Microsoft has deployed a cluster of more than 4,600 Nvidia GB300 NVL72 systems.

In a blog post, Microsoft revealed its new cluster featuring Nvidia Blackwell Ultra GPUs and the next-generation Nvidia InfiniBand network which is set to support OpenAI workloads.

According to the company, this is the "first of many," with Microsoft aiming to scale to hundreds of thousands of Blackwell Ultra GPUs across the company's AI data centers globally.

Each rack of the GB300 NVL72 consumes 142kW of power. It features 72 Nvidia Blackwell Ultra GPUs and 36 Arm-based CPUs, NVLink bandwidth of 130TBps, and up to 40TB of fast memory. Compared to the Blackwell GPUs, Blackwell Ultra GPUs have 2x the attention-layer acceleration and 1.5x more AI compute floating-point operations per second (flops) - up to 1.4 exaflops of FP4 performance.

Ian Buck, VP of hyperscale and high-performance computing at Nvidia, said of the deployment: "Microsoft Azure’s launch of the Nvidia GB300 NVL72 supercluster is an exciting step in the advancement of frontier AI. This co-engineered system delivers the world’s first at-scale GB300 production cluster, providing the supercomputing engine needed for OpenAI to serve multitrillion-parameter models. This sets the definitive new standard for accelerated computing."

In the blog post, Microsoft notes that the deployment required the company to "reimagine every layer of the stack - computing, memory, networking, data centers, cooling, and power - as a unified system."

Within the rack, the chips are connected via NVLink and NVSwitch, and between the racks, the company is using the Nvidia Quantum-X800 Gbps Infiniband to enable customers to scale up how many GPUs are used.

Additionally, Microsoft is using standalone heat exchanger units and facility cooling to minimize water usage, and the company is continuing to "develop and deploy new power distribution models" to support the clusters.

Microsoft claims the deployment is the first "at scale" Blackwell ultra cluster, though AI cloud provider Lambda revealed last month that it had deployed at least one GB300 NVL72 system.