Here is a number that should change how you think about building intelligence, and it has nothing to do with a model.
A team at Texas Tech University, led by an energy engineer named Brandon McCoy, spent its first year running a real HVAC sequence-optimization program. They went looking for energy savings and they found them: nearly a million dollars. The opportunities were real, measured, sitting right there in the building data.
They captured $97,000 of it. Less than ten percent.
The other nine hundred thousand dollars did not disappear because the analysis was wrong, or because the sensors lied, or because they needed a smarter algorithm. It disappeared because, in the words of the recap I read this week, "the optimization opportunities were all there, but the hands to execute them were not." They could see the money. They could not reach it.
I have been building monitoring systems for small buildings for long enough to know that this is the most important sentence in the entire physical-AI conversation, and almost nobody says it out loud.
Detection is the cheap part now
Everything in this field is pushing the cost of finding the problem toward zero. This week alone: Mistral shipped a robotics model that navigates a room from a single camera and a plain-language prompt. Luxonis raised $14 million to make edge cameras cheaper. NVIDIA keeps dropping perception models that run on a box the size of a paperback. A $50 vibration sensor and a small model can now tell you a bearing is going to fail three weeks before it does.
We are very, very good at detection. We are drowning in it. And the Texas Tech number is what happens when detection outruns everything else: you generate a million dollars of findings and you have the hands to act on ninety-seven thousand.
Why the big portfolio loses the money
Think about why a sophisticated team at a research university, with a budget and real expertise, still left ninety percent on the table. It is not incompetence. It is structure.
A large portfolio finds savings faster than any team can execute them. One energy group covers dozens of buildings. Each optimization is a work order, a contractor, a schedule, an approval, a follow-up. The findings pile up in a spreadsheet faster than anyone can burn them down, and the ones at the bottom of the list are worth exactly nothing until someone gets to them. The detection scaled. The hands did not. So the capture rate collapses, and a million becomes ninety-seven thousand.
This is the same failure I watch kill "smart building" pilots. The dashboard lights up. The alerts fire. And then the alert lands in an inbox that belongs to no one in particular, in an organization where the person who could fix the boiler is three tickets and two departments away from the person who got the notification. The signal is perfect. The loop never closes.
A small building is built to close the loop
Now here is the part that sounds backwards until you have lived it. The building most people would call too small to bother with is the one that captures the value the university couldn't.
In a church basement, a three-unit rental, a small machine shop, there is no portfolio, no energy team, no ticket queue. There is one boiler and one person who is responsible for it. When my monitor notices the boiler drifting, it does not file a finding into a backlog. It texts that one person. And that one person, more often than not, is also the person who will go down and fix it, or make the single phone call that gets it fixed, this week.
The find-to-fix gap that swallowed Texas Tech's nine hundred thousand dollars is, in a small building, about the width of a text message. The value is small in absolute terms, a few hundred dollars a month in avoided damage and wasted energy, but the capture rate is enormous, because the distance between the signal and the hands is almost zero.
| Large portfolio | One small building |
|---|---|
| Finds $1M, captures <10% | Finds less, captures nearly all of it |
| Findings pile into a backlog | Finding goes straight to a phone |
| The person alerted can't fix it | The person alerted owns the boiler |
| Detection scaled, hands didn't | Detection and hands are the same size |
So when someone tells me a single building is not a serious market for physical AI, I think about that ninety-seven thousand out of a million. Scale is what breaks the loop. Smallness is what keeps it whole.
What I actually sell, in this light
It would be easy to describe my service as "we watch your pump." That is the detection half, and detection, as we have established, is the cheap and abundant half. It is not the thing worth paying for.
What is worth paying for is the loop. A small detector, built from off-the-shelf sensors and a small model on a local box, learns one machine's normal and watches for the drift that shows up weeks before a breakdown. And then, the part that matters, it tells a specific human who knows the equipment, in plain language, in time to act. Not a dashboard nobody logs into. Not an alert into an anonymous inbox. A message to the person who can walk downstairs.
Anybody can build the part that finds the problem. The business is the part that gets it fixed. In a building small enough that the owner is the maintenance crew, that part is nearly free, which is exactly why it works.
Texas Tech proved, at real scale and with real money, that finding the savings is not the hard part and never was. The hard part is capturing them, and the capture rate is a function of how far the signal has to travel to reach a pair of hands. My whole model is to make that distance as short as physically possible: one asset, one human, one message, closed loop.
Detection is cheap. The closed loop is the product.
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 your machine's normal, and watches vibration, temperature, and current for the drift that shows up weeks before a breakdown. Then it tells a specific person who can actually fix it, in time to matter. Nothing leaves the building. $99 to $199 per month, hardware under $3,000. We do not just find the problem. We close the loop.
See how it worksSources: Nexus Labs, "Owner Signal: Two Things Mature Energy Programs Do Better" (email briefing, July 8 2026), recapping its June NexusCast on HVAC sequence optimization: Brandon McCoy's team at Texas Tech University, in its first year running a sequence-optimization program, found nearly $1M in energy savings and captured $97,000 (under 10%), "the optimization opportunities were all there, but the hands to execute them were not"; Nexus Labs' five-level program maturity model. Supply-side context: Mistral "Robostral Navigate" single-camera navigation model (Bloomberg, July 8 2026); Luxonis $14M Series A for OAK edge cameras (Weekly Robotics #367, July 6 2026). Field deployments at The Intersecto Watertown sump-pump site and Northampton 40-device building. Companion brief: /Users/tdeshane/lobster/research/physical-ai-brief-2026-07-09.md.