ToolNavs Find Useful AI Tools
Submit Sign in
Back to AI information
Vera Rubin Arrives on Microsoft Azure: Next-Generation AI Chips Start in the Cloud, Enterprises Pilot First

Vera Rubin Arrives on Microsoft Azure: Next-Generation AI Chips Start in the Cloud, Enterprises Pilot First

AI information • Admin • • 28 views

On November 19, 2026, NVIDIA announced at Microsoft Ignite that Vera Rubin NVL72 systems are already running in Microsoft Azure data centers, and access to Azure virtual machines powered by the systems is expanding in preview. The same day, NVIDIA's Nemotron open models moved more broadly into Microsoft Foundry. The partnership bundles 'next-generation compute' with 'open-model customization,' both aimed at enterprise AI adoption.

Running in the cloud first; no price or timeline yet

Microsoft became the first cloud to bring up a Vera Rubin NVL72 system for validation back in March, and delivered the first production systems in August. NVIDIA disclosed no virtual machine series, region, price or general-availability date this time; preview remains the operative word. The performance case comes from MLPerf v6.1: Azure trained DeepSeek-V3 on 10,240 GB200 NVL72 GPUs, the largest scale in that round, and GB300 NVL72 led cloud submissions on DeepSeek-R1 interactive throughput by more than 10%. On the model side, Foundry Managed Compute now offers the Nemotron 3.5 lineup: Super for long-running agents, Lightning for fast responses and tool use, with Bolt and Transcribe to follow. EY has already used Foundry's Programmed Training to fine-tune Nemotron for classifying corporate purchasing records, improving accuracy on its evaluation set by 55% at 15 times lower cost than proprietary frontier models.

What enterprises get is a pipeline, not just a chip

For enterprise users, the value of this update is not a single performance number but a connected pipeline: training and inference in the cloud, on-premises deployment and Windows endpoints all using the same models and tools. When Vera Rubin topped the MLPerf training charts on debut, doubled single-card performance had already been proven; this round adds 'where it runs and how it is customized.' Azure Local will support Nemotron 3.5 and RTX PRO 6000 Blackwell GPUs, Jetson Thor and Jetson Orin cover industrial scenarios, RTX Spark Windows PCs can run large language models locally, and Direct Stream for Windows 365 Cloud PCs has entered limited preview. The boundary is clear, though: the Vera Rubin virtual machines are preview-only, with no price or availability timeline, so what enterprises can do now is validate and pilot, not buy at scale.

Microsoft is itself a user of the stack: Dynamo inference software supports reinforcement learning after training, and Bing search models run on RTX PRO 6000 Blackwell GPUs. Two things to watch next: when the preview turns into general availability, and whether enterprise adoption data for Nemotron on Foundry can support the 'open replaces proprietary' claim.

Recommended Tools

More