Market and industryNovember 20, 2026Редакция 105.kz
NVIDIA and Microsoft Expand Local AI: Azure Local, Jetson Thor and PCs for Industry
NVIDIA and Microsoft expanded their collaboration at Ignite: Nemotron 3.5 models in Foundry Local, RTX PRO 6000 Blackwell GPUs, and Jetson Thor and Orin edge platforms for machine vision and robotics. We look at what of this is applicable at a plant in Kazakhstan.
Key takeaway. NVIDIA and Microsoft have expanded the local AI line-up: Foundry Local and Azure Local gain Nemotron 3.5 models, RTX PRO 6000 Blackwell GPUs and the Jetson Thor and Orin edge platforms for machine vision and robotics. For a plant in Kazakhstan this widens the set of options for running AI inside its own perimeter, but applicability to specific units, legacy process control systems and Russian-language documentation has to be verified in a pilot on your own data.
What happened
At the Microsoft Ignite conference, NVIDIA and Microsoft announced an expansion of their collaboration in accelerated computing — from Azure infrastructure to industrial systems and Windows PCs. According to the companies, NVIDIA Vera Rubin NVL72 systems already run in Azure data centers and are designed for compute-intensive reasoning tasks, agentic applications, and large-scale training and inference of models. In parallel, they announced an expansion of confidential computing to bare metal and rack-scale AI infrastructure: attestation lets the compute environment be verified before models, data or keys are passed into it. Work on energy efficiency with NVIDIA MaxLPS was named as a separate track. In the MLPerf Training v6.1 round, Azure scaled to 10,240 NVIDIA GPUs on GB200 NVL72 systems while training DeepSeek-V3 — the largest scale in the round; in MLPerf Inference v6.1, GB300 NVL72 systems in cloud submissions outperformed other solutions in interactive throughput on DeepSeek-R1 by more than 10% (according to Signal65, as cited in the publication).
On the tooling side, Microsoft Foundry added a serverless Programmed Training service for managing post-training and Foundry Managed Compute with support for deploying your own model weights. The NVIDIA Nemotron 3.5 open model line includes Super (long-running agents, code, research and business workflows), Lightning (fast responses and tool calling), Embed and Rerank for search, Parse for extracting structured data from documents, and Content Safety; Bolt and Transcribe are listed as 'coming soon'. One example cited is EY: adapting open Nemotron models to classify corporate procurement records yielded a 55% accuracy gain on an evaluation set and a 15x reduction in cost compared with proprietary frontier models. For sites outside the cloud: Foundry Local and Azure Local gain support for Nemotron 3.5, including retrieval models that find answers in text, document images, charts and tables; Azure Local will support NVIDIA RTX PRO 6000 Blackwell GPUs from hardware partners and will extend to the Jetson Thor and Jetson Orin platforms — for machine vision, robotics and autonomous operations. On PCs with NVIDIA RTX Spark, local large language models were demonstrated; for Windows 365 Cloud PCs, Direct Stream streaming was announced (up to two 4K streams at 60 frames per second), with sign-up open for a limited preview and public availability planned for early 2027. Microsoft also uses NVIDIA Dynamo for inference during reinforcement learning and RTX PRO 6000 Blackwell GPUs for Bing search models.
What this means for enterprises in Kazakhstan
Plants and manufacturing enterprises. The practical meaning of the announcement is not the release itself but the framing of three tiers of AI execution: public cloud, a shop-floor server with a GPU, and an edge device directly on the line. For a plant this widens the solution space across two classes of tasks. The first is finding answers in operating documentation: datasheets, process regulations, drawings, work orders, parameter tables. The declared support for retrieval models over text, document images, charts and tables maps directly onto the structure of typical plant documentation (estimate). The second is video-based control of operations and positioning in hazardous areas, where an edge device is the logical choice. The data for such tasks already exists inside the perimeter: process control system and SCADA historian data, work permit logs, a library of regulations and datasheets, and the video archive from control posts. On-premise remains the norm here: the information security requirements of national companies and personal data localization do not envisage moving production data into a public cloud, which is why the local segment of the announcement matters.
Energy. The declared work on compute energy efficiency (NVIDIA MaxLPS) is relevant when planning an in-house shop-floor or corporate compute node: for an energy company, the cost of ownership of AI infrastructure includes electricity consumption (estimate). The source gives no specific savings figures, so this can only be treated as a direction of work, not as a basis for calculations.
Uranium and mining and metallurgy. More significant are the Jetson Thor and Orin edge platforms declared for machine vision, robotics and autonomous operations. For a concentration stage or an open pit, this is the class of tasks covering PPE compliance, fencing off hazardous areas, and recording violations on conveyor lines (estimate). The limitation is the same everywhere: integration with legacy process control systems and SCADA requires separate engineering, and the source contains neither protocol descriptions nor claims of compatibility with GOST or industry standards.
A separate note on language. The source does not claim support for Russian or Kazakh and gives no quality measurements on non-English documents. For an operator advisor this is a key parameter: search and answer quality in Russian has to be measured on your own regulations rather than carried over from demonstrations (estimate).
How to apply it: an 8–12 week pilot
- Pick one process and one class of task. For example, answering questions from operating documentation in one area, or monitoring a single operation by video. Record the baseline metric 'as is': time to find a document, share of picking errors, number of recorded violations per shift.
- Audit the data inside the perimeter. Determine which reference books, regulations, datasheets, historian data and video files are available, what condition they are in, who owns them and what access rights are needed. The NDA and security review belong at this step, not after the model is chosen.
- Choose where it runs. A shop-floor server with a GPU or an edge device on the line; the criteria are acceptable latency, data volume, security requirements and availability of maintenance staff. This is a management decision, not a technical detail.
- Compare the chosen approach with the baseline on your own dataset. Acceptance is by a metric on a holdout set of real documents and records, not by a demonstration. The sample size and acceptance criterion are fixed before the start.
- Measure the effect and record the limitations. What worked, on which units, where it failed, what needs rework. The scaling decision is made on the measured result.
Baseline approaches to this kind of work are described on the Corporate LLMs and Machine Vision for HSE pages.
Limitations and risks
- Almost everything listed is a statement of availability and plans. The public preview date for Direct Stream (early 2027) and the 'coming soon' status of the Bolt and Transcribe models show that part of the announcement refers to the future.
- The EY case concerns classification of procurement records and was measured on an evaluation set, not in industrial operation. The 55% gain cannot be transferred to equipment diagnostics or safety monitoring tasks.
- The MLPerf numbers measure the throughput of cloud systems on specific models, not the quality of a solution for a plant.
- The source contains no data on cost, delivery times or availability of RTX PRO 6000 Blackwell hardware and Jetson platforms in Kazakhstan, and no information on support for Russian and Kazakh or on compatibility with legacy process control systems.
- The facts in the source are sufficient to assess the direction but not to make an engineering choice of a specific configuration: the infrastructure decision will have to be made after an in-house measurement at a pilot area.
Sources
- NVIDIA and Microsoft Expand Accelerated Computing From Azure to Windows PCs · 2026-11-19T20:00:15.000Z
Azure Local is expanding to NVIDIA Jetson Thor and Jetson Orin edge AI platforms to support industrial applications such as machine vision, robotics and autonomous operations.
- Фото: August Geyler — FOT3917 Schönbachstraße 91 (Leipzig) – ehemalige Walzenfabrik Felix Böttcher.jpg (CC BY-SA 4.0, Wikimedia Commons)
This material was prepared with AI from the sources listed above. Spotted an inaccuracy? Write to info@105.kz.