Blog Post

Internet of Things Blog
13 MIN READ

Azure IoT Operations 2610 is now available: Scaling the foundation for Physical AI

seanparham's avatar
seanparham
Icon for Microsoft rankMicrosoft
Nov 17, 2026

AI is moving out of dashboards and onto factory floors, oil fields, and production lines, where it can sense, reason, and act in the physical world. Most industrial organizations are only beginning to make that shift, and they need a platform that lets them start with confidence and keep going as their ambitions grow. Azure IoT Operations 2610 is built for exactly that.

For the first time, Azure IoT Operations ships on two channels at once. The generally available (GA) release delivers production-ready capabilities that help organizations understand their operations across many more sites, at lower cost. A new public preview channel gives customers hands-on access to the capabilities that are coming next, including governed access to operations for AI agents and AI inference running directly inside edge data pipelines. Taken together, the two channels show where the platform is today and where it is heading.

This release is part of a broader set of Physical AI announcements at Microsoft Ignite. Microsoft's Ignite Physical AI announcement (https://aka.ms/Ignite2026/PhysicalAIBlog) describes how IoT, Industrial AI, and Physical AI connect data, decisions, and action: IoT helps organizations understand their operations, Industrial AI helps them reason, and Physical AI helps them act, together forming a continuous learning loop. Within Microsoft's adaptive cloud approach, Azure IoT Operations is the on-site foundation for that loop. This post goes one level deeper into what 2610 delivers and what it makes possible.

Physical AI and where Azure IoT Operations fits

Physical AI is AI that works in the physical world. It perceives what is happening across an operation, reasons about what it means, and acts to improve the outcome. Getting there is a journey rather than a single leap, and customers should not feel they have to take on everything at once. Azure IoT Operations is designed to make that journey practical, so organizations can start simply, prove value, and evolve toward far more sophisticated applications of AI and agents at their own pace.

The journey begins with easy integration into the physical operation customers already run. Many start with perception: connecting equipment, bringing in data from sensors and cameras, and seeing what is happening in real time. From there, they move into reasoning, with AI models and agents that diagnose issues and recommend improvements. Ultimately, they move into action, where approved decisions reach equipment through partner control systems. Each step builds on the one before it, on the same platform.

Azure IoT Operations is also designed from the ground up to protect each customer's proprietary operational data. Customers decide who can access which data and how, and that applies to AI agents as much as to people and applications. They also decide what data leaves the site, if any at all. That control lets customers choose the deployment model that fits their operation: cloud-first deployments that rely on AI in the cloud, hybrid deployments that combine cloud AI with local AI models and agents, and fully local deployments where AI and agents run only on site and data never leaves it.

Whichever path and deployment model a customer chooses, the role of Azure IoT Operations comes down to three core functions and two key intersections with the systems around it.

Three core functions

Understand the operation. Azure IoT Operations keeps trusted operational context close to the process: what each asset is, what its data means, how equipment relates, its current and historical state, observations from cameras and other sensors, and the capabilities and constraints of the process. Standards such as OPC UA and W3C Web of Things are the tools. The outcome is context that people and AI can rely on.

Optimize the operation. AI models and agents run at the edge, next to the operational data and the process, where they can perceive, diagnose, recommend, optimize, and coordinate. Microsoft Foundry supplies the models and agent development capabilities. Azure IoT Operations supplies the operational environment in which those agents become useful.

Govern action. When an agent proposes an action, Azure IoT Operations turns it into an authorized, bounded, and observable operation. It authenticates the actor, evaluates policy and current conditions, obtains approval where required, invokes an authoritative command path, records the decision, and verifies the result.

Two key intersections

Where AI meets control: Azure IoT Operations optimizes, partners actuate. Azure IoT Operations decides, predicts, and recommends, sometimes in the form of a control input such as a new setpoint, and issues that intent through governed interfaces. Partner control systems and robots run the deterministic real-time loop and own safety, for people and machines, at all times. As a simple example, Azure IoT Operations can set the thermostat, but the thermostat and the HVAC system get the room to the correct temperature.

