In factories, warehouses, retail stores, and field sites, a new generation of systems can perceive what's in front of them, reason about it, and act: a robot arm that adapts to each part it handles, a camera that catches a defect before it ships, a sensor that hears a failing bearing before it breaks. Advances in models, sensors, and edge compute hardware are making this possible.
Building one of these systems is exciting. Running hundreds or thousands of them is a different challenge. Every device needs to be provisioned, secured, monitored, and updated – often under tight constraints on size, power, and cost, and ideally without sending anyone on-site.
If you know Azure Local, you probably know it running on datacenter-class servers. But you can't bolt a server rack onto a robot. That's why earlier this year we introduced a new deployment option of Azure Local for small form factor hardware. There's no hypervisor: each device runs Linux directly on bare metal to host containers, and shows up in Azure as a Provisioned Machine that looks and behaves a lot like an Azure VM. In effect, you can manage edge machines like cloud machines.
Industry example: Danfoss
Let's see this in action.
Consider a company like Danfoss, a leading manufacturer of valves, pumps, and drives. They make over 800,000 unique products, often in small batches. With that much variety, cost-effective automation can be hard to achieve. Traditional robotics, with pre-defined movements that repeat exactly, is too inflexible when every product goes together a little differently.
So Danfoss teamed up with Universal Robots and Microsoft to explore a different approach: generalized physical intelligence. Instead of scripted movements, what if a robot could take a plain-language instruction, like "fit the heat sink into the bottom plate," and figure it out?
Watch the video to see a UR7 partially assemble one of Danfoss's iC7 variable frequency drives, using a vision-language-action model (Pi 0.5) to predict the robot's next joint positions based on what it sees. All the physical intelligence required runs locally with Foundry Local, on an NVIDIA Jetson Thor managed by Azure Local.
To be clear, this is a technology demonstration hosted in a Microsoft test lab, not something running on Danfoss's factory floor today. But it's a vivid glimpse of what's becoming possible.
What's new in 2610
This week at Microsoft Ignite 2026, we've released an updated preview of Azure Local for small form factor hardware, called version 2610, alongside updates to Foundry Local and other services. Here are the highlights:
Azure Local expands to arm64
By far the biggest news: for the first time, you can run Azure Local and many of Azure's edge services on arm64 devices.
Azure Local for small form factor hardware now works with NVIDIA Jetson AGX Thor Development Kit, in public preview.Many of the most capable platforms for physical intelligence are built on arm64, like NVIDIA Jetson Thor, a go-to platform for robotics development. With 2610, we're adding the Jetson AGX Thor Developer Kit in preview, using a Microsoft-provided Ubuntu image. Getting here took a ton of engineering work behind the scenes, together with our partners at NVIDIA. The result is full feature parity with the Intel-based devices we already offered, from Azure-based provisioning and configuration to security and monitoring, all natively in the Azure portal, CLI, and APIs.
And arm64 support extends beyond the infrastructure to the services that run at the edge, including Foundry Local (as seen in the Danfoss video), Azure IoT Operations, and Azure Kubernetes Service, opening the door to a whole new class of edge opportunities.
GPUs as first-class resources
For intelligent systems, the GPU is often the most important (and most expensive) part of the machine. Now, when you provision an edge device with Azure Local, the Microsoft-provided OS image includes NVIDIA drivers baked in, so your GPU is automatically discovered and ready to use.
To give it first-class management, the GPU is represented as its own resource in Azure Resource Manager – a child resource of the Provisioned Machine, just like disks and network interfaces. Standard (free) metrics for GPU utilization, temperature, and power flow automatically into Azure Monitor, so you get observability without wiring anything up yourself. Use them for alerts, or visualize them with Azure Managed Grafana and similar tools.
Wireless networking (Wi-Fi)
Not every intelligent system lives at the end of an Ethernet cable. Mobile robots and devices spread across large facilities or remote sites often depend on wireless connectivity. With 2610, you can configure the Wi-Fi interface from Azure alongside your other network settings, with secrets like passphrases stored securely in Azure Key Vault, so you keep centralized control no matter how a device connects.
Use Azure portal to view and configure the network interfaces of a Provisioned Machine, including wireless (Wi-Fi) interfaces.Image-based A/B updates
At the edge, a failed update can be one of the costliest problems to fix: a device that doesn't come back online may need a site visit or a replacement. Azure Local for small form factor devices guards against that by updating the bare metal OS using an A/B image swap approach.
The new image installs on an inactive partition while the current OS keeps running. On reboot, the device switches over – and if the new image fails to boot, it automatically rolls back to the original OS, which is known to work. Starting with 2610, we'll ship some otherwise-minor updates to preview participants as whole new images, to put this mechanism through its paces before general availability.
And so much more
All of this builds on the capabilities launched in 2604 and 2607, and is inspired by your feedback – so please keep it coming! This post only covers the big, glamorous features, but there are 249 other enhancements and fixes across the product in version 2610. And the team is already hard at work on the next release, version 2701, coming soon.
Get started
We're excited to extend Azure-managed infrastructure across a greater diversity of edge silicon and hardware than ever before, bringing the power of the cloud closer to where physical work happens. You can get started today:
1. Try Azure Local on small form factor hardware. This deployment option is currently in public preview, with no charge during the preview. You don't necessarily need an NVIDIA Jetson Thor: supported devices start as low as around $1,000 each. Instructions to get started are on Microsoft Learn.
2. Build physical AI and robotics scenarios with Microsoft. Sign up through the Microsoft Robotics interest form to connect with the team.
We genuinely had a blast building this demo with Universal Robots and Danfoss. There's something magical about watching a robot work out an assembly step that nobody explicitly programmed. Whether you're managing a fleet of edge machines or teaching one to think for itself, we can't wait to see what you build.
<3 Cosmos
on behalf of our global team