When AI Truly Grows Into the Controller: EdgePLC and ARMxy at CIIF 2026
At the 2026 Shanghai Industrial Fair, the most important signal in industrial automation was not another cloud platform or standalone AI box. It was the emergence of the “AI PLC”—a controller in which AI inference, real-time control, and event scheduling are natively integrated.
EdgePLC: Not “PLC + AI,” but “AI PLC”
At the CODESYS booth, EdgePLC demonstrated CODESYS Runtime and edge AI inference running together on one device.
Take the flagship BL245. It uses a Rockchip RK3588J with a heterogeneous architecture: 4× Cortex-A76 + 4× Cortex-A55 + 3× Cortex-M0. The A cores run CODESYS and Linux applications; the M0 cores handle real-time I/O; and the integrated 6 TOPS NPU runs vision models such as YOLO.
The key point: inference results do not leave the device. They enter CODESYS control logic directly and drive I/O execution. AI is not a peripheral—it is a computing unit inside the controller. Control, computing, and inference close the loop on the same SoC.
The booth also mapped EdgePLC models to scenarios:
ARMxy: Modularity Matters More Than Miniaturization
ARMxy industrial computer series appeared alongside EdgePLC. Its size draws attention, but its design logic is more important: decoupling.
ARMxy uses a modular architecture of host unit + SoM core board + X board + Y board:
These combine freely into a large number of hardware configurations. The answer to an old industrial question—why custom hardware for every new project?—is simple: change the SoM to upgrade compute; change the X/Y boards to adapt interfaces. Hardware becomes a configurable, evolvable, reusable platform rather than a one-off engineering project.
Domestic AI Chips Are Moving to the Industrial Site
CIIF 2026 reinforced this direction. For the first time, the fair hosted an independent integrated circuit exhibition, the “Guoxin Exhibition,” with more than 300 exhibitors covering chip design, wafer manufacturing, packaging and testing, equipment, and materials. AI chip companies such as Muxi, Enflame, Tianshu Zhixin, and Dongfang Suanxin participated, and the top three domestic EDA companies were present.
More importantly, their focus has shifted from cloud training supernodes to industrial sites:
A notable detail: many domestic AI chip solutions at CIIF chose ARM architecture or on-device SoC routes rather than traditional x86 industrial control. Chip vendors are moving toward the edge; controller vendors are moving toward AI. The two lines have converged.
ARM Is Becoming the New Baseline for Industrial Control
Viewed together, EdgePLC, ARMxy, and the domestic AI chip trend point to a larger fact.
By 2025, ARM platforms accounted for 65% of the embedded development market, while traditional x86 solutions had fallen to about 12%. Industrial MPUs are replacing x86 as the core of edge controllers, industrial gateways, and HMIs.
One upgrade case: after an edge controller switched from x86 to a quad-core Cortex-A53, overall power consumption dropped from 35W to 8W, cost fell by 40%, and the control cycle stabilized at 500 microseconds.
The lesson: industrial controllers do not need “stronger general-purpose compute.” They need to fit real-time control, edge computing, AI inference, and industrial communication into one compact physical form at the right power and cost. The ARM SoC heterogeneous architecture—A core + M core + NPU—matches this requirement. EdgePLC’s BL245 and ARMxy’s modular platform are product choices built on this trend.
An Underestimated Signal
The grand narrative at CIIF belongs to AI platforms, humanoid robots, and industrial large models. But what determines whether industrial AI actually runs on the shop floor is often less conspicuous: an NPU added inside a controller, a swappable core board, a software runtime that migrates across SoCs, and domestic AI chips turning toward industrial sites.
The signal is clear: industrial AI is moving from “bringing compute into the factory” to “building compute into the controller itself.”
