Grafy News
Try Grafy free

NVIDIA and Microsoft Extend Accelerated Computing from Azure Cloud to Windows PCs

NVIDIA and Microsoft announced new Azure VM options, edge AI support and Windows PC models.

Illustration for the news story: NVIDIA and Microsoft Extend Accelerated Computing from Azure Cloud to Windows PCsSource: NVIDIA Press Room

NVIDIA and Microsoft revealed a suite of new accelerated‑computing services that move powerful AI workloads from Azure cloud servers to on‑premise Windows PCs and edge devices. The expansion matters because it gives enterprises the flexibility to run large language models and inference workloads closer to their data, reducing latency and addressing regulatory constraints.

What new Azure virtual machines are now available?

At Microsoft Ignite the partners announced that Azure virtual machines will be powered by NVIDIA Vera Rubin NVL72 systems. These GPUs are built on the Blackwell architecture and are designed for demanding reasoning, agentic applications, and large‑scale model training and inference. In the latest MLPerf Training v6.1 benchmark, Azure scaled to 10,240 NVIDIA GPUs using the GB200 NVL72 configuration on the DeepSeek‑V3 model, marking the largest scale run in that round. For inference, the GB300 NVL72 systems led cloud submissions on the DeepSeek‑R1 interactive throughput benchmark, delivering more than a 10 % advantage over competing submissions, according to the NVIDIA blog.

How are NVIDIA and Microsoft strengthening confidential computing?

The collaboration also extends confidential‑computing technology to bare‑metal and rack‑scale AI infrastructure. NVIDIA Confidential Computing and attestation let customers verify the integrity of the compute environment before releasing sensitive models, data, or encryption keys. This protects proprietary or regulated information while it is being processed, a crucial capability for sectors such as finance, healthcare and defense. The partners are working to propagate these protections across the entire AI factory, ensuring that as workloads span more interconnected accelerators the same security guarantees apply.

What customization capabilities does Microsoft Foundry now offer for open models?

Microsoft Foundry’s new services let enterprises customize and deploy open‑source models without managing the underlying hardware. The Programmed Training offering gives developers control over post‑training workflows, while Foundry Managed Compute adds support for deploying custom model weights. EY used Programmed Training to adapt NVIDIA’s Nemotron open models for classifying corporate purchasing records, achieving a 55 % accuracy lift and cutting costs by a factor of 15× compared with proprietary frontier models. The Nemotron 3.5 family now includes several variants – Super for long‑running agents, Lightning for fast‑response tool‑use, Bolt for high‑throughput low‑latency reasoning, and upcoming Transcribe for speech recognition. Complementary models such as Nemotron Embed, Rerank, Parse and Content Safety help agents retrieve information, extract structured data, and filter harmful content.

How is AI being brought closer to enterprise data on‑premise and at the edge?

Beyond the cloud, the announcement adds Foundry Local support for Nemotron 3.5, enabling reasoning, coding and tool‑using agents to run in Azure Local environments. Retrieval models can now search across text, document images, charts and tables stored on‑premise. Azure Local will also accept NVIDIA RTX PRO 6000 Blackwell GPUs supplied by hardware partners, expanding on‑premise deployment options. For edge scenarios, support is being added for NVIDIA Jetson Thor and Jetson Orin platforms, targeting industrial use cases such as machine vision, robotics and autonomous operations. Physical AI models like Cosmos are also being made available on Foundry Local. On Windows PCs equipped with NVIDIA RTX Spark GPUs, the companies demonstrated local large‑language‑model experiences, showing that developers can run sophisticated AI applications directly on a developer’s workstation.

What does the new Direct Stream capability mean for Cloud PCs?

Microsoft introduced Direct Stream for GPU‑accelerated Windows 365 Cloud PCs. The feature supports demanding graphics workloads with up to two simultaneous 4K streams at 60 fps and low latency, even over challenging networks. A limited preview is open for sign‑ups, with a public preview slated for early 2027. This capability is particularly relevant for designers, engineers and creators who need high‑resolution, real‑time rendering on remote desktops.

How will developers and enterprises benefit from these announcements?

The expanded hardware and software stack gives developers a broader palette of compute options—from massive Azure GPU clusters to edge‑optimized Jetson devices and desktop‑class RTX Spark cards. Enterprises can keep sensitive data on‑premise while still leveraging the latest model families, thanks to confidential‑computing extensions. The ability to fine‑tune open models with Foundry’s Programmed Training reduces reliance on costly proprietary APIs and opens the door to domain‑specific AI solutions. In practice, a financial services firm could run a Nemotron‑based compliance assistant on a secure on‑premise server, while a media company could prototype visual effects on a local RTX Spark workstation and then scale the same workload to Azure’s Vera Rubin fleet for full‑length production.

How creators can explore these developments with Grafy

Grafy users can experiment with the concepts announced by Microsoft and NVIDIA without needing access to the hardware itself. The Grafy Chat app lets you converse with multiple AI models side‑by‑side, mirroring the multi‑model environment described in Foundry Managed Compute. You can also use Grafy Reader to ingest and query PDFs of the NVIDIA blog post or related research papers, helping you understand the technical details of Vera Rubin performance or confidential‑computing attestation. For developers interested in code, the Grafy Code studio provides an AI‑assisted coding environment where you can prototype scripts that interact with Azure APIs or simulate model fine‑tuning workflows. These tools give creators a low‑barrier way to explore the same ideas—model customization, edge inference and secure compute—before deploying them on the full NVIDIA‑Microsoft stack.

Frequently asked questions

What is the Vera Rubin NVL72 system and why is it important?

The Vera Rubin NVL72 is a Blackwell‑based GPU system that powers new Azure virtual machines. It delivers high‑throughput training and inference performance, demonstrated by scaling to 10,240 GPUs in MLPerf benchmarks and achieving a 10 % lead in interactive throughput, making it a cornerstone for large‑scale AI workloads.

How does Microsoft Foundry help businesses customize open‑source models?

Foundry offers Programmed Training for post‑training workflow control and Managed Compute for deploying custom model weights. This lets enterprises fine‑tune models like Nemotron 3.5 on their own data, reducing reliance on expensive proprietary services while improving task‑specific accuracy.

Can AI models run locally on Windows PCs, and what hardware is required?

Yes. NVIDIA RTX Spark GPUs enable local large‑language‑model experiences on Windows PCs. Developers can run sophisticated AI applications directly on the device, leveraging the same model families that are available in Azure, which opens up low‑latency, on‑device AI use cases.


Source: NVIDIA Press Room — NVIDIA and Microsoft Expand Accelerated Computing From Azure to Windows PCs