Open up almost any connected industrial device and you will find a small square board soldered near the antenna. It is the module: a pre-certified package containing the radio, the modem and just enough processing to get data onto a network. For most of the past two decades, its job has been narrow. It moved bytes. Anything clever happened somewhere else, on a separate processor inside the device or, more often, in the cloud.
That is changing. In the space of a few weeks in September 2026, a module maker launched a vision module with its own AI processor, another released software to run machine learning models directly on cellular modules, a chip company and a software firm teamed up to manage AI workloads on edge hardware, and fresh market research showed Bluetooth chips taking on AI tasks as well.
Taken one at a time, each announcement is a product launch. Taken together, they describe a shift in where intelligence lives in an IoT system. The smallest, most standardised component in the device is starting to think for itself.
From Modem To Computer
To understand why this matters, it helps to recall why modules exist at all. Designing a cellular radio from scratch is expensive and slow. Every new design needs regulatory approval and certification from mobile operators before it can be sold. Modules solve that problem. A device maker buys a pre-approved block from a supplier such as Quectel, Telit Cinterion or Fibocom, adds sensors and a case, and gets to market far faster.
Over time, modules picked up extra duties. GNSS positioning moved inside. So did Wi-Fi and Bluetooth in some designs. Some “smart modules” began running full operating systems such as Linux or Android, turning the module into the device’s main computer rather than a peripheral.
AI is the next step on that path. Instead of streaming raw sensor data or video to the cloud for analysis, the module can run a model locally and send only the result: a reading, an alert, a classification. The data leaving the device shrinks from a stream of raw measurements to a short statement of what happened.
What Launched This Month
Four September announcements show how the idea is being applied across different kinds of hardware.
Quectel’s Vision Module
On 21 September, Quectel introduced the SE200ZC-AP, a smart module built around Rockchip’s RV1126B platform. According to IoT Business News, it pairs a quad-core Arm processor with a neural processing unit rated at 3 TOPS, a measure of how many trillion AI operations a chip can perform each second.
The module supports up to five cameras, records 4K video, runs a preloaded Linux distribution and connects to external Wi-Fi, Bluetooth and LTE Cat 1 modems. It measures 40 by 40 millimetres and comes in an industrial version rated from -35°C to 80°C. Quectel is aiming it at security cameras, industrial automation, vehicles, robot vision and smart home devices. It was listed as coming soon for customer sampling.
Zeljko Maric, a product development manager at Quectel, described the goal as giving customers “a single, ready-to-deploy platform that combines vision, AI, and connectivity so they can focus on their application instead of reinventing the hardware underneath it.”
Telit Cinterion Puts AI On The Modem
Telit Cinterion took a software route. Earlier in the month it detailed an Edge AI SDK that lets developers run machine learning models directly on its cellular modules, as reported by IoT Tech News. The kit embeds LiteRT, Google’s lightweight runtime formerly known as TensorFlow Lite, so teams can use models trained in familiar tools.
The SDK covers 4G, 5G RedCap and higher-performance 5G modules. Example uses include analysing vibration and sound from machines, detecting breaking glass or alarm tones, reading old analogue meters with a small camera, and basic image classification and object detection. Telit Cinterion says computer vision tasks peaked at 17% of the module’s processor capacity in its tests, low enough to avoid heat problems that could disrupt the cellular connection. The SDK is due in the fourth quarter of 2026.
Marco Argenton, the company’s senior vice president of product management, summed up the pitch: “Industrial IoT teams should not have to redesign their entire device architecture to add practical AI capabilities.”
Bluetooth Chips Take On More Work
The shift is not limited to cellular. Writing for IoT Business News, Takeshi Niwa of Techno Systems Research described how the Bluetooth chip market is shifting toward IoT data applications and away from its traditional base in audio.
The firm expects Bluetooth chip shipments to grow from 8.3 billion units in 2025 to 10.3 billion in 2032, with data applications drawing level with audio by the end of that period. Growth areas include continuous glucose monitors, industrial and logistics tags, digital car keys and item finders. Its analysis makes the AI connection explicit: “Edge AI is increasing the value of BLE SoCs beyond connectivity, especially in industrial, healthcare and smart-home applications.”
The market structure is worth noting too. Nordic Semiconductor led Bluetooth Low Energy chip shipments in 2025 with 21% of units and 36% of revenue, while a group of Chinese suppliers took 29% of shipments but only 9% of revenue. Low-cost volume and higher-value features, including on-chip AI, are pulling the market in two directions.
Cameras That Inspect On Their Own
On the factory floor, the same idea shows up in cameras. Electronics manufacturer USI launched an AI smart camera designed to detect defects, check assembly, read text and guide robots without sending video elsewhere for analysis, IoT Tech News reported. It comes with no-code tools for collecting images, training models and deploying them.
USI says the camera is already in use in its own factories, and that an earlier automated inspection system it ran in 2023 and 2024 reached 85% defect detection. The company did not disclose the camera’s processor, AI performance, pricing or availability. Justin Chang, a director at USI, explained the thinking: “Today’s manufacturers are seeking more than just camera hardware.”
Why Put AI Inside The Module
None of these companies is adding AI purely for novelty. There are practical reasons device makers want intelligence closer to the sensor.
Fewer Chips, Fewer Designs
Every extra chip on a circuit board adds cost, power draw, board space and engineering time. If the module can already run a vision model or analyse vibration data, a device maker may not need a separate AI processor or an extra microcontroller. For companies building thousands or millions of units, removing one component can change the economics of a product.
