· via dev.to (home feed)
Nvidia and Microsoft push open models across Azure, edge and Windows as EY cuts AI costs 15x
Nvidia says EY raised accuracy 55% and cut costs 15x by tuning Nemotron open models, as the companies expand open AI options from Azure data centers to edge devices and Windows PCs.

What was announced
At Microsoft Ignite, Nvidia and Microsoft laid out new ways for businesses to run Nvidia's open AI models, spanning Azure data centers, on-premises hardware, industrial equipment and Windows PCs. According to Fathom News, the announcement was anchored by a case study: accounting firm EY adapted Nvidia's Nemotron models for classifying corporate purchasing records and reported a 55% accuracy improvement on its evaluation dataset, along with a 15x cost reduction compared with proprietary frontier models.
The EY result
EY carried out the tuning with Programmed Training, a serverless capability in Microsoft Foundry that gives developers control over post-training workflows while Foundry manages the underlying infrastructure. Microsoft Foundry is simultaneously expanding customer availability for Nvidia's open models and adding Managed Compute support for deploying custom model weights.
The Nemotron family covers a spread of enterprise tasks, per Fathom News. Nemotron 3.5 Super is aimed at long-running agents handling coding, research and business workflows; Lightning is built for fast responses and tool use; Embed and Rerank help agents find relevant information; Parse extracts structured data from documents; and Content Safety flags harmful material. Two further models, Bolt and Transcribe, are listed as coming soon for high-throughput, low-latency reasoning and speech recognition.
New Azure infrastructure
Azure virtual machines powered by Nvidia's Vera Rubin NVL72 systems are now running in Azure data centers, positioned for demanding reasoning, agentic applications and large-scale training and inference. The companies are also extending confidential computing to bare-metal and rack-scale AI infrastructure: Nvidia Confidential Computing and attestation let customers verify the computing environment before releasing sensitive models, data or keys, a safeguard aimed at organizations with regulated information.
On performance, Nvidia said Azure scaled to 10,240 Nvidia GPUs with GB200 NVL72 systems on the DeepSeek-V3 model in MLPerf Training v6.1, which it described as the largest scale in that round. In MLPerf Inference v6.1, Azure's GB300 NVL72 systems led cloud submissions on DeepSeek-R1 interactive throughput by more than 10%, according to Signal65 figures cited by Fathom News. Power efficiency work is continuing using Nvidia MaxLPS.
Microsoft is also a customer of the technology it is reselling: the company uses Nvidia's Dynamo inference software for post-training and RTX PRO 6000 Blackwell GPUs for its Bing search models.
Reaching local and edge systems
Expanded Nemotron 3.5 support in Foundry Local brings reasoning, coding and tool-using agents into Azure Local environments, together with retrieval models that locate evidence across text, document images, charts and tables. Nvidia's Cosmos physical AI models are also being brought to Foundry Local.
Azure Local will additionally support RTX PRO 6000 Blackwell GPUs from hardware partners for on-premises deployment, and will expand to Jetson Thor and Jetson Orin edge platforms for industrial applications such as machine vision, robotics and autonomous operations.
On the desktop side, the companies plan to showcase local large language model experiences on Windows PCs powered by Nvidia RTX Spark. Windows 365 Cloud PCs are getting Direct Stream in limited preview, supporting demanding graphics workloads with up to two 4K streams at 60 frames per second and low latency even on difficult networks; a public preview is planned for early 2027.
For scientific research, Nvidia's BioNemo Agent Toolkit and ALCHEMI Toolkit are entering 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 laboratory evaluation.
Why it matters
The EY numbers are the part to watch because they make the open-model argument concrete. An organization can download a model, tune it on its own records, run it on infrastructure it controls, and outperform a frontier model on a narrow task at a fifteenth of the cost. If that pattern repeats, more businesses will question why routine work such as sorting purchase records is sent to the most expensive closed models.
The arrangement also serves both vendors. Microsoft keeps customers inside Azure whether they choose closed frontier models or cheaper Nvidia alternatives, while Nvidia supplies silicon at every tier, from Vera Rubin racks in the cloud down to workstation and edge GPUs. The open question is adoption: whether more enterprises publish cost and accuracy results like EY's, and whether the promised Bolt and Transcribe models arrive in Foundry in time to shift additional everyday workloads away from pricier closed models.
- #nvidia
- #microsoft
- #azure
- #open-models
- #edge-computing