This week Siemens and Arm laid out how they are doing predictive maintenance now, and for the first half of it I could have written the press release myself.
The intelligence runs on the sensor. Armv9-based AI chips clamp onto factory equipment and run their models right there on the edge device, no cloud round trip. They watch the three signals that matter: vibration, temperature, and energy draw, on motors, conveyors, and actuators. When a bearing starts to overheat, the sensor knows before anyone on the floor does.
That is the exact architecture I have been installing in a church basement in Watertown and a 40-device building in Northampton, and getting the occasional look that says why not just use the cloud like everyone else. This week the largest industrial automation company on earth shipped my answer for me. Intelligence on the sensor, in the building, watching the machine's own signals, is not the frugal option anymore. It is how serious predictive maintenance is done.
So I agree with the whole front half of Siemens's system. Where we part ways is what happens next.
Siemens closed the loop. That's the part to think hard about.
In the Siemens system, the sensor does not just notice the overheating bearing and tell someone. It acts. The system automatically slows the motor, rebalances the load, or kicks on a cooling cycle, with no human in the loop. The intelligence moved onto the sensor, and then the sensor got handed the lever.
For a Siemens factory, that is a reasonable call. There are engineers on site. The process is controlled and instrumented. There is redundant equipment, and the actuators are designed to fail safe. If the model makes a wrong automatic decision, the worst case is usually a paused line and an annoyed shift supervisor who walks over and restarts it.
A small building has none of the factory's safety net. No engineer standing by. No redundant pump. No controlled process. Just one sump pump, one boiler, one compressor, and a lot of hours when the building is empty.
The failure mode inverts
Picture the same closed loop in a basement during a storm. Water is coming in fast, so the pump is working hard, so it is drawing more current than usual. A monitor with its hand on the lever, trained to treat high current as a fault, decides the pump is in trouble and shuts it off to protect it.
It just flooded the basement it was installed to protect. And there is no engineer to walk over and restart it, because the whole reason the monitor exists is that nobody is down there.
This is the thing about autonomous action in a small building: the cost of a wrong decision is not symmetric with the cost in a factory. The same model, making the same mistake, produces an annoyance in one place and a disaster in the other. The intelligence is identical. The consequence of letting it act on its own is not.
The question stopped being "is the model smart enough to act?" and became "who pays when it acts wrong?" In a building with no operator on site, the answer is: the owner, in water damage. So the model doesn't get the lever.
What I build: the front half of Siemens, plus a human on the lever
So the design I sell is deliberately Siemens's front half and deliberately not its back half.
| Part of the system | Siemens factory | Small-building monitor |
|---|---|---|
| Where the intelligence runs | On the sensor, on the edge | On the sensor, on the edge |
| What it watches | Vibration, temperature, current | Vibration, temperature, current |
| Who decides to act | The system, automatically | A person who knows the machine |
| Cost of a wrong action | A paused line | A flooded basement |
The monitor catches the drift three to six weeks early, the way a good edge system does, and then it does one thing: it tells a person. It does not slow the motor. It does not cut the pump. It raises its hand to someone who knows the equipment and can walk down and look.
I used to describe the human-on-top part almost apologetically, like it was the piece I had not automated yet. This week reframed it. "The model watches, a person decides" is not an unfinished feature. In a building with no operator on site, it is the safety property. It is the reason a wrong reading costs you a false alarm and a phone call instead of a wrecked basement.
Why the "boring" version is the one that's already solved
There is a reason the frontier of physical AI is pouring effort into machines that act on their own, and it is not that acting is easy. It is that their problems are open-ended.
The same week as the Siemens news, NVIDIA showed off a system called ENPIRE, where a coding agent wraps a real robot in a self-improvement loop: it runs the robot, reads the failure logs, rewrites its own training code, and tries again, over and over, until the robot can do a brand-new task like sorting pins into a box or cutting a zip tie. It got to 99 percent. It is genuinely impressive. It also took a full software harness, real robots running in parallel, and a frontier model rewriting code, all to teach a robot one skill it had never done before.
Your pump's skill is "keep this basement dry." It does not change. The building hands you fresh ground truth every single day, in the form of a basement that is either dry or not. You do not need a self-improvement loop, because the task is fixed and reality grades your monitor for free every morning. The problem is boring, and boring is exactly why it is already solved. You are not on the expensive frontier of physical AI. You are on the side where the hard part is done and the only real decision left is a design choice: let the model act, or keep a human on the lever.
For a small building, that is not a close call. The intelligence goes on the sensor, in the building, the way Siemens just confirmed it should. The lever stays with a person who knows the machine.
Edge-native monitoring, with a human on the lever.
Each critical asset, your pump, your boiler, your compressor, gets its own small detector built from off-the-shelf sensors, running a small model on a local device. It clamps on in an afternoon, learns that machine's normal, watches vibration, temperature, and current, and catches drift three to six weeks before a failure. Then it does one thing: it tells a person who knows the equipment. It never shuts your pump off on its own. Nothing leaves the building. $99 to $199 per month, hardware under $3,000.
See how it worksSources: Arm Newsroom, "Siemens Reinvents Factory Reliability with Edge AI-Driven Predictive Maintenance" (Armv9-based AI sensors running generative AI models on-device; SIMATIC S7-1500 PLCs, IoT2040 edge devices, MindSphere and Industrial Edge; monitoring vibration, temperature, and energy draw on motors, conveyors, and actuators; automatic corrective action including slowing motors, rebalancing load, and triggering cooling cycles with no human in the loop), July 2026. Import AI 463, "Self-improving robots" (NVIDIA ENPIRE: coding-agent self-improvement loop with Environment, Policy-Improvement, Rollout, and Evolution modules; 99 percent success on dexterous tasks including pin sorting, zip-tie cutting, and GPU insertion; GPT-5.5 in Codex and Opus 4.7 in Claude Code lead), June 29 2026. Cloud-inference latency of 50 to 500 milliseconds and BAS-overlay deployment pattern via July 2026 edge predictive-maintenance market coverage. Field deployments at The Intersecto Watertown sump-pump site and Northampton 40-device building. Companion brief: /Users/tdeshane/lobster/research/physical-ai-brief-2026-07-05.md.