Vera Rubin NVL72 in Azure preview: NVIDIA and Microsoft expand partnership
NVIDIA and Microsoft expanded their AI collaboration at Microsoft Ignite on November 19, 2026. The announcement spans four layers: Vera Rubin NVL72 systems in Azure, confidential computing at rack scale, Nemotron 3.5 models in Microsoft Foundry — and deployment beyond the cloud, from industrial edge devices to Windows 365 Cloud PCs. Here is what was actually announced, and what remains uncertain.
What was announced
According to Unite.AI's coverage of NVIDIA's announcement, the expanded collaboration concerns "Vera Rubin NVL72 systems running in Azure data centers, broader support for Nemotron open models in Microsoft Foundry, and deployment options for on-premises systems, industrial edge devices, and Windows Cloud PCs."
The announcement came at Microsoft's annual Ignite conference. All details in this story come from Unite.AI's rendering of NVIDIA's own description — neither NVIDIA's newsroom nor Microsoft's blog is available as primary sources here, and the two Unite.AI pages that exist are duplicates of the same article.
Layer one: Vera Rubin NVL72 in Azure
The most concretely technical element is expanded access to virtual machines powered by Vera Rubin NVL72 through Azure's Vera Rubin preview. According to the coverage, the systems are designed for reasoning and agentic applications, as well as large-scale model training and inference. That ties the preview to both model development and model deployment.
Two caveats are worth noting. The first is that this is described as a preview, not general availability. The second is that details on timing and availability cannot be cross-checked beyond the single secondary source — the other Ignite coverage available deals with Copilot, MAI models and other topics, not the NVIDIA collaboration.
Layer two: confidential computing heading toward rack scale
The second layer targets customers with proprietary or regulated information. NVIDIA Confidential Computing with attestation is intended to let customers verify the computing environment before sensitive models, data or keys are released. Microsoft and NVIDIA are, according to the coverage, working to extend these protections to bare-metal and rack-scale AI infrastructure — that is, across interconnected "AI factory" equipment.
It is important to read this as work in development, not as something delivered. The description explicitly states the extension is under way, and there is no available documentation on when or in what form it will become available. For regulated industries, the ceiling is real, but it is a promise of direction, not a product launch.
Layer three: Nemotron 3.5 in Microsoft Foundry
The third layer makes NVIDIA's open Nemotron family more broadly available in Foundry Managed Compute. According to the coverage, the lineup includes:
- Super — for long-running agents working across coding, research and business workflows
- Lightning — for applications requiring fast responses and tool use
- Embed and Rerank — for information retrieval
- Parse — for extracting structured information from documents
- Content Safety — for detecting harmful content
Two models are listed as "coming soon": Bolt for high-throughput, low-latency reasoning and Transcribe for speech recognition. The distinction between available now and upcoming is worth holding onto — "supported model" lists in cloud platforms often blur across timelines.
The coverage also mentions internal use at Microsoft itself: Microsoft AI uses NVIDIA Dynamo for rollout inference in reinforcement learning during post-training — the compute-intensive process that generates the experiences used to improve models. Dynamo coordinates inference across a large GPU fleet and handles requests, model updates and error recovery in a deployment built on joint engineering work between Microsoft AI and NVIDIA. Microsoft also uses RTX PRO 6000 Blackwell GPUs for Bing's search models (Unite.AI).
The EY example: numbers that deserve caution
The most specific result in the announcement comes from EY. NVIDIA stated that EY used Programmed Training to adapt Nemotron open models for classifying the company's procurement records, with two results: a 55 percent accuracy improvement on EY's own evaluation dataset, and a 15-fold cost reduction compared with proprietary frontier models.
These figures are NVIDIA's own statements, relayed via Unite.AI. The methodology is not available in the sources: we do not know which models were compared, how accuracy was measured, or how costs were calculated. A 55 percent accuracy improvement on one company's own dataset and a 15x cost reduction against frontier models is, in itself, a credible profile for customized open models on a narrow, specialized assignment — but it is an example supplied by the model maker itself, not an independent benchmark. Read it as a documented example of adapting to a specialized business task, not as a general promise of such numbers.
Layer four: beyond Azure
The fourth layer moves the collaboration out of the data center:
Azure Local on Jetson Thor and Jetson Orin. Azure Local is being extended to NVIDIA's Jetson devices for industrial use cases such as machine vision, robotics and autonomous operations. In practice, that means the same software platform should be able to run both in hyperscale data centers and on edge devices in factories and facilities.
Cosmos in Foundry Local. NVIDIA's Cosmos models for physical AI are being brought to Foundry Local, so that physics-oriented models can run locally rather than in the cloud.
Direct Stream for Windows 365. For GPU-accelerated Windows 365 Cloud PCs, Direct Stream reportedly supports up to two 4K streams at 60 frames per second with low latency, even over demanding networks. A limited preview is open for sign-up, and a public preview is planned for early 2027. This is one of the few concrete timelines in the announcement — and even it concerns only a preview, not a launch.
What is verified — and what is not
What is clear: NVIDIA and Microsoft announced an expanded collaboration agreement on November 19, 2026 at Ignite, and it covers concrete hardware in Azure preview, work in progress on rack-scale confidential computing, a Nemotron 3.5 lineup in Foundry where most models are available now, and deployment on the industrial edge, locally and in Cloud PCs.
What is company-reported: the EY figures (55 percent, 15x) and Direct Stream's performance specifications come from NVIDIA's own description. Methodology and independent verification are absent from the available source material.
The open questions: When will the Vera Rubin preview become generally available, and at what price? How far will the extension of confidential computing to bare metal and rack scale go, and when? When will Bolt and Transcribe arrive in Foundry? None of these points can be answered from the available sources. For readers weighing the individual pieces, the safest approach is to treat each announcement on its own timeline: the models in Foundry apparently exist now, the hardware is in preview, rack-scale confidential computing is a goal under way, and Direct Stream is an early-2027 question — as a preview.

