/nd-gb300-v6-series. https://insightnodes.com/claims/e/nvidia-azure-ndv6-gb200-gb300-vm-series-2025 (accessed October 29, 2026).
This page has a permanent address — link to it from notes, articles or models. Not investment advice.