On July 29, Procore agreed to buy DroneDeploy for $845 million in cash.
The obvious read is that a construction software company bought a drone company. That read is wrong in a way that matters if you run sensors in a building, so it's worth slowing down on.
DroneDeploy is a reality capture platform. It ingests imagery from drones, ground robots, phones, fixed cameras, and helmet cams, and turns it into three-dimensional documentation of a jobsite. It has been used on more than three million sites. The visual record it has accumulated covers roughly twenty trillion square feet.
Twenty trillion square feet for $845 million works out to about four one-thousandths of a cent per square foot. That number is so small it tells you something: the square footage isn't what got bought.
What $845 Million Actually Buys
Two other figures appear in the announcement, and they are much smaller and much more expensive.
Tens of millions of user-generated annotations. And more than one hundred thousand labeled safety issues.
An annotation is a person looking at a picture of a real place and writing down what is wrong with it. A labeled safety issue is a person looking at a real place and saying that specific thing, right there, is a hazard, and here is why. That is not capture. That is judgment, attached to a coordinate and a timestamp, one at a time, by someone who was standing there.
Procore is bringing its own pile to the merge: roughly 400 million photos, over 126 million drawings, and over 10 million RFIs, submittals, and inspections in the last year alone. Their stated plan is to train on the combined set so the platform can spot a problem in site imagery and kick off a response automatically.
To do that, they need images where somebody already said what the problem was. The drone is how the pixels arrived. The pixels are cheap. A capable drone costs a few hundred dollars, and half of DroneDeploy's capture modalities are a phone in someone's pocket. The labels are the part nobody can download.
The Same Week, the Other Half Got Cheaper
Here is the part that makes the timing feel less like a coincidence.
The day after the Procore announcement, Google DeepMind shipped Gemini Robotics 2. Three models: a vision-language-action model for whole-body humanoid control, an embodied reasoning model for planning, and Gemini Robotics On-Device 2, which runs locally on the robot with no internet connection at all.
The claim on that last one is the one to write down. It adapts to a brand-new two-armed robot body in a few hours, typically with fewer than two hundred examples, even when the new robot has a drastically different shape, different sensors, and a different number of joints.
Two hundred examples. A few hours. On the device, offline.
And the same week, Import AI documented something with a similar shape. Anthropic's Claude Opus 4.1, in August 2025, failed autonomous robot tasks outright. Opus 4.7, in May 2026, completed nearly all of them in under ten minutes. The writeup is blunt about why that happened:
"This progress is not the result of a concerted effort to improve the robotics capabilities of our models."
Nine months, no robotics program, and the capability showed up anyway as a side effect of general scaling. Sunday Robotics got to the same place independently with a model that folds garments at 99.1% success across garment types, using pretraining scale plus a small amount of their own data.
Put the week together. The ability to look at a physical thing and know what to do about it is depreciating on a nine-month clock that nobody is steering. The record of what specific physical places looked like, and what was actually wrong with them, sold for $845 million in cash.
I've Been Collecting the Cheap Half
I monitor a sump pump in a basement, and I have telemetry on it going back to 2024. Voltage, current draw, cycle count, cycle duration, motor temperature, water level, outdoor temperature, rainfall. Every reading timestamped, nothing dropped. It's a clean dataset and I'm reasonably proud of it.
It is also, by the standard the market just set, the worthless half.
What makes that data worth anything is a much smaller set of moments. The float switch that started sticking intermittently in March 2025 and took me four days to characterize. The night the cycle duration stretched 12% and it turned out to be the outdoor temperature crossing 28°F, not a failing motor. The 3 AM alert I got out of bed for that turned out to be nothing at all.
Three real labels. Somewhere in the low dozens if I'm honest about how many I could reconstruct across the pump and the forty-device building I also watch.
And here's the uncomfortable part: I know those three because I happened to remember them. Two of them exist as a sentence in a note I wrote afterward. Not one of them is attached to the timestamp in the telemetry where the anomaly actually appears. If I handed my dataset to somebody else tomorrow, they'd get eighteen months of very tidy numbers and no idea which minutes mattered.
The pump data is capture. My memory of the float switch is a label sitting in the wrong place.
What Two Hundred Examples Means for a Boiler Room
Follow the On-Device 2 number forward eighteen months, which is roughly how long it takes edge hardware to absorb something the labs demonstrate today.
A model that adapts to a novel physical system from under two hundred examples, running locally on the sensor, offline, is going to show up in a boiler room. Not as a research project. As a firmware update on a $50 vibration sensor.
When it does, the operator who can hand it two hundred labeled examples of their pump behaving badly gets a working site-specific model on a Tuesday afternoon. The operator with three years of immaculate unlabeled telemetry gets a very well-informed shrug.
That's the whole bet, and it inverts what most people assume they're accumulating. The scarce input isn't years of data. It's a couple hundred moments where somebody wrote down what was actually going on.
Three Things I'm Changing This Week
None of this is expensive. That's rather the point.
1. One append-only events file, sitting next to the telemetry
Seven columns: timestamp, system, what I observed, what I concluded, what I did, what happened after, and how confident I was. A CSV. It does not need to be a database and it definitely does not need a UI. The requirement is that the timestamp lines up with the telemetry, so the label points at the exact readings it explains.
2. Backfill everything I can still reconstruct
Notes, texts, git history, invoices from the plumber, memory. Every service call, every alert, every anomaly I chased. I'll get maybe forty rows across both sites and some of them will have honest confidence values of "fairly sure." Forty labeled events is 20% of the way to the number DeepMind just published as sufficient to teach a model a physical system it has never seen.
3. Log the non-events too, especially those
This is the one I've been getting wrong. Every time an alert fires and it turns out to be nothing, that's a labeled negative, and it's the class I have almost none of. A model trained only on things that broke will tell you everything is breaking. The 3 AM alarm that was just a heavy rain is worth exactly as much as the failure, and it costs one line to record while you're standing there anyway.
The Cheap Part Is Getting Cheaper
Every trend in this business is pushing capture toward free. Sensors that cost $600 three years ago cost $50. Cameras are in everyone's pocket. Models that couldn't see last summer can see now, without anyone having tried to make them. In eighteen months the anomaly detection I currently run on a local server will run on the sensor itself.
All of that is genuinely good, and none of it is a moat. It arrives for everyone on the same schedule.
What doesn't arrive for everyone is a record of your building: what went wrong, when, what it looked like in the data beforehand, what fixed it, and which of the scary-looking nights turned out to be nothing. Nobody can generate that for you. No model release changes how much of it you have. It accumulates one honest line at a time, and it is the only part of this that gets more valuable while everything else gets cheaper.
Procore didn't buy a drone company. They bought a hundred thousand times somebody looked at a real place and wrote down what was wrong.
You can start doing that today, in a text file, for free.
Monitoring that leaves you with a record, not just a dashboard
If you run a facility, the monitoring is the easy part. The asset is the labeled history of what actually went wrong in your building, and it only exists if somebody is writing it down. That's a deliverable, and you should own it.
See how it works