Vera Rubin on Azure: running, but still a preview
NVIDIA says Vera Rubin NVL72 systems are running in Azure data centers. They are built for demanding reasoning, agentic applications, and large-scale model training and inference. The wording matters, because several outlets framed this as availability. NVIDIA described Azure's Vera Rubin availability as a preview, with access through Azure virtual machines expanding. One explainer put the practical question plainly: can you rent Vera Rubin on Azure today? NVIDIA calls it a preview and says access is expanding.
The announcement leaves out several things buyers will want:
- No pricing.
- No VM series name.
- No region list.
- No general-availability date.
- No eligibility criteria for the preview.
Microsoft's own timeline gives some background. In March, Microsoft became the first cloud to bring up an NVIDIA Vera Rubin NVL72 system for validation. In August, it delivered the first production Vera Rubin systems to its datacenters. Those are Microsoft's earlier internal milestones. They are not a customer launch date.
Confidential computing and power efficiency
Microsoft and NVIDIA are also collaborating to extend confidential computing to bare-metal and rack-scale AI infrastructure. The idea is attestation. Customers verify the computing environment before they release sensitive models, data or keys. The companies describe this as work in progress across interconnected "AI factory" infrastructure, not a finished service. Teams in regulated industries should treat it as a roadmap item until Microsoft publishes configuration guidance.
The companies are also working on power efficiency through NVIDIA MaxLPS, aimed at getting more performance from the power a data center already has. No measured savings were given.
The benchmark context
NVIDIA cites MLPerf v6.1 results as evidence of the co-engineering. In training, Azure scaled to 10,240 NVIDIA GPUs using GB200 NVL72 systems on DeepSeek-V3, the largest scale in that round. In inference, Azure's GB300 NVL72 systems led cloud submissions on DeepSeek-R1 interactive throughput by over 10%, according to Signal65. Both results come from Blackwell-generation hardware. They say nothing directly about Vera Rubin performance, so don't read them that way.
Foundry: train, customize, deploy
For developers, the more immediately usable part is in Microsoft Foundry. Microsoft's serverless Programmed Training lets developers control post-training workflows while Foundry manages the infrastructure. Foundry Managed Compute adds support for deploying custom model weights. Together they link customizing a model to deploying it.
The headline proof point is EY. It used Programmed Training to adapt Nemotron open models to classify corporate purchasing records. NVIDIA reports 55% higher accuracy on EY's evaluation dataset and 15x lower cost than proprietary frontier models. Treat that as a vendor-reported result for one task on one dataset. It is not a general benchmark. Some commentary reads it as an argument for open models over closed APIs. That is a fair hypothesis, but one case study won't settle it.
Nemotron 3.5 lineup on Managed Compute
| Model | Stated purpose | Status |
|---|---|---|
| Super | Long-running agents (coding, research, business workflows) | Available |
| Lightning | Fast responses and tool use | Available |
| Embed and Rerank | Information retrieval | Available |
| Parse | Structured extraction from documents | Available |
| Content Safety | Detecting harmful content | Available |
| Bolt | High-throughput, low-latency reasoning | Coming soon |
| Transcribe | Speech recognition | Coming soon |
Beyond the cloud: Azure Local, edge and Windows
- Foundry Local and Azure Local: Expanded Nemotron 3.5 support brings reasoning, coding and tool-using agents to Azure Local. Retrieval models can find evidence across text, document images, charts and tables.
- On-premises GPUs: Azure Local will also support NVIDIA RTX PRO 6000 Blackwell GPUs from hardware partners. The partners and server configurations were not named.
- Industrial edge: Azure Local is expanding to Jetson Thor and Jetson Orin for machine vision, robotics and autonomous operations. NVIDIA's Cosmos physical AI models are coming to Foundry Local.
- Windows PCs: The companies will demonstrate local large-language-model experiences on Windows PCs powered by NVIDIA RTX Spark. This is a demo of developers running AI apps on-device. It is not a shipping Windows feature, and no supported-PC list was given.
- Windows 365: Direct Stream is coming to GPU-accelerated Cloud PCs. It supports up to two 4K streams at 60 frames per second with low latency, even over challenging networks. A limited preview is open for sign-ups, with public preview planned for early 2027.
For Windows admins, Direct Stream is the item to watch. It targets graphics-heavy workloads such as design and engineering. The preview is limited, so don't plan a rollout around it yet. Early sign-up is cheap, though.
Microsoft's own use of NVIDIA tech
Microsoft AI uses NVIDIA Dynamo for reinforcement-learning rollout inference during post-training. Dynamo coordinates inference across a large GPU fleet and helps manage requests, model updates and failure recovery. Microsoft also uses RTX PRO 6000 Blackwell GPUs for Bing search models. No scale or results were given for either.
NVIDIA's BioNeMo Agent Toolkit and ALCHEMI Toolkit are also expanding into the Microsoft Discovery catalog, for life sciences and materials and chemical science. An initial integration uses RFD4-Proteina, built with the University of Washington's Institute for Protein Design. It generates candidate protein structures and sequences for further evaluation and lab validation. These are candidates, not validated results.
What IT teams should take from this
- Preview is not production. Vera Rubin VMs and Windows 365 Direct Stream are both preview-stage. Plan around the fact that pricing, regions and SLAs are undisclosed.
- Customization is the practical hook. Programmed Training plus custom weights in Managed Compute is a concrete workflow for teams with domain data. Validate it on your own data, not EY's.
- Check licensing and limits yourself. The announcement doesn't cover model licensing terms or service limits.
- Confidential computing at rack scale is a direction. Don't base compliance decisions on it until it ships with documentation.
- Source caution. Nearly everything here comes from NVIDIA's own blog. Secondary outlets largely repeat it, and one of the submitted articles (Breakingthenews.net) could not be retrieved for review. Independent Microsoft documentation on pricing and availability would be the real test.
The broader picture is one vendor pairing spanning silicon, cloud, edge and the desktop. The details that decide whether it's useful, such as availability, cost and supported hardware, mostly haven't arrived yet.
References
- Microsoft to bring Nvidia's Vera Rubin systems to Azure - Breakingthenews.net Breakingthenews.net · Thu, 19 Nov 2026 20:45:00 GMT
- Nvidia and Microsoft put Vera Rubin systems in Azure preview and add Nemotron models to Foundry at Ignite mlq.ai
- NVIDIA Vera Rubin on Azure: Ignite 2026 Announcements - explainx.ai explainx.ai · 2026-11-19T00:00:00+00:00