Partners are already proving this pattern. Azure IoT Operations integrates with Rockwell Automation's FactoryTalk Optix and is part of Schneider Electric's EcoStruxure Automation Expert. In Schneider's green-hydrogen pilot with H2E Power, AI agents validated control logic in simulation before deploying it to hardware; the system ran for more than 6,000 hours autonomously and cut the levelized cost of hydrogen by roughly 10%.

Where edge meets cloud: local intelligence at the edge, enterprise intelligence in the cloud. Azure IoT Operations integrates local operational intelligence at each site. Microsoft Fabric integrates enterprise-wide and cross-site intelligence. The two are federated, so when customers choose to share data beyond the site, the context created at the edge flows to the cloud without being rebuilt.

These three core functions follow the same path customers take into Physical AI: understanding the operation turns raw data into trusted context, optimizing it is where AI reasons over that context, and governing action is how decisions safely reach the physical world. Clear roles at each intersection are what make AI credible to the people who run plants. Customers get AI that is useful on the floor without trading away the control systems, certifications, and safety practices they already trust.

What 2610 delivers

Azure IoT Operations already provides the foundation many industrial customers run in production today: governed cloud-to-edge management actions, no-code data flow graphs, broad industrial connectivity, and unified health reporting. Release 2610 builds on that foundation in two directions.

The GA release deepens the first core function, understanding the operation, and makes it practical at fleet scale. It also strengthens the intersection where edge meets cloud by landing edge data directly in the lakehouse. The preview release takes the first concrete steps into the second and third core functions: running AI against live operational data at the edge, and giving AI agents governed, auditable access to operations.

Core function

Generally available in 2610

In preview with 2610

Understand the operation

Essentials SKU for large fleets, ODBC-based connector for SQL Server, smarter OPC UA collection, leaner data flow writes to the lakehouse

Asset context with Web of Things and the edge registry, type-based OPC UA discovery, live data in the operations experience, Azure Marketplace connectors and the Akri Connector Builder, Microsoft Fabric integration

Optimize the operation

Service calls from custom WebAssembly modules, AI-assisted data flow authoring in VS Code

Inference connector with Foundry Local and Azure Machine Learning, file-backed graph processing

Govern action

Management actions for HTTP/REST assets

MCP server with policy-governed tool access, data-driven management actions, application lifecycle for agents and MCP servers

Two release channels: what you can run today, and what is coming next

Starting with 2610, every Azure IoT Operations release ships on two channels.

  • GA (stable) is production-ready and fully supported. This is what you run on the factory floor.
  • Preview includes everything in GA plus new capabilities that are ready for hands-on feedback. Use it in test or pilot environments to evaluate what is coming and help shape it before it reaches GA.

You choose a channel when you deploy, through the Azure portal deployment wizard or the Azure CLI, and you upgrade within that channel. Switching an existing instance between GA and preview is not supported, so production deployments stay on a predictable, stable path. Connector and transform versions follow the same rule: GA instances see only GA versions, while preview instances see both.

 

 

For technology leaders, the second channel changes how you can plan. You no longer have to wait for GA to evaluate a major capability or sign up for a one-off private preview. The preview channel is a working view of the roadmap, available every release, that your teams can put in front of real equipment and real operators while production keeps running.

Generally available: understanding operations at scale

Physical AI is only as good as the operational context beneath it, and that context has to exist at every site, not just at the flagship plant. The GA capabilities in 2610 turn on-site data into actionable intelligence: they make operational context cheaper to establish, broader in what it covers, and easier to hand up to enterprise analytics.

Bring Physical AI to more sites: Azure IoT Operations Essentials

A new Essentials SKU sits alongside the full-featured offering, now called Standard. Essentials delivers the core of Azure IoT Operations, the edge MQTT broker, Akri connectors, and data flows, on single-node clusters at a lower price point. Many organizations need a reliable edge data layer at hundreds or thousands of small sites, such as stores, substations, wind turbines, or remote equipment, that do not need the full industrial feature set. Essentials makes those fleet-wide rollouts cost-effective. Standard remains the choice for full industrial deployments that need OPC UA connectivity or multi-node clusters. You choose the SKU when you create an instance, and moving to Essentials requires a new deployment. Together with ARM64 support, generally available since release 2607, Essentials lowers both the cost and the hardware bar for line-level and pilot deployments.

