AI · Nov 5, 2026
NVIDIA’s four open Cosmos 3 Nano models let robot and driving teams simulate worlds that reactAt Ignite, Nvidia and Microsoft pitch one stack from Vera Rubin racks to RTX Spark laptops
· Updated · Signal Desk newsroom
Analysis of reporting by NVIDIA Blog. How we work

Vera Rubin NVL72 is running in Azure, Nemotron 3.5 lands in Foundry and on-prem Azure Local, and Microsoft’s own teams lean on Nvidia software. It’s an enterprise lock-in story dressed as choice.
Key takeaways
- Azure is running Nvidia Vera Rubin NVL72 systems and expanding VM access, with confidential computing extended to bare-metal and rack scale.
- EY says Programmed Training in Microsoft Foundry adapted Nemotron to classify purchasing records with 55% better accuracy and 15x lower cost than proprietary frontier models.
- Nemotron 3.5 Super and Lightning are on Foundry Managed Compute; Bolt and Transcribe are coming soon.
- Azure Local gains RTX PRO 6000 Blackwell GPUs and Jetson Thor and Orin edge support; Windows 365 Cloud PCs get Direct Stream in limited preview.
- Microsoft AI uses Nvidia Dynamo for RL rollout inference, and RTX PRO 6000 GPUs serve Bing search models.
One partner, every layer
Nvidia’s Ignite post reads like a map of where it wants to sit inside Microsoft’s business: data center, enterprise model shop, factory floor, Windows desktop. The infrastructure news leads. Vera Rubin NVL72 systems are now running in Azure, with Azure VM access expanding. The two companies are extending confidential computing and attestation to bare-metal and rack-scale systems so regulated customers can check the environment before releasing models, data or keys. They’re also working on power efficiency with something Nvidia calls MaxLPS.
This follows Google Cloud’s own Vera Rubin general-availability announcement, so the hyperscalers are racing to sell the same rack. Nvidia backs Azure’s pitch with MLPerf: Azure scaled GB200 NVL72 training to 10,240 GPUs on DeepSeek-V3, and its GB300 systems led cloud submissions on DeepSeek-R1 interactive throughput by over 10%, per Signal65. Those are Blackwell results used to sell a Rubin preview, which is fair marketing but not proof of Rubin performance.
The open-model pitch is the interesting part
The more telling claim is EY’s. Using Microsoft Foundry’s serverless Programmed Training, EY adapted Nemotron open models to classify corporate purchasing records. It reports 55% better accuracy on its evaluation set and 15x lower cost than proprietary frontier models. That is the argument Nvidia keeps making: for narrow, high-volume enterprise jobs, a tuned open model beats a rented frontier model on cost and often on accuracy. We find it plausible for classification work, but it’s a single vendor-reported case on a private evaluation set.
Nemotron 3.5 Super, aimed at long-running agents, and Lightning, for fast tool use, are on Foundry Managed Compute, with Embed, Rerank, Parse and Content Safety around them. Bolt and Transcribe are listed as coming soon. Foundry Managed Compute can now deploy custom weights, closing the loop from tuning to serving. For Microsoft this hedges against depending on any single frontier lab.
Edge and desktop: the long tail
The rest pushes Nvidia hardware closer to where data lives. Nemotron 3.5 comes to Foundry Local on Azure Local, Azure Local supports RTX PRO 6000 Blackwell GPUs from partners, and it extends to Jetson Thor and Orin for robotics and machine vision, with Cosmos physical-AI models headed to Foundry Local. Windows PCs on the newly shipping RTX Spark hardware will show local LLM demos. GPU-backed Windows 365 Cloud PCs get Direct Stream, up to two 4K streams at 60 frames per second, in limited preview, with public preview planned for early 2027.
Microsoft is also a customer. Microsoft AI uses Nvidia Dynamo to run RL rollout inference during post-training, and Bing search models run on RTX PRO 6000 Blackwell GPUs. And BioNeMo and ALCHEMI toolkits join Microsoft Discovery, starting with an RFD4-Proteina protein-design integration built with the University of Washington’s Institute for Protein Design. The pattern is the same everywhere: Nvidia software as the default runtime under Microsoft’s product names.
Who it touches
- Regulated enterprises
- Confidential computing at rack scale and on-prem Azure Local options make sensitive AI workloads easier to justify.
- Enterprise AI teams
- A supported path to tune and deploy open Nemotron models inside Foundry, with EY’s case as the sales pitch.
- Frontier-model vendors
- Microsoft offering cheap tuned open models in its own catalog chips away at routine API revenue.
- Industrial and robotics firms
- Jetson support in Azure Local ties edge deployments to Microsoft’s management tools.
What to watch
- When Vera Rubin VMs on Azure move from preview to broad availability, and at what price.
- Independent reproductions of open-model tuning savings like EY’s.
- Ship dates for Nemotron 3.5 Bolt and Transcribe on Foundry.
- Windows 365 Direct Stream’s public preview, planned for early 2027.
- How much of Microsoft’s own model training runs on Nvidia’s Dynamo stack.
FAQ
Is Vera Rubin generally available on Azure?
Nvidia describes Azure’s Vera Rubin offering as a preview with expanding VM access; systems are running in Azure data centers.
What did EY achieve with Nemotron?
EY reports 55% higher accuracy on its evaluation set and 15x lower cost than proprietary frontier models for classifying purchasing records, according to Nvidia.
What’s new for on-premises customers?
Azure Local adds RTX PRO 6000 Blackwell GPU support, Jetson Thor and Orin edge platforms, and Nemotron 3.5 in Foundry Local.
More on NVIDIA
AI · Nov 16, 2026
Rubin HPC is Nvidia’s bet that science agents will be bottlenecked by simulation, not reasoningAI · Nov 9, 2026
Nvidia’s Korean startup bets span interconnect, open models, robot hands and document AIAI · Nov 8, 2026
Nvidia’s Dynamo recipes speed four Korean open models to production, tuned for B200sAI · Nov 5, 2026
Nvidia profiles Chris Alexiuk, the YouTube teacher making the case that open AI means weights, data and recipes