NVIDIA, Microsoft Extend Full-Stack AI Collaboration From Azure to Windows PCs

NVIDIA and Microsoft are pushing accelerated computing from Azure data centers onto Windows PCs and factory floors, with Vera Rubin NVL72 systems now running in Azure and new local AI deployment options announced at Microsoft Ignite this week. The expansion, detailed by NVIDIA’s newsroom, puts enterprise AI agents, on-premises inference, and even RTX-powered local LLMs under one full-stack umbrella.

The announcement, made at Microsoft Ignite, spans three tiers: hyperscale cloud infrastructure, mid-market enterprise customization, and the edge. Microsoft is also eating its own cooking — running NVIDIA Dynamo inference software for post-training its own models and deploying RTX PRO 6000 Blackwell GPUs to accelerate Bing search. That last detail matters. When Microsoft accelerates its own flagship services on NVIDIA silicon, it signals the partnership is infrastructural, not just a resale arrangement.

Vera Rubin lands in Azure before most enterprises have finished with Blackwell

Vera Rubin NVL72 systems are already running in Azure data centers, according to NVIDIA, aimed at demanding reasoning workloads, agentic applications, and large-scale training and inference. The preview builds on codesign work the two companies have been doing for years, and the benchmark numbers back the claim up.

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. On the inference side, Azure’s GB300 NVL72 systems led cloud submissions on DeepSeek-R1 interactive throughput by over 10%, per Signal65. Those are audited, comparable numbers, not vendor cherry-picking, which is precisely why the MLPerf results carry weight with infrastructure buyers.

There’s a security angle too. Microsoft and NVIDIA are extending confidential computing to bare-metal and rack-scale AI infrastructure — hardware-isolated execution environments paired with cryptographic attestation, so customers can verify the computing environment before releasing sensitive models, data, or keys. As AI factories sprawl across interconnected accelerators, extending those guarantees rack-wide is the hard part, and it’s where regulated industries have been holding back.

Nvidia and microsoft expand accelerated computing from azure to windows pcs
Photo: wpnews.pro

Power is the other constraint. The companies are collaborating on efficiency using NVIDIA MaxLPS, because in 2026 the binding constraint on AI capacity is increasingly megawatts, not GPUs.

The Nemotron bet: open models, 15x cheaper than frontier

The enterprise software tier is where the economics get interesting. NVIDIA’s open Nemotron 3.5 lineup is expanding on Microsoft Foundry, with Super built for long-running agents across coding, research, and business workflows, and Lightning tuned for fast responses and tool use. Complementary models cover retrieval (Embed and Rerank), document extraction (Parse), and content safety, with Bolt for high-throughput, low-latency reasoning and Transcribe for speech coming soon.

The proof point comes from EY, which used Microsoft’s serverless Programmed Training in Foundry to adapt Nemotron open models for classifying corporate purchasing records. Per NVIDIA, accuracy on the evaluation dataset improved 55% while cost dropped 15x versus proprietary frontier models. That’s the argument in a single data point: for narrow, specialized tasks, a small open model you fine-tune beats a frontier model you rent.

Foundry Managed Compute now supports deploying custom model weights, closing the loop between customization and production deployment inside Microsoft’s managed services.

From the data center to the factory floor — and your desk

The edge tier is the genuinely new territory. Azure Local gains support for Nemotron 3.5, bringing reasoning, coding, and tool-using agents to on-premises environments, and will support RTX PRO 6000 Blackwell GPUs from hardware partners.

Azure Local is also expanding to NVIDIA’s Jetson Thor and Jetson Orin platforms for industrial work — machine vision, robotics, autonomous operations — with NVIDIA’s Cosmos physical AI models heading to Foundry Local. Cyberoly’s coverage of the announcement frames this as the practical payoff: businesses running AI where their data actually lives, without shipping everything to a hyperscaler.

On the consumer-adjacent side, Windows PCs powered by NVIDIA RTX Spark will showcase local LLM experiences at Ignite — AI running directly on the device, no round trip to a data center.

And for organizations using Cloud PCs, Direct Stream capabilities are coming to GPU-accelerated Windows 365, supporting two 4K streams at 60 frames per second with low latency over imperfect networks. A limited preview is open now, with public preview planned for early 2027.

Science gets a slice too. The BioNeMo Agent Toolkit and ALCHEMI Toolkit are entering Microsoft Discovery’s catalog, with an initial integration using the RFD4-Proteina model — developed with the University of Washington’s Institute for Protein Design — to generate candidate protein structures for lab validation.

What this changes for who

For infrastructure buyers, the Vera Rubin preview plus rack-scale confidential computing removes two of the biggest blockers to putting regulated workloads on Azure. For developers, the Nemotron-plus-Foundry stack offers a credible open-model alternative to closed frontier APIs — with EY’s 15x cost reduction as the exhibit. For industrial operators, Jetson integration into Azure Local means edge AI without a bespoke stack.

It also tightens the platform question. The tighter NVIDIA’s silicon binds to Microsoft’s cloud and OS layer, the harder it becomes for rivals — AMD’s Instinct accelerators, Google’s TPUs on GCP, AWS’s Trainium — to make their case inside enterprises already standardized on this stack. Open models running on locked-in infrastructure is an increasingly common compromise.

NVIDIA, founded in 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem, began as a gaming graphics company — one of roughly 70 startups chasing GPU acceleration in the late 1990s, of which only NVIDIA and ATI survived. The company’s early ideas were sketched out in a meeting at a Denny’s diner in East San Jose, and it went public in January 1999, releasing the GeForce 256, its first product marketed as a GPU, that same year.

From those gaming roots, NVIDIA broadened into AI, professional visualization, and supercomputing; in 2025 it became the first company to surpass US$4 trillion and then US$5 trillion in market capitalization, driven by demand for its AI data center hardware.

The open question

What the announcements don’t resolve: real-world pricing for Vera Rubin VMs, and whether the local PC AI experiences graduate from Ignite demos into shipping products. NVIDIA’s booth at Ignite, featuring a GB300 NVL72 rack and RTX Spark devices running demos from enterprise applications to robotics, is where those answers start — with Direct Stream capabilities coming to GPU-accelerated Windows 365 Cloud PCs.

Photo of author

Technology Desk

Technology Desk at Archyde.

Baillie Gifford: Anthropic IPO One Step in Long-Term AI Investment Story

Newcastle Target Spurs-Sold Striker Lankshear as Tottenham Hold Matching Rights and Sell-On Clause