 

 

Data flow graphs that land data directly in your lakehouse

Data flow graphs write directly to Azure Data Lake Storage Gen2 in Delta or Parquet, Microsoft Fabric OneLake tables in Delta, and Azure Data Explorer or local storage in Parquet. Teams can branch, window, filter, and run custom WebAssembly transforms at the edge, then write the result straight into analytics-ready tables with no extra messaging hop and no second pipeline to maintain. In 2610, Delta Lake writes use less memory and carry less transaction-log overhead, with checkpoints running in the background. Less, cleaner data goes to the cloud, which lowers storage and ingestion costs. This is where edge meets cloud, in practice: the edge does the local work, and Fabric receives context it can use for cross-site analysis.

More of the plant, connected

The new ODBC-based connector for SQL Server brings data from the databases behind MES, quality, batch, and historian systems into the same pipelines as OPC UA and MQTT telemetry. Teams configure a SQL Server connection and its assets in the operations experience and define datasets with SQL queries, with no custom extraction code. Management actions now extend to assets connected through the HTTP/REST connector as well, so teams can collect data from and act on REST-connected equipment alongside everything else. That brings governed management actions to more of the plant.

 

 

Smarter OPC UA data collection

OPC UA is a widely used open standard that lets industrial machines share their data, and 2610 makes the Azure IoT Operations connector for it smarter about what it collects and sends. It can now ignore values that barely change, drop readings the equipment itself flags as unreliable, and sample data at a fixed interval. Filtering out noise and bad readings at the source means less data to move and store, lower cloud costs, and cleaner input for analytics and AI. When a machine restarts or the network drops, applications can request a complete snapshot of the machine's current state, or receive one automatically when a key value changes, instead of waiting for values to trickle back in. Optional health signals confirm that the connection is alive and that data has been delivered, and an optional fast failover gets data flowing again sooner when a connection fails.

Faster to build, easier to operate

Several GA capabilities shorten the path from idea to running pipeline. Developers can describe a data flow in plain language to GitHub Copilot in VS Code, validate it locally, and deploy it. Custom WebAssembly modules in data flow graphs can now call request-response services over MQTT while processing a message, for enrichment, lookups, routing decisions, or asset management actions. The edge state store now supports paged key scans and multi-key reads, so applications can discover and retrieve the context available to them. Live data views in the Azure IoT Operations Data Flow extension for VS Code speed up commissioning and troubleshooting, and an updated navigation layout in the operations experience makes instances and their resources easier to find.

Public preview: AI that reasons over operations, with action that stays governed

The preview channel is where 2610 reaches furthest toward Physical AI. These capabilities put agentic AI to work on live plant data at the edge, within the governance controls the customer sets, and give agents a governed path into operations. Customers can evaluate them now in test and pilot environments.

Connect AI agents to edge operations, with guardrails

Azure IoT Operations now includes a Model Context Protocol (MCP) server that gives compatible AI agents and assistants authenticated, authorized access to live MQTT telemetry and state store data at the edge. A single endpoint aggregates supported first-party and third-party MCP tools, so teams can register their own MCP servers behind the same entry point instead of managing separate client connections. Authentication and authorization policies that you control govern access to every registered tool, so each agent gets only the tools its role allows, and publishing MQTT messages requires Admin permissions. New application lifecycle capabilities also let teams deploy and manage agents and MCP servers on the cluster, alongside connectors.

This is the third core function, govern action, taking shape. Teams want AI copilots that can answer "why is line 3 down?" or "which assets are unhealthy?" without giving AI broad access to production systems. With a single governed entry point, organizations bring the agent they already use and decide exactly what that agent can see and do.

