Skip to content
Weekly digest · Issue 2

Edge AI chips multiply, LoRaWAN goes to orbit, and the Cyber Trust Mark looks for buyers

Issue 2: why every microcontroller vendor now ships a neural accelerator, what satellite LoRaWAN is good for, and whether a U.S. security label changes purchasing.

This week in brief
  • Neural accelerators are becoming standard in mid-range microcontrollers, which changes what 'low power' inference means.
  • Satellite backhaul for LoRaWAN and NB-IoT is moving from demos to commercial plans for agriculture, shipping and utilities.
  • The U.S. Cyber Trust Mark is voluntary; its effect depends on whether retailers and procurement teams ask for it.
  • Open-source smart home platforms keep absorbing features vendors used to charge for, including local voice control.

Every MCU vendor has an NPU now

A year ago a neural processing unit on a microcontroller was a differentiator. This year it is a checkbox: most of the major vendors have announced or shipped mid-range parts with a small accelerator next to the Cortex-M core, and the toolchains have caught up enough that a quantized model can be compiled for the accelerator without hand-written kernels. The practical change is not peak performance but energy per inference. A keyword-spotting or anomaly-detection model that used to drain a coin cell in weeks can run continuously for much longer when the accelerator handles the multiply-accumulate work and the main core sleeps.

Why it matters: the hardware is ahead of the product thinking. Most teams still ship the model they trained in the cloud, shrunk; the opportunity is in models designed for the constraints from the start. Our TinyML guide covers quantization, pruning and the frameworks involved.

LoRaWAN via satellite is a real product category

LoRaWAN was designed for terrestrial gateways, but the same chirp modulation can be received from low Earth orbit, and several operators now sell or plan direct-to-satellite coverage for sensors in places a gateway will never be installed: pasture, open ocean, pipelines, remote substations. Throughput is tiny and latency is measured in orbital passes, which rules out anything interactive but suits a soil probe that reports twice a day.

The same story is playing out for cellular, where 3GPP’s non-terrestrial network work lets NB-IoT devices talk to satellites with standard modems. See the satellite IoT entry and the agriculture guide for the use cases that justify the cost.

Cyber Trust Mark: a label waiting for demand

The U.S. Cyber Trust Mark is a voluntary consumer label for connected devices that meet a baseline derived from NIST guidance: no default passwords, secure updates, data protection, a stated support period. The program’s effect depends entirely on whether retailers feature it and whether procurement teams write it into requirements. Compared with the mandatory Cyber Resilience Act in Europe and the UK’s PSTI rules, it is the lightest-touch of the three, which is both its appeal and its limitation.

Why it matters: manufacturers that already comply with ETSI EN 303 645 for the European market will find the overlap large. The regulation guide maps the three regimes side by side.

Open-source smart home platforms keep eating features

Home Assistant and the projects around it have spent the year absorbing capabilities that were once paid or cloud-only: local voice assistants, energy dashboards that speak to inverters and EV chargers directly, and first-class support for Matter and Thread. For readers of this newsletter the interesting part is less the hobbyist angle than the pressure it puts on vendors: when a free, local platform does energy management well, charging a subscription for the same feature gets harder. Our smart home guide discusses the ecosystem trade-offs.

Also worth your time

Sources

Topics: edge AI, NPU, LoRaWAN, satellite IoT, Cyber Trust Mark