NVIDIA and Microsoft say Vera Rubin NVL72 systems are now running in Azure data centers, and that Nemotron 3.5 open models are available through Microsoft Foundry. The claims come from a NVIDIA blog post published November 19, 2026, during Microsoft Ignite. The post spans cloud hardware, model customization, edge devices and Windows PCs.
This article sorts the announcements by status: what is available, what is preview, and what is "coming soon." It also flags the one customer number that has no published method behind it. Our full Ignite 2026 coverage tracks the Microsoft side of the week.
TL;DR: what NVIDIA and Microsoft announced
| Item | What was said | Status per the sources |
|---|---|---|
| Vera Rubin NVL72 on Azure | Systems running in Azure data centers; expanding access to VMs | Preview, per NVIDIA |
| Confidential computing | Extending to bare-metal and rack-scale AI infrastructure | Collaboration, no date |
| Power efficiency | NVIDIA MaxLPS work with Microsoft | Collaboration, no date |
| Nemotron 3.5 on Foundry Managed Compute | Super, Lightning, plus Embed, Rerank, Parse, Content Safety | Available, per NVIDIA |
| Nemotron 3.5 Bolt and Transcribe | High-throughput reasoning; speech recognition | Coming soon |
| Programmed Training | Serverless post-training in Foundry | Offered, per NVIDIA |
| Foundry Managed Compute | Adds custom model weight deployment | Added, per NVIDIA |
| Azure Local | Adds RTX PRO 6000 Blackwell, Jetson Thor and Orin | Expanding |
| Windows 365 Direct Stream | Up to two 4K streams at 60 fps | Limited preview; public preview early 2027 |
| Microsoft Discovery | BioNeMo Agent Toolkit and ALCHEMI Toolkit join the catalog | Expanding |
What is new with Vera Rubin NVL72 on Azure?
NVIDIA says Vera Rubin NVL72 systems are running in Azure data centers, built for reasoning, agentic applications and large-scale training and inference. It also says the companies are expanding access to Azure virtual machines powered by those systems.
Microsoft's own pre-Ignite infrastructure post gives the history. It says Microsoft became the first cloud to bring up a Vera Rubin NVL72 system for validation in March, and that it delivered its first production Vera Rubin systems in August. That Microsoft post does not describe VM availability or a preview, so the VM detail rests on NVIDIA's blog. If you plan to rent capacity, check the Azure portal and your quota, not the press releases.
NVIDIA ties the preview to benchmark results from Blackwell systems. In MLPerf Training v6.1, it says Azure scaled to 10,240 GPUs with GB200 NVL72 systems on DeepSeek-V3, the largest scale in the round. In MLPerf Inference v6.1, it says Azure's GB300 NVL72 systems led cloud submissions on DeepSeek-R1 interactive throughput by more than 10%, per Signal65. These are Blackwell results, not Rubin results.
For the chip itself, see our posts on Rubin HPC at SC26 and the vLLM Rubin throughput claim.
Confidential computing and power
Two infrastructure items have no dates. NVIDIA and Microsoft are working to extend confidential computing to bare-metal and rack-scale AI infrastructure. The aim is to let customers verify the environment through attestation before releasing sensitive models, data or keys, across more interconnected accelerators. They are also working on power efficiency using NVIDIA MaxLPS, to extract more performance from the power a data center has.
Which Nemotron models are on Foundry, and what is Programmed Training?
NVIDIA lists the Nemotron 3.5 lineup on Foundry Managed Compute:
- Nemotron 3.5 Super: long-running agents across coding, research and business workflows.
- Nemotron 3.5 Lightning: fast responses and tool use.
- Nemotron Embed and Rerank: find relevant information for agents.
- Nemotron Parse: extract structured data from documents.
- Nemotron Content Safety: detect harmful content.
- Nemotron 3.5 Bolt and Transcribe: coming soon.
Microsoft's serverless Programmed Training gives developers control of post-training while Foundry manages the infrastructure. Foundry Managed Compute now supports deploying custom model weights, so a tuned model can go from training to managed deployment on one platform.
For background on the open-model push around Nemotron, see Jensen Huang's exit from NVIDIA and the open-source Nemotron effort and the Interconnects open model coalition piece.
The EY result: what is claimed and what is not shown
NVIDIA says EY used Programmed Training to adapt Nemotron open models for classifying corporate purchasing records. It reports 55% better accuracy on EY's evaluation dataset and 15x lower cost than proprietary frontier models.
The post does not say which frontier models were the baseline, which Nemotron model was tuned, how large the evaluation set was, or how cost was measured (per token, per record or total). "55% better" could be relative or in percentage points. The result is plausible, since narrow classification tasks are where small tuned models tend to win, but it is a vendor-published customer figure. If you plan to repeat it, build your own evaluation set first, and compare against the cheapest frontier model that meets your accuracy bar.
