Edge AI being operationalised in factories

Factories are starting to run artificial intelligence next to the machines, not in distant cloud data centers, so operators can troubleshoot production problems in real time. (premioinc.com) In one case published April 13, 2026, Premio said a manufacturer deployed its LLM-1U-RPL server on site so operators could search programmable logic controller alarms, inspection-camera results, and maintenance records through a local interface. Premio said the system was built around localized inference because existing factory systems were not designed for graphics processing unit workloads and the plant wanted to keep sensitive operational data on site. (premioinc.com) The hardware details point to why these projects are moving off the cloud and into cabinets and control rooms. Premio said its server used an NVIDIA RTX PRO 4500 Blackwell graphics processing unit, redundant 600-watt power supplies, hot-swappable fans, and multiple 2.5 gigabit Ethernet ports, and that deployment took four to five weeks. (premioinc.com) Edge computing means processing data where it is generated, on the shop floor or inside the plant network, instead of sending every query away for analysis. Siemens says its Industrial Edge platform is designed to collect, process, and analyze industrial data on the factory floor and connect operational technology systems with information technology systems. (siemens.com) That setup fits the way factory data is actually stored. Amazon Web Services said on July 15, 2025 that manufacturers typically collect data from multiple automation systems with different formats and protocols, and a Forrester Consulting survey of 500 manufacturing leaders found 98% reported at least one data issue inside their organizations. (aws.amazon.com) The new wrinkle is that language models are being asked to work like a search-and-summary layer for messy plant records rather than as a replacement for control systems. Premio described operators using local large language model inference to review engineering documentation and machine history during active shifts, while Siemens markets edge software for real-time performance monitoring, maintenance, and defect detection. (premioinc.com) (siemens.com) Smaller edge boxes are part of the same push. NVIDIA’s Jetson Platform Services documentation says the software stack supports Jetson Orin AGX, NX16, NX8, and Nano 8 gigabyte devices and packages video analytics, text prompting, monitoring, alerts, and application programming interfaces into containerized services for on-device deployment. (docs.nvidia.com) Industrial vendors are also keeping some artificial intelligence tightly coupled to controllers rather than sending it to a remote model. Rockwell Automation says FactoryTalk Analytics LogixAI runs predictions at the edge using streaming controller data from ControlLogix and now supports CompactLogix 5380, with the goal of making process predictions at speeds required for process-critical control. (rockwellautomation.com) Analysts have been describing the same shift in broader terms. Gartner said in research published December 9, 2024 that edge computing in manufacturing is being driven by low-latency needs, with predictive maintenance and smart manufacturing among the main use cases, and ARC Advisory Group wrote in December 2025 that industrial artificial intelligence is moving toward physically deployed systems because of cost, latency, governance, and trust barriers. (gartner.com) (arcweb.com) What is emerging is less a single product category than a factory architecture: controllers keep running the line, while nearby servers and embedded nodes read alarms, logs, images, and manuals fast enough to help the next technician make a decision before the line stops. (premioinc.com) (rockwellautomation.com)