AI & ComputeLOT 261120 · 3 min read
Nvidia puts Vera Rubin NVL72 systems into Azure preview
Nvidia and Microsoft opened an Azure preview for Vera Rubin NVL72, added Nemotron 3.5 models to Foundry and extended confidential computing to rack-scale systems.
Key takeaways
- Azure previewed Vera Rubin NVL72 systems at Microsoft Ignite on Nov. 19.
- Azure scaled to 10,240 GB200 NVL72 GPUs in MLPerf Training v6.1.
- Foundry Managed Compute now offers Nemotron 3.5 Super, Lightning, Embed, Rerank, Parse and Content Safety.
- Nvidia said EY reported a 55% accuracy gain and a 15-fold cost cut using Programmed Training.
- Direct Stream supports two 4K streams at 60 fps; public preview is planned for early 2027.

By Grace Kim
Nvidia and Microsoft opened an Azure preview for Vera Rubin NVL72 rack-scale systems at Microsoft Ignite on Thursday, Nov. 19. The systems are running in Azure data centers. Nvidia also said Azure scaled to 10,240 GPUs on GB200 NVL72 systems in MLPerf Training v6.1. The company called that the largest scale in the round.
What does the Azure preview include?
Nvidia described Vera Rubin availability in Azure as a preview. Access through Azure virtual machines is expanding. The companies confirmed that the NVL72 systems are operating inside Azure data centers. Their status remains a preview, not a general release.
At Microsoft Ignite, the companies extended confidential computing to bare-metal and rack-scale AI systems, using Nvidia attestation to check an environment before sensitive models, data or keys are released to it.
Attestation provides the release check for the protected environment. Models, data and keys enter only after the environment passes that check.
Which Nemotron models arrive on Foundry?
Foundry Managed Compute now offers Nvidia’s Nemotron 3.5 family. Availability differs by model.
- Nemotron 3.5 Super is for long-running agents.
- Nemotron 3.5 Lightning is for fast responses and tool use.
- Nemotron Embed and Rerank are retrieval models.
- Nemotron Parse handles document extraction.
- Nemotron Content Safety is the safety model.
Foundry Managed Compute also now accepts custom model weights. Microsoft added Programmed Training, a serverless post-training service. Together, the announcements separate custom model intake from serverless post-training.
Nemotron 3.5 Bolt and a Transcribe speech model are coming soon. Neither model has an announced release date.
Nvidia said EY used Programmed Training to adapt Nemotron models to classify purchasing records, reporting “a 55% accuracy gain on its test set” and “a 15-fold cost cut versus proprietary frontier models.”
The figures apply to EY’s purchasing-record test and its comparison with proprietary frontier models. They are company-reported results, not a general Nemotron benchmark.
What changes for edge and on-premises systems?
Azure Local will support RTX Pro 6000 Blackwell GPUs from hardware partners and expand to Jetson Thor and Jetson Orin platforms for robotics and machine vision.
Azure Local is Microsoft’s on-premises version of Azure. The announcement covers GPU hardware and edge platforms.
Foundry Local will gain wider Nemotron 3.5 support. Nvidia’s Cosmos physical AI models are coming to Foundry Local. These items remain separate from the models already available on Foundry Managed Compute.
What changes for Windows cloud PCs?
The companies demonstrated local language-model experiences on Windows PCs using Nvidia’s RTX Spark chip, which reached mini desktops from Asus, Dell and MSI this week.
Direct Stream for GPU-accelerated Windows 365 Cloud PCs supports up to two 4K streams at 60 frames per second. That is the stated ceiling for each cloud PC.
A limited preview is open for sign-ups. Public preview is planned for early 2027. The two stages separate current sign-up access from the planned wider release.
What does the MLPerf result show?
Azure scaled to 10,240 GPUs on GB200 NVL72 systems in MLPerf Training v6.1, which Nvidia described as the largest scale recorded during the benchmark round.
The benchmark used GB200 NVL72 systems, not the Rubin preview systems. The result establishes Azure’s training scale on GB200 hardware. It does not provide a Vera Rubin capacity figure.
Bing search models run on RTX Pro 6000 Blackwell GPUs. That deployment runs alongside, rather than through, the Rubin preview.
The announcements span training scale, cloud preview, managed model services, on-premises hardware and local PCs. Their status differs: some are running now, while others remain preview, coming-soon or early-2027 items.
EY’s reported cost advantage gives Nemotron a pricing argument. Wider use now depends on Azure expanding Rubin preview access and Nvidia delivering Bolt and Transcribe.
mlq.ai (Source)
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,166 articles