Run predictive AI models in the data pipeline

The new Inference connector runs registered machine learning models, using predictive models served by Foundry Local on the cluster or models hosted in Azure Machine Learning, with multiple models running concurrently. Large inputs such as high-resolution images are read from files rather than passed through the message broker, and new graph-level file stores let custom WebAssembly transforms read and write those files on persistent cluster storage. Moving a model from a notebook onto a live production line usually means custom integration code for every model. With the Inference transform, a model becomes another step in the pipeline, and its results can feed downstream transforms or trigger management actions on connected assets, closing the loop from perception through reasoning to action. This is the second core function, optimize the operation, running where the data is.

 

 

Give data meaning: asset context and the edge registry

Raw tag values are not enough for analytics or AI to reason about an operation. A new WoT Enrich transform uses Web of Things mappings to give asset messages a consistent structure and add context such as what a value means and its units, and a new Microsoft Fabric integration formats and batches that telemetry for Eventhouse or Lakehouse. The new edge registry stores JSON Schemas and Web of Things Thing Models and Thing Descriptions on the cluster, where local workloads can reach them, and keeps them in sync with the cloud registry. Type-based OPC UA discovery finds assets and event groups from OPC UA type information, and assets such as Modbus pumps or S7 PLCs can be onboarded through OPC UA servers that implement the OPC UA Web of Things Connectivity API. When context travels with the data from the source, both Fabric reports and AI agents understand what they are looking at, even when cloud connectivity is interrupted.

See data as it flows, and connect more sources

Live data in the operations experience lets authorized users inspect incoming asset messages and validate data as it moves through transforms, WebAssembly modules, and data flow graphs, which answers the first question in every rollout: is the data arriving, and does it look right? To broaden connectivity, customers on the preview channel can discover and deploy supported third-party connectors from Azure Marketplace through the Azure portal or Azure CLI. For protocols that no connector covers yet, the Akri Connector Builder, a GitHub Copilot agent plugin, turns protocol specifications into connector code, tests, and documentation, and validates the result against a generated protocol simulator and a local Azure IoT Operations environment.

 

 

Azure IoT Operations in Azure Government

Azure IoT Operations is now in public preview in Azure Government (US Gov Virginia), running the 2610 GA feature set. US government agencies and their partners can evaluate edge data collection and processing in a government cloud, deploying through Azure CLI and using the operations experience. The preview channel is not available in Azure Government.

Physical AI in action

Customers are already putting this foundation to work. Unilever, working with Microsoft, NVIDIA, and Accenture, has built an advanced digital twin that captures the behavior of a whole factory as a connected system. The solution uses Azure IoT Operations to ingest and integrate sensor data, contextualize it at the edge, and feed it to edge and cloud AI services. By identifying bottlenecks, testing interventions, and recommending actions before issues affect production, it helps unlock greater efficiency, faster changeovers, and reduced waste, while giving engineers and factory teams AI-powered insight to make faster, better-informed decisions.

What comes next

Each Azure IoT Operations release now shows two things: what is ready for production, and what is coming next. Over the coming releases, the platform will keep moving from perceiving and informing, to reasoning and recommending, to orchestrating action through partner control systems. As it does, the human role shifts from approving every step to overseeing the system with the ability to intervene, and the platform's two key intersections, with partner control systems and with the cloud, stay the same throughout.

Get started

  • Deploy Azure IoT Operations 2610 with the Azure portal deployment wizard, Azure CLI, or Bicep, and choose the GA or preview channel and the Essentials or Standard SKU at deployment.
  • Upgrade existing instances with Azure CLI, following the release-specific upgrade instructions.
  • To try preview capabilities, deploy a new instance on the preview channel in a test or pilot environment.
  • Review the public documentation (https://aka.ms/aiodocs) and release notes (https://github.com/Azure/azure-iot-operations/releases) for full details.
  • Read Microsoft's Ignite Physical AI announcement (https://aka.ms/Ignite2026/PhysicalAIBlog) for the full picture across Azure IoT Hub, Azure Device Registry, Azure Local, Microsoft Fabric, and our partner ecosystem.

Azure IoT Operations 2610 strengthens the foundation that is in production today and shows where Physical AI is heading. Wherever your organization is starting from, we look forward to building the next era of Physical AI with you.

Updated Nov 16, 2026
Version 1.0
No CommentsBe the first to comment