Hello readers!

This week, we’re talking about the next challenge in edge AI, healthcare IoT’s sneaky connectivity problem, and more!

For most of edge AI's short history, the hard problem was compute — fitting a usable model onto silicon that couldn't spare much power or memory. That problem is largely solved. Model compression, purpose-built NPUs, and frameworks designed to run on hardware with only a few kilobytes of RAM have pushed inference onto devices that would have been impossible five years ago. What's left is the layer beneath the model: getting sensor data onto the chip fast enough for that inference to be worth running at all.

Robots and vision systems hit this wall before they hit a compute wall. Multiple camera and LiDAR streams can overwhelm a connection long before the inference workload becomes the limiting factor, and the standard fix has been to hand-stitch whatever interfaces a given sensor happened to ship with — MIPI here, I2C there, a proprietary connector for the camera, UART for something else. It works on a prototype. It doesn't survive contact with a second design, let alone a fifth camera added mid-project, because every new sensor means re-deriving the wiring from scratch.

That's the gap a new class of hardware is starting to close. Microchip's second-generation PolarFire FPGA Ethernet Sensor Bridge, released earlier this month, replaces that scattered mix of interfaces with a single 10Gb Ethernet connection built on NVIDIA's Holoscan Sensor Bridge. A camera plugs in the way a network device does. The board also supports twice the camera count of its predecessor at 60 percent of the size, and it carries onboard circuitry for measuring latency from sensor capture through inference — a detail aimed less at the demo table and more at systems where that number has to be provable, like surgical robotics or industrial vision.

The same instinct is reshaping the layer above the sensor. As edge deployments scale from a handful of prototype devices to fleets in the thousands, ad hoc management stops working and organizations move toward unified orchestration platforms that can absorb any device or protocol without a rebuild each time. Sensor connectivity is following the identical path, just a step behind — standardization creeping down the stack as each layer above it stops tolerating bespoke wiring.

Anyone evaluating an edge AI platform this year would do well to look past the inference benchmark. How many sensors can it take on, over what interface, and how long does adding one more actually take? That number, more than the chip's throughput, tends to decide whether a pilot ships on schedule or spends its budget on cabling and driver work instead.

Connected devices are leaving controlled environments and showing up in ambulances, clinics, patients’ homes, and everywhere in between. They’re collecting data, communicating with platforms, and supporting care that depends on connectivity being available when it matters. Yet connectivity is often treated like a commodity.

Read more

Voice is becoming the preferred interface for connected hardware — and most teams are discovering that the hard part isn’t the model; it’s everything around it. Orchestrating your voice AI pipeline — from the time your user speaks, the words are transcribed, an LLM decides how to respond, and the reply comes back as natural speech — connectivity, silicon: up to five vendors, five integration projects, and a product that still degrades the moment the network does.

Agora’s Physical AI solution offers the industry’s only end-to-end, real-time conversational AI and AIoT ecosystem:

3B+ devices already run on Agora, powering use cases from educational toys and AI companions to elderly care devices. Explore how we can bring your devices to life.

In this episode of the IoT For All Podcast, Wienke Giezeman, CEO and co-founder of The Things Industries, joins Ryan Chacon to discuss how IoT is finally delivering what it promised ten years ago. The conversation covers what changed technically and commercially, the ROI of IoT, why deployments failed in the early days, criticism of IoT, what companies still get wrong about LoRaWAN, and The Things Conference 2026.

YouTube • Spotify • Apple • YT Music • Amazon