Run AI at the edge

Factories, stores, vehicles and field sites produce data that is costly or slow to send to the cloud. Some sites lose connectivity, some decisions must happen in milliseconds, and some data must stay in a country or on premises by law or contract.

Running AI locally brings its own work: hardware choices, model packaging, updates across many sites and keeping a central view of what runs where.¹

• Single device or app: Foundry Local runs models on the device with no Azure subscription and no per-token costs, on Windows, macOS (Apple silicon) and Linux, using GPUs or NPUs when available and falling back to CPU. It offers an OpenAI-compatible API but is not designed as a server inference stack.
• Windows PCs with NPUs: Copilot+ PCs have NPUs rated above 40 TOPS; Microsoft recommends Windows ML for inference on the device NPU, and exposes built-in AI features through Microsoft Foundry on Windows APIs.
• On-site infrastructure: Azure Local runs Azure capabilities on customer-owned hardware, managed through Azure Arc, connected or disconnected; Microsoft prices it per physical core plus consumption charges. Disconnected operations have eligibility requirements, and Arc-enabled Kubernetes in disconnected mode is in preview.
• Industrial data: Azure IoT Operations runs on Arc-enabled Kubernetes with an MQTT broker, connectors including OPC UA, and data flows to Microsoft Fabric OneLake, Event Hubs, Azure Data Lake Storage and Azure Data Explorer. Microsoft states it can operate offline for up to 72 hours, with possible degradation.
• Fleet management: Azure Arc projects servers and Kubernetes clusters into Azure Resource Manager; core control plane features carry no extra cost, while services such as Defender for Cloud or Azure Monitor are billed normally.

When a simpler route is enough: if sites are reliably connected and latency is tolerable, calling a cloud model endpoint is easier to run than an edge stack.¹²³⁴⁵⁶