Chips & Compute

Microsoft expands Vera Rubin and Nemotron access beyond Azure

Microsoft and NVIDIA expanded their AI partnership on November 19, spanning Vera Rubin NVL72 Azure previews, Nemotron 3.5 tools and local deployments on RTX PRO systems.

NVIDIA Expands Microsoft AI Collaboration with Vera Rubin and Nemotron
NVIDIA Expands Microsoft AI Collaboration with Vera Rubin and NemotronAI-generated
By Sophie Lindqvist6 min read

Why it matters

  • November 19, 2026: Microsoft and NVIDIA announced expanded Vera Rubin, Nemotron and confidential-computing support.
  • NVIDIA reported a 55% accuracy improvement and a 15x cost reduction for EY’s Programmed Training case.
  • Azure’s Vera Rubin preview targets reasoning, agentic applications, model training and inference.
  • Direct Stream supports up to two 4K streams at 60 fps, with public preview planned for early 2027.
  • Nemotron 3.5 lists Super, Lightning, Embed, Rerank, Parse, Content Safety, Bolt and Transcribe.

By Sophie Lindqvist

Microsoft and NVIDIA announced on November 19, 2026, at Microsoft Ignite, that Azure virtual machines powered by NVIDIA’s Vera Rubin NVL72 would become available through the Azure Vera Rubin preview.

The same announcement broadened Nemotron 3.5 support in Microsoft Foundry and added deployment paths to on-premises systems, industrial edge devices and Windows Cloud PCs. For regulated customers, the stakes are practical: they need to verify computing environments before releasing sensitive models, data or keys, while enterprises can choose cloud, edge or local hardware for different workloads.

What changes in Azure?

Microsoft is expanding access to Azure virtual machines powered by Vera Rubin NVL72 through the Azure Vera Rubin preview. The systems run in Azure data centers and target reasoning and agentic applications, as well as large-scale model training and inference, Microsoft and NVIDIA said.

The announcement positions the preview as a bridge between model development and deployment. It ties the hardware to a model toolchain instead of treating Vera Rubin as a stand-alone compute upgrade. Foundry Managed Compute supports custom model weights as part of that path.

The Azure phase starts with new hardware access and a wider model menu.

Can customers verify confidential AI workloads?

Microsoft and NVIDIA are extending confidential computing to bare-metal and rack-scale AI infrastructure used with proprietary and regulated information. NVIDIA Confidential Computing and attestation help customers verify the computing environment before they release sensitive models, data or keys, the companies said.

Work is under way to extend those protections across interconnected AI factory infrastructure. The companies are also collaborating on power efficiency through NVIDIA MaxLPS, which NVIDIA said can help data centers obtain “more performance from available power.”

For enterprise buyers, verification happens before a sensitive asset enters the workflow. Customers can check whether the computing environment meets their requirements before releasing models, data or keys.

Customers can check the environment before handing over their most sensitive AI assets.

What is Microsoft already running?

Microsoft AI uses Dynamo for reinforcement-learning rollout inference during post-training, Microsoft said. The company described post-training as the computationally intensive process that generates experiences used to improve models.

Dynamo coordinates inference across a large GPU fleet, optimizing performance and managing requests, model updates and recovery from failures. The deployment combines joint Microsoft AI and NVIDIA engineering, while Microsoft also uses RTX PRO 6000 Blackwell GPUs for Bing search models.

Microsoft is both a customer of the infrastructure and a channel for it. The orchestration layer runs inside Microsoft’s own AI development work, alongside hardware used by Bing.

What did EY achieve with Nemotron?

Microsoft’s serverless Programmed Training gives enterprise developers control over post-training workflows while Foundry manages the infrastructure. Microsoft said Foundry Managed Compute also supports deploying custom model weights, connecting model adaptation with deployment through its managed services.

NVIDIA reported that EY used Programmed Training to adapt Nemotron open models for classifying corporate purchasing records. NVIDIA reported a “55% accuracy improvement on EY’s evaluation dataset and a 15x cost reduction compared with proprietary frontier models.”

