Microsoft Azure launches huge Nvidia GB300 cluster for OpenAI
Microsoft has launched the first supercomputing-scale production cluster of Nvidia GB300 NVL72 systems on its Azure cloud platform, built to run OpenAI's most demanding artificial intelligence workloads.
The new ND GB300 v6 virtual machine series packs more than 4,600 Nvidia Blackwell Ultra graphics processors into a single cluster, connected by Nvidia's Quantum-X800 InfiniBand networking. Microsoft says the system is the first of many, with plans to scale to hundreds of thousands of Blackwell Ultra GPUs across its AI data centres around the world.
What the cluster delivers
Each GB300 NVL72 rack combines 72 Nvidia Blackwell Ultra GPUs with 36 Grace CPUs, delivering 37 terabytes of fast memory per rack, according to Microsoft. The chips inside a rack are linked by NVLink and NVSwitch, while racks are connected to each other through InfiniBand networking at 800 gigabits per second per GPU.
Microsoft says the hardware can reach up to 1.4 exaflops of FP4 Tensor Core performance, and that Blackwell Ultra GPUs offer twice the attention-layer acceleration and 1.5 times the AI compute of the previous Blackwell generation. Each rack consumes 142 kilowatts of power, and Microsoft says it has reimagined every layer of the system, from computing and networking to cooling and power, as a unified design, using standalone heat exchangers and facility cooling to minimise water use.
Built for reasoning models
The cluster is designed for the latest reasoning models, agentic AI systems and multimodal generative AI, which need huge amounts of memory and fast communication between processors. In MLPerf Inference v5.1 benchmarks, Microsoft and Nvidia report that GB300 NVL72 systems delivered up to five times higher throughput per GPU on the 671-billion-parameter DeepSeek-R1 reasoning model than the previous Hopper architecture.
Ian Buck, Nvidia's vice president of hyperscale and high-performance computing, said the launch provides the supercomputing engine needed for OpenAI to serve multitrillion-parameter models, calling it the world's first at-scale GB300 production cluster.
Microsoft says the new systems will let customers train models in weeks rather than months, and will be the first to support training models with hundreds of trillions of parameters.
