The most interesting AI news this week was not a bigger model. It was a bunch of smaller ones, and an argument about where they should live. Apple is building hybrid local/cloud models with Google, splitting the work between something that runs on the device and something that runs in the cloud. Google shipped Gemma 4 12B, a multimodal model small enough to call "laptop-sized." Apple put out research on rethinking mixture-of-experts specifically to save local memory. Even the new Claude release got framed as capability "for the masses."
Strip the product names away and there is one question underneath all of it: what should run on-device, and what should run in the cloud? The frontier is treating that as a hard, open design problem, and pouring real engineering into the answer. Which is funny, because a sump pump in a basement settled the same question a long time ago, and the answer turns out to be the whole business model.
The split a building already runs
A fixed-asset monitoring system has exactly two kinds of work, and they want to live in two different places.
The first kind is the watching. A pump, an air handler, a compressor, a circulating loop, each one gets observed every minute of every day. The detector reads the asset's measured behavior, compares it against what that asset normally does, and asks one question on a loop: is today still matching yesterday? This work is constant, it is latency-sensitive, and it touches data the owner would rather not stream off the property. It has to be cheap, private, and always on. That is the textbook definition of what belongs on the edge, and it is where it runs: a small box on the wall, doing local inference, with no per-token bill and nothing leaving the building.
The second kind is the explaining. When the detector flags a drift, the pump cycle running long, a schedule that quietly slipped, a current draw creeping up, something has to turn that signal into a sentence a human will act on. Summarize what changed, put it in context, draft the note that goes to the owner or the service tech. That work is heavier, but it is also rare. It happens on an exception, not on a loop. And rare, heavy work is exactly what you reserve the cloud for.
This is not a clever trick. It is just where the line naturally falls once you notice that monitoring is mostly silence punctuated by the occasional event. You do not pay a streaming inference bill to confirm, ten thousand times a day, that the pump is fine. You pay almost nothing to watch locally, and you spend a little, on the few occasions it matters, to explain.
Why that line is the margin
I have written before about the cost asymmetry between the open-world frontier and a closed building. The local/cloud split is where that asymmetry shows up on an invoice.
Run the watching in the cloud and you are paying, continuously, to ask a remote model a question whose answer is "still fine" the overwhelming majority of the time. That is a variable cost that scales with every asset and every minute, and it quietly eats a monitoring business alive. Run the watching local and that cost collapses to a one-time hardware purchase, under three thousand dollars, that then runs for years. The cloud only gets touched on a flagged exception, which for a healthy building is a handful of calls a week. Pennies, not a meter running around the clock.
So when a building owner asks the two questions they always ask, "what does this cost to run, and is my data leaving the building," the hybrid split is the answer to both at once. The constant part is local, so it is cheap and it stays on-site. The cloud part is rare, so it never becomes the bill. The same architecture the labs are validating for privacy and latency on a phone is what holds gross margin on a monitoring contract.
The new temptation: an agent that does more than watch
There was a second current in the week's news worth naming. NVIDIA's latest Jetson software push, JetPack 7.2, leans hard into "agentic AI skills" on edge hardware, and the broader drumbeat is to use agents that do work, not just chat. That is genuinely useful, and it is coming to the edge fast.
But for building monitoring, the moment you give an edge agent the ability to act on the asset, change a setpoint, cycle the equipment, drive the pump, you have traded away the two things that made the product safe and cheap to sell. I made this case a few days ago and the agentic-hardware news only sharpens it. The right first job for an edge agent on a building is to draft the alert and the work order, the explaining, not to touch the controls. Keep the agent on the local/cloud line where it belongs: it can help write the sentence, it should not turn the valve.
I did not design this. The work did.
None of this came from a whiteboard. Our sump pump system in a Watertown basement runs its watching locally because the alternative, streaming a pump's heartbeat to the cloud forever, was obviously absurd the first time I priced it. The cloud gets involved only when the pump does something worth a sentence. The 40-device site in Northampton is the same split, forty times over: forty assets watched locally, one shared place to explain the rare event that actually matters.
The frontier labs arrived at the same shape this week from the opposite direction, optimizing models small enough to live on-device and reserving the cloud for the heavy lift. It is a good week when the most advanced AI organizations on earth converge on the architecture your boiler room has been quietly running the whole time. It means the boring, shippable end of physical AI was right about something, and it was right cheaply.
The question worth asking your building
If you are weighing any kind of monitoring or "smart" system for your property, the useful question this week is not which model it uses. It is: what runs on-site, and what leaves the building, and what does the part that leaves cost me every month?
If the answer is "everything runs in the cloud and the bill scales with how much we watch," you are paying frontier prices to confirm a pump is fine. If the answer is "the watching is local and only the rare exception reaches out," you have the architecture Apple and Google just spent a week building, sized for a building instead of a phone, and priced like a boiler room instead of a data center.
Local for the watching. Cloud only for the rare exception. That is the whole pricing story.
Each asset that matters gets its own small detector trained on its own measured history, running on an edge box on the wall, local, with no cloud dependency and no per-token bill for the 24/7 watching. The heavier work, turning a flagged drift into a clear, human-readable alert, happens only on an exception, so it never becomes the monthly bill. It never touches the controls. $99 to $199 per month, hardware under $3,000.
See how it worksSources: DeepLearning.AI Data Points, 2026-06-08 (Apple building hybrid local/cloud models with Google; Gemma 4 12B "laptop-sized" multimodal model) and 2026-06-10 (Apple rethinks mixture-of-experts to save local memory); The Batch, 2026-06-12 (Claude Fable 5 "Mythos for the masses"; encouragement to use agents that do real work). NVIDIA JetPack 7.2 agentic AI skills on Jetson (COMPUTEX 2026). NEURA Robotics raising up to $1.4B Series C for physical AI (The Robot Report); Apptronik $520M / ~$5B valuation, Feb 2026. Predictive maintenance market ~$10.9B (2024) to $70B+ by 2032, CAGR >26% (MarketsandMarkets / Coherent). IoT sensor unit costs below $1 (OxMaint 2026 guide). Field deployments at The Intersecto Watertown sump-pump site and Northampton 40-device building. Companion brief: /Users/tdeshane/lobster/research/physical-ai-brief-2026-06-14.md.