It also shortens development. Because modules are pre-certified, putting more functions inside them reduces the amount of new hardware that must be tested and approved.
Less Data To Send
Connectivity costs money and power. A camera streaming video over a cellular link uses far more data than one that sends a short message saying a defect was found. A vibration sensor that uploads raw waveforms drains its battery faster than one that reports “bearing wear detected.”
This matters most for the low-bandwidth networks many IoT devices rely on. As we explained in our guide to the networks built to whisper, low-power networks are designed for small, infrequent messages. Running a model locally lets a device do useful analysis and still fit its findings into a few bytes.
Decisions That Survive A Dropped Connection
Networks fail. Coverage drops in basements, rural areas and moving vehicles. A device that depends entirely on the cloud to interpret its data becomes blind when the connection goes. A device that runs its own model can keep detecting problems, raising local alarms and storing results until the link returns.
Latency is the related issue. Some decisions, such as rejecting a faulty part on a production line, need an answer in milliseconds. Sending an image to the cloud and waiting for a response adds delay that local processing avoids. We looked at why factories value predictable, low-delay connections in our explainer on industrial IoT and industrial connectivity.
The Limits Of Small AI
It would be easy to overstate what a module can do. The AI running on these devices is very different from the large models running in data centres.
Power And Heat
A module’s first job is still connectivity. Every cycle spent on AI is a cycle not spent managing the radio, and every extra watt becomes heat inside a small, often sealed enclosure. Telit Cinterion’s decision to publish a figure for processor use during vision tasks reflects this. If AI processing makes a module too hot, it can slow itself down or disrupt the cellular link, which defeats the purpose.
Battery-powered devices face an even tighter budget. A sensor designed to run for ten years on one battery cannot run a model continuously. In practice, many designs wake the AI only when a simpler trigger, such as a motion or sound threshold, suggests something worth analysing.
What 3 TOPS Can And Can’t Do
A 3 TOPS neural processor, like the one in Quectel’s new module, is well suited to tasks such as detecting objects in a camera frame, classifying images or recognising specific sounds. It is not designed to run large language models or complex reasoning of the kind now appearing on much larger edge computers.
That is not a weakness so much as a design choice. Most industrial AI tasks are narrow: is this part defective, is this machine vibrating abnormally, is there a person in this zone. Small, focused models handle those questions well. The risk lies in marketing that blurs the difference between a module that can classify an image and a system that can understand a scene.
Managing AI Across Thousands Of Devices
Adding AI to one device is an engineering task. Keeping AI working across tens of thousands of devices spread over many sites is an operations task, and it is harder.
Orchestration Moves To The Edge
That is the problem Ambarella and ZEDEDA set out to address. On 16 September, the two companies announced they had brought ZEDEDA’s EVE-OS operating system to Ambarella’s N1 family of edge AI chips, validated first on the N1-655 processor. The combination is meant to let companies deploy and manage vision and language models on field devices remotely, with secure setup, monitoring and support for containers and Kubernetes.
Early access began at announcement, with production development kits planned for the fourth quarter of 2026. ZEDEDA’s founder and chief executive, Said Ouissal, described the aim as giving developers “a cloud-like experience for deploying and operating AI.” Ambarella’s chief growth officer, Muneyb Minhazuddin, pointed to “intelligent cameras, robots, vehicles, and industrial machines” as the targets.
Updating Models In The Field
Models are not static. A defect detector trained on one product needs retraining when the product changes. A sound classifier may perform differently in a new building. Over time, real-world conditions drift away from the data a model was trained on.
That makes model updates a core part of device management, alongside firmware updates and security patches. Teams deploying AI on modules need to plan how models will be tested, rolled out in stages, monitored for accuracy and rolled back if something goes wrong. The tooling for this is still maturing, which is one reason announcements like Ambarella and ZEDEDA’s are appearing now.
What This Means For Device Makers And Buyers
For companies designing connected products, AI-capable modules open options that were previously reserved for more expensive hardware. A meter reader, a machine monitor or a small camera can now make simple judgements on its own without a separate AI chip. The trade-off is that the device team takes on new work: choosing and training models, testing them under real conditions and maintaining them over the product’s life.
For buyers, the questions change as well. It is no longer enough to ask which network a device uses and how long its battery lasts. It becomes important to ask what the device decides locally, how accurate those decisions are in conditions like yours, how models are updated, and what happens to the raw data that the device no longer sends.
Several of the products described here are not yet shipping in volume. Quectel’s module was listed for sampling, and both the Telit Cinterion SDK and the Ambarella and ZEDEDA development kits are due in the fourth quarter. Real-world performance will only become clear as customers deploy them.
A Quiet Change In The Device
The module has always been one of the least visible parts of IoT: a certified block on a circuit board that most end users never think about. That anonymity is part of why this shift matters. When the component that every device already contains gains the ability to analyse data, AI stops being a separate feature that has to be designed in and becomes something close to a default.
It will not replace cloud analytics or larger edge computers. Complex reasoning, fleet-wide analysis and model training will still happen elsewhere. But the first judgement about what a sensor is seeing, whether a sound is normal, whether a part is faulty, whether anything is worth reporting at all, is moving onto the device itself. For an industry that has spent years building pipelines to move data to where the intelligence was, that is a meaningful reversal.
