Private Equity Is Buying My Customers

I spent this week reading two things that have nothing to do with each other, and they turned out to be the same story.

The first was a chip announcement. On August 3rd, Microchip introduced a reference design called EnviroSense AI. It runs environmental classification on a microcontroller — a Cortex-M7, three ordinary sensors for temperature, humidity, and ambient light — and sorts conditions into five categories. Sunny. Cloudy. Humid. Rainy. Cool indoor. It also tells morning from noon from evening from night. No neural processing unit. No cloud. It is a downloadable example project.

The second was a set of deal trackers for HVAC service company acquisitions.

Here is why they belong in the same paragraph. The first one says the thing I sell is becoming free. The second one says something adjacent to it is getting more expensive by the month. I have been building on the wrong side of that line.

The Number That Reorganized My Week

In 2023, private equity accounted for roughly eight percent of HVAC service acquisitions.

Through early June of this year, one tracker counted 77 HVAC transactions. Thirty-nine of them were private-equity-led. That is 50.6 percent.

Eight percent to more than half, in three years. A separate tracker, using different inclusion rules, counted 149 HVAC services transactions year to date, up 12.9 percent over the prior year, which works out to roughly one deal every two to three days. Those two counts are not comparable to each other and I am not going to pretend they are. The share is the number that matters, and the share moved from a rounding error to a majority.

If you own a mechanical contracting business, you already know this, because your phone has been ringing.

But the part that actually reorganized my thinking was not the volume. It was the price list.

Same trade. Same trucks. Same technicians. Roughly double the price, on the basis of how the revenue arrives.

That is not an opinion about the market. It is what buyers actually paid, across dozens of closed transactions, in a sector where deals are closing every couple of days. And it is not a subtle preference buried in a spreadsheet somewhere — the published buyer checklists put recurring revenue at the top, followed by technician retention that survives the founder leaving, geographic density for dispatch efficiency, and then, in plain language, technology stack. ServiceTitan and BuildOps get named out loud.

I Have Been Pitching the Wrong Person

For a year I have walked into small buildings and explained to owners that a fifteen-dollar smart plug watching a sump pump's current draw will tell them the motor is dying before the basement floods. It is true. I have the logs. It has been running for two years without me touching it.

The problem with that pitch is not that it's wrong. It's that I am competing against doing nothing, and doing nothing is free.

An owner weighs a hundred and fifty dollars a month against a flood that has not happened yet. Some of them say yes. Most of them say "let me think about it," which is a polite no with a calendar attached. I have gotten pretty good at that conversation and it is still, structurally, an uphill one. I am asking someone to spend real money to avoid a hypothetical.

The contractor is a completely different conversation, and I did not see it until this week.

For a contractor, a monitoring agreement is not a cost. It is the cheapest available machine for turning a break-fix customer into a recurring-revenue customer. And recurring revenue, on his P&L, is worth roughly double per dollar when somebody comes to buy him.

I don't have to convince him that monitoring prevents floods. He has been pulling dead pumps out of basements for twenty years; he believes me before I open my mouth. What I have to show him is that the monitoring is what makes the customer sign an annual agreement instead of calling whoever answers first at 2 a.m.

The owner buys peace of mind, which is a soft thing to price. The contractor buys a revenue line with a multiple attached to it. One of those two people has a spreadsheet where my product shows up as an asset.

Why the Chip Announcement Belongs in This Post

Because it tