Nvidia and Microsoft are pushing businesses to run cheaper open AI models on their own terms, from Azure data centers to factory machines and Windows PCs, and Nvidia says accounting giant EY cut its costs 15x by making the switch.
Key Takeaways
Nvidia says EY tuned its Nemotron open models to raise accuracy 55% and cut costs 15x compared with proprietary frontier models
Azure virtual machines powered by Nvidia Vera Rubin NVL72 systems are running in Azure data centers
Microsoft Foundry is adding customer availability for Nvidia open models and customization capabilities
Azure Local will support RTX PRO 6000 Blackwell GPUs and expand to Jetson Thor and Jetson Orin
GPU-accelerated Windows 365 Cloud PCs are getting Direct Stream in limited preview, with public preview planned for early 2027
Nvidia announced the expansion at Microsoft Ignite, where the companies outlined new access points spanning Azure virtual machines, local infrastructure, industrial systems and Windows PCs.
Microsoft is also using Nvidia technologies in its own AI work, including Dynamo inference software for post-training and RTX PRO 6000 Blackwell GPUs for Bing search models.
From Azure To Local Systems
Azure virtual machines powered by Nvidia Vera Rubin NVL72 systems are now running in Azure data centers. The infrastructure is designed for demanding reasoning, agentic applications, and large-scale model training and inference.
The companies are also collaborating to extend confidential computing to bare-metal and rack-scale AI infrastructure. Nvidia Confidential Computing and attestation help customers verify the computing environment before releasing sensitive models, data or keys, a safeguard aimed at organizations handling proprietary or regulated information.
Nvidia and Microsoft are also working on power efficiency using Nvidia MaxLPS. In MLPerf Training v6.1, Azure scaled to 10,240 Nvidia GPUs with GB200 NVL72 systems on the DeepSeek-V3 model, which Nvidia described as the largest scale in the round. In MLPerf Inference v6.1, Azure’s Nvidia GB300 NVL72 systems led cloud submissions on DeepSeek-R1 interactive throughput by over 10%, per Signal65.
Also Read: Nvidia GeForce Mod Promises Up To 2X AI Speed On Home PCs
How Foundry Adds Open Models
Through Microsoft Foundry, customer availability is expanding for Nvidia open models and customization capabilities. Microsoft’s serverless Programmed Training gives developers control over post-training workflows while Foundry manages the underlying infrastructure, and Foundry Managed Compute adds support for deploying custom model weights.
EY used Programmed Training in Microsoft Foundry to adapt Nvidia Nemotron open models for classifying corporate purchasing records. EY improved accuracy on its evaluation dataset by 55% and lowered cost by 15x compared with proprietary frontier models.
Nemotron ModelStated Enterprise UseNemotron 3.5 SuperDesigned for long-running agents working across coding, research and business workflowsLightningBuilt for applications that need fast responses and tool useEmbed And RerankHelp agents find relevant informationParseExtracts structured information from documentsContent SafetyDetects harmful contentBolt And TranscribeComing soon for high-throughput low-latency reasoning and speech recognition
Expanded support for Nemotron 3.5 in Foundry Local brings reasoning, coding and tool-using agents to Azure Local environments. Retrieval models help applications find evidence across text, document images, charts and tables, and Nvidia physical AI models including Cosmos are also being brought to Foundry Local.
Why This Matters
The EY result is the part to watch. Open models can be downloaded, trained on a company’s own records and run on its own hardware, instead of paying per request for a closed model from a single provider. If a tuned open model can beat a frontier model on a narrow job at a fifteenth of the cost, many businesses will ask why they are paying top prices for everyday tasks like sorting purchase records.
That puts Microsoft in an interesting spot. It sells access to the biggest closed models through Azure, and it is now also helping customers swap in cheaper Nvidia models for specific work. Nvidia wins either way, because both paths run on its chips, from the Vera Rubin NVL72 racks in Azure to RTX GPUs in office PCs.
What To Watch Next
Azure Local will also support Nvidia RTX PRO 6000 Blackwell GPUs from hardware partners, giving organizations additional options for deploying AI on premises. It is also expanding to Nvidia Jetson Thor and Jetson Orin edge AI platforms for industrial applications such as machine vision, robotics and autonomous operations.
On Windows PCs powered by Nvidia RTX Spark, the companies will showcase local large language model experiences at Ignite. For organizations using Windows 365 Cloud PCs, Direct Stream supports demanding graphics applications with up to two 4K streams at 60 frames per second and low latency, even over challenging networks. The limited preview is open for sign-ups, with public preview planned for early 2027.
The bigger test is adoption. Watch for more companies to publish cost and accuracy numbers like EY’s, and for whether Bolt and Transcribe, the two Nemotron 3.5 models listed as coming soon, arrive in Foundry in time to push more work off pricey closed models.
For scientific research, Nvidia BioNemo Agent Toolkit and Nvidia ALCHEMI Toolkit are expanding into the Microsoft Discovery catalog. An initial integration uses the RFD4-Proteina model, developed with the University of Washington’s Institute for Protein Design, to generate candidate protein structures and amino acid sequences for further evaluation and laboratory validation.
Read Next: Cohere Hides Patient And Bank Data From Outsiders On Azure