Microsoft and Nvidia expand confidential AI from cloud to workstation

Microsoft and Nvidia broaden their AI partnership with deployments on Azure, on-premises, and PCs. The agreement combines computing, data security, and open models for regulated enterprises.

Microsoft and Nvidia announced on November 19, 2026, an expansion of their AI collaboration, featuring Vera Rubin NVL72 systems in Azure, better-integrated Nemotron models in Foundry, and deployment options on-premises, at the industrial edge, and on Windows Cloud PC. The challenge goes beyond mere computing power: both companies also aim to make execution environments verifiable before using sensitive data, a crucial point for regulated sectors.

From cloud to factory, without losing control over data

The most notable development is the extension of confidential computing to larger-scale infrastructures. NVIDIA Confidential Computing and attestation mechanisms enable clients to verify the computing environment before transmitting models, data, or keys. In other words, trust no longer relies solely on a contract or logo but on the technical state of the machine. This effort also targets interconnected AI infrastructures, not just a single card or isolated server.

At the same time, Microsoft is expanding access to Azure virtual machines powered by Vera Rubin NVL72. These systems are designed for reasoning tasks, agents, and large-scale model training. The announced availability remains gradual, which is logical for a platform of this type: the promise is broad, and deployment will be slower than the launch communication, as is often the case with cloud giants.

Nemotron gains strength in Foundry

On the models side, Microsoft is expanding Nemotron support in Foundry, with a suite focused on customization and agent execution. Foundry Managed Compute adds support for custom model weights, while the serverless Programmed Training offering provides greater control over post-training steps. Post-training, recall, includes adjustments made after initial training to adapt a model to a specific task.

NVIDIA also cites a use case at EY. The firm reportedly used Programmed Training to adapt Nemotron models for classifying enterprise purchase data, achieving a 55% improvement in accuracy on its evaluation set and a 15-fold cost reduction compared to leading proprietary models. These figures illustrate the central commercial argument of the agreement: making customization more financially accessible, especially when the task is specialized and repetitive.

What Microsoft already does with its own GPUs

The partnership not only fuels client-facing offerings. Microsoft also uses Nvidia technologies in its own operations. Microsoft AI relies on Dynamo to perform reinforcement learning during post-training, coordinating inference across a large GPU fleet. The company also reports using RTX PRO 6000 Blackwell GPUs for Bing search models.

This type of production deployment shows that the agreement is not just a communication exercise between two already highly visible brands. It also aims to industrialize common building blocks, from workload management to disaster recovery. In the sector, demos shine quickly; pipelines, however, require a bit more coffee.

Use cases extend beyond the data center

Another key focus of the announcement concerns moving AI computing beyond the centralized cloud alone. Foundry Local supports more reasoning, coding, and agent scenarios in Azure Local environments, with retrieval models capable of searching for evidence in text, document images, tables, or charts. Microsoft also plans support for RTX PRO 6000 Blackwell GPUs provided by hardware partners for on-premises deployments.

In the industrial field, Azure Local also opens up to Jetson Thor and Jetson Orin for computer vision, robotics, and autonomous operations, while Cosmos physical models arrive in Foundry Local. The goal is clear: bring AI closer to data generated by machines, rather than sending everything back to a distant data center. On the workstation side, both companies also want to demonstrate local large model experiences on Windows PCs equipped with RTX Spark, with a limited preview already open and a public preview planned for early 2027.

Key points

  • Announcement made on November 19, 2026, at Microsoft Ignite.
  • EY reports 55% more accuracy on its tests.
  • Costs reportedly dropped by a factor of 15 according to Nvidia.
  • Microsoft tests Direct Stream on Windows 365 Cloud PCs.
  • Jetson Thor and Jetson Orin target industrial AI.
  • Nvidia and Microsoft extend confidential AI to rack-scale.

By the numbers

  • November 19, 2026 — date of the announcement at Microsoft Ignite.
  • 55% — accuracy gain reported by EY on its evaluation set.
  • 15x — cost reduction announced compared to leading proprietary models.
  • 2 streams of 4K at 60 fps — performance announced for Direct Stream on Windows 365 Cloud PCs.
  • early 2027 — expected window for the public preview of Windows 365 Cloud PCs.

To read