Last week a vice president at a chip company wrote down the thing I have been telling building owners for two years, and put it in the trade press where it carries more weight than anything I could say.
The essay, published in The Robot Report on May 23, argued that the future of physical AI is not the humanoid robot. It is "millions of intelligent systems," each one built to do a single defined task, optimized for reliability, safety, and cost. Humanoids, the author wrote, will stay stuck in "niche, high-cost applications" for the foreseeable future, because the thing holding robots back was never the AI. It was hardware complexity, dexterity, and economics.
He listed examples of the machines that are actually scaling right now: robot vacuums, AI-enabled lawn mowers, autonomous drones, warehouse robots. Cheap, single-purpose, already deployed by the million.
I read the list twice. One category was missing.
The Machine Nobody Puts on the List
A building monitor is one of those millions of intelligent systems. It does exactly one job: it watches a physical process and tells you when reality stops matching normal. It is cheap, it is task-specific, and it runs the same edge-AI architecture the chip vendors are now describing as the winning pattern. But it never makes the example list, because a sensor watching a sump pump is less photogenic than a robot folding a shirt.
I have been deploying these systems since 2024. First a sump pump edge AI system in my own building. Then a 40-device deployment across a 40,000 square foot community center, watching temperature, humidity, CO2, power draw, water presence, and door contacts. Neither one looks like a robot. Both of them are exactly the kind of machine the article is describing: a defined task, an edge model, a local control loop.
The most important sentence in the whole essay was the one about that control loop:
"Latency, connectivity gaps, or unpredictable delays cannot be part of a control loop."
I have been making that argument to skeptical building owners for two years, and it always sounded like my personal preference. A chip vendor just made it the spec.
Why "The Cloud Is Optional" Stopped Being My Opinion
When I built the sump pump system, I made a decision that some people found strange. The safety function does not depend on the internet. The sensor checks the water level every 30 seconds on the local network. If the level crosses the threshold, the alert fires from the local network, with no cloud API call in the critical path. The cloud is used for dashboards, history, and remote access. It is never used to decide whether the basement is flooding.
At the time, that was a practical call about 2024 hardware and a homeowner who did not want a flooded basement waiting on an AWS region. It turns out it is also the architecture the entire industry is now converging on. When a perception-chip vendor writes that connectivity gaps "cannot be part of a control loop," he is describing the design I shipped two years ago, for the same reason: a control loop that handles a physical event cannot pause to ask the internet for permission.
The Intelligence Is Moving Into the Controller
The same week, a separate walkthrough of edge AI for industrial systems described where building mechanical systems are headed, and it reads like a description of the sump pump's internals. Boiler systems, the piece explained, are evolving from manual gauge-reading, to cloud dashboards, to "intelligence within the controller itself."
The architecture they described: a small model runs locally on the device. It predicts what the sensor readings should be. A second step compares the prediction against the actual reading. When the gap between predicted and actual grows past a threshold, that gap, the residual, is the anomaly. Multiple sensor streams get fused together in time, so the system reasons about the machine as a whole, not one gauge at a time.
That is not a future roadmap for me. That is how the sump pump already works. The edge model learns what a normal pump cycle looks like, and it does not just classify the cycle, it emits a residual that spikes the moment the world stops matching the model. A float switch that sticks, a pump that runs long, a cycle that fires too often, each one shows up as a gap between predicted and observed. The industry is now describing this predict-then-compare-the-residual pattern as the default for building systems. It has been running in a basement on a Raspberry Pi the whole time.
What This Means If You Operate a Small Building
The takeaway is not that you should wait for smarter hardware. It is that the architecture you should build on is no longer a bet. The whole industry, from the chip vendors up, has settled on the same answer: put the intelligence as close to the physical event as possible, keep the control loop local, and use the cloud for the things that can tolerate a delay.
If you instrument your building on that pattern now, you are not making a wager that might age badly. You are building on the consensus design, two years before most of the products that embody it will reach the shelf. And every day the system runs, it accumulates the one thing you cannot buy or backfill: a baseline of what normal looks like in your specific building, with its specific quirks, its specific pump, its specific drafty corner.
The alert decision happens locally. The cloud handles dashboards and history. This is no longer a preference, it is the design the chip industry says is required.
The valuable signal is not "temperature exceeded X." It is "this reading no longer matches what normal looks like here." Predict normal, then watch the gap.
A model that knows your building is only as good as the history it learned from. You cannot build that history retroactively. You build it by starting.
The List Will Get Longer
The chip vendor's list of task-specific machines that are quietly winning, vacuums, mowers, drones, warehouse bots, is going to keep growing. Building monitors belong on it. They are cheap, they do one job, they run an edge model, and they keep working when the internet does not. They just do not look like robots, so nobody points the camera at them.
I do not need them to point the camera at the basement. The sump pump has been running this architecture for two years, on a board that costs less than dinner. The industry spent this week describing it as the future. From down here, it looks like Tuesday.
Want one of those intelligent systems watching your building?
I design physical AI monitoring for small commercial buildings using open protocols and a local-first control loop, the architecture the chip industry just endorsed. If you want to know what a first deployment looks like for your building, let's talk.
See the Physical AI practice