The two figures measure different things: task accuracy on EY’s purchasing records and the cost of reaching that result. NVIDIA presented them as results from a specialized business task using Nemotron open models.

The case gives enterprises a concrete cost-and-accuracy test for adapting open models.

Which Nemotron 3.5 models are listed?

NVIDIA listed the expanded Nemotron 3.5 lineup on Foundry Managed Compute:

  • Super: Supports long-running agents across coding, research and business workflows.
  • Lightning: Supports applications that need fast responses and tool use.
  • Embed and Rerank: Support information retrieval.
  • Parse: Extracts structured information from documents.
  • Content Safety: Detects harmful content.

NVIDIA marked Bolt for high-throughput, low-latency reasoning and Transcribe for speech recognition as coming soon. The division is practical: Super covers long-running work, while Lightning covers rapid responses and tool use. The named retrieval and safety models handle narrower supporting jobs.

The lineup now separates agent work, retrieval, document parsing and safety into named services.

Where can local AI run?

Expanded Nemotron support in Foundry Local brings reasoning, coding and tool-using agents to Azure Local environments, NVIDIA said. The retrieval models help applications find evidence across text, document images, charts and tables.

NVIDIA said Azure Local will support RTX PRO 6000 Blackwell GPUs from hardware partners for on-premises AI deployments. Azure Local is also expanding to Jetson Thor and Jetson Orin for machine vision, robotics and autonomous operations.

Cosmos physical AI models are being brought to Foundry Local. The local catalog now spans Windows PCs, on-premises systems, Jetson edge devices and Windows 365 Cloud PCs.

Foundry Local is the connective layer in this part of the rollout. It brings Nemotron agents, retrieval and Cosmos physical AI models into environments that can run on local hardware rather than only in centralized Azure regions.

The collaboration is moving AI closer to industrial data while keeping the same model and agent tools in view.

What can Windows Cloud PCs do?

NVIDIA said the companies will demonstrate local large-language-model experiences on RTX Spark-powered Windows PCs at Ignite. For GPU-accelerated Windows 365 Cloud PCs, Direct Stream supports up to two 4K streams at 60 frames per second with low latency over challenging networks, Microsoft said.

A limited preview is open for sign-ups, and Microsoft plans a public preview for early 2027. Cloud PC users get a bounded preview now and a defined route to broader access.

What changes for scientific research?

NVIDIA’s BioNeumo Agent Toolkit and ALCHEMI Toolkit are expanding into the Microsoft Discovery catalog, adding models for life sciences, drug discovery, materials science and chemical science. The announcement places them in Microsoft’s agentic research platform.

BioNeumo lets researchers use natural language to run computational-biology workflows, with agents coordinating the required tools and model calls. The mechanism matters for research groups because agents can manage the sequence of tool and model calls needed to complete a workflow.

The addition puts scientific models alongside Microsoft’s agentic research tools, where natural-language requests can trigger specialized workflows. Scientific users gain a workflow layer, not just additional standalone models.

When will the rollout become broadly available?

The rollout remains staged, according to the announcement. Vera Rubin systems already run in Azure, Azure virtual machines are available through preview, Direct Stream has a limited sign-up preview, and additional models remain on the roadmap.

Microsoft runs NVIDIA technology inside its own model development and search operations while offering related infrastructure and model services to enterprises. The staged schedule matters because it separates products already running from features still waiting on a preview or a future release.

The next test is whether the preview, local tools and confidential-computing extensions become generally available services rather than staged demonstrations.

Source: unite.ai

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

823 articles

Related articles

  1. Baseten Lands Four Open Models in Microsoft Foundry as First-Party Preview
  2. Dell and Microsoft bring Foundry Local AI into the private cloud
  3. Microsoft's Licensed Weights Lets Firms Custom-Train Its AI Models
  4. Nscale activates first NVIDIA Vera Rubin NVL72 AI racks
  5. NVIDIA Ships Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model

« Previous articleNext article »