Similar small-model-beats-frontier claims have appeared elsewhere this month. See Genspark's 17x cheaper claim with Fireworks.
Edge and Windows: Azure Local, Jetson, RTX Spark
NVIDIA and Microsoft are widening where models can run.
- Azure Local will support NVIDIA RTX PRO 6000 Blackwell GPUs from hardware partners, and expands to Jetson Thor and Jetson Orin for machine vision, robotics and autonomous operations.
- Foundry Local gets expanded Nemotron 3.5 support, bringing reasoning, coding and tool-using agents to Azure Local. Retrieval models find evidence across text, document images, charts and tables. NVIDIA's Cosmos physical AI models are also coming to Foundry Local.
- Windows PCs with RTX Spark: the companies will demonstrate local LLM applications at Ignite. Our post on RTX Spark compact desktops shipping covers the hardware.
- Windows 365 Direct Stream brings up to two 4K streams at 60 frames per second to GPU-accelerated Cloud PCs. A limited preview is open for sign-ups, and public preview is planned for early 2027.
Science and Microsoft's own models
NVIDIA's BioNeMo Agent Toolkit and ALCHEMI Toolkit are joining the Microsoft Discovery catalog, covering life sciences, drug discovery and materials science. The first integration uses the RFD4-Proteina model, built with the University of Washington's Institute for Protein Design, to propose candidate protein structures and sequences "for further evaluation and laboratory validation," NVIDIA says. That wording matters: these are candidates, not validated proteins.
On the Microsoft side, NVIDIA says Microsoft AI uses Dynamo for reinforcement learning rollout inference in post-training, and RTX PRO 6000 Blackwell GPUs for Bing search models. We explain how Dynamo works in RL in our NVIDIA Dynamo RL explainer. Microsoft's own model news from the week is in MAI-Code-2 and seven models at Ignite.
Why does this partnership post matter?
Partnership posts like this one are easy to dismiss as co-branded marketing. Three details make this one worth reading closely.
First, it describes a full stack in one place. Hardware (Vera Rubin NVL72 racks), models (Nemotron 3.5), tooling (Programmed Training, custom weights on Managed Compute), edge devices (Jetson, RTX PRO 6000) and clients (RTX Spark PCs, Cloud PCs) all appear together. For a buyer, that means one vendor pair can cover a workload from training to deployment. For a competitor, it shows how much of the pipeline NVIDIA wants to own.
Second, it shows where NVIDIA's open models fit. Nemotron appears as the default open choice inside Microsoft's own model catalog. That gives NVIDIA distribution beyond its own channels, and gives Microsoft customers an alternative to proprietary frontier models for narrow jobs.
Third, it exposes the gap between announcement and availability. Within one post you find "running," "expanding," "added," "limited preview," "coming soon" and "planned for early 2027." Readers who plan budgets should read each verb literally.
Questions people ask
Can I rent Vera Rubin on Azure today? NVIDIA calls it a preview and says access is expanding. No pricing, region or VM size was published in the sources we checked. Ask Azure directly.
Is Nemotron 3.5 free? The post does not state licensing or pricing for Foundry Managed Compute. NVIDIA calls the models open, but Foundry hosting costs follow Azure pricing. Read each model card before you build on it. Our guide to open-weight model licenses explains what to look for.
Does this replace GPT-style models in Foundry? No. The post adds options. It does not say any proprietary model is removed.
What should builders and buyers do?
- Check Rubin access separately. NVIDIA says VMs are expanding; Microsoft's post does not confirm them. Ask your Azure rep about regions, quota and pricing, none of which were given.
- Try a Nemotron 3.5 model on a narrow task. If you have a classification or extraction job, the EY example is a template, not proof. Run your own side-by-side.
- Treat "coming soon" as not shipped. Bolt, Transcribe and public Direct Stream preview have no firm dates, except early 2027 for Direct Stream.
- If you have regulated data, watch the confidential computing work. It is a collaboration with no date.
What is confirmed and what is not
From NVIDIA's own post: Vera Rubin NVL72 systems run in Azure; Nemotron 3.5 is on Foundry Managed Compute; Programmed Training and custom weights are available; Direct Stream public preview is planned for early 2027.
From Microsoft's own post: first Vera Rubin bring-up in March and first production systems in August.
Not published: Rubin VM pricing, regions and sizes; the EY method; any independent benchmark of the Nemotron 3.5 models on Foundry.
Related reading
- Microsoft Ignite 2026 complete coverage
- NVIDIA Dynamo for RL rollouts
- NVIDIA Rubin HPC at SC26
- vLLM Rubin 93x throughput claim
- NVIDIA RTX Spark compact desktops
- MAI-Code-2 and seven models at Ignite
- Nemotron and the open model coalition
Sources: NVIDIA blog, Nov 19, 2026; Microsoft Azure Infrastructure blog, Join NVIDIA and Microsoft at Ignite 2026.
Details are accurate as of November 19, 2026. Preview and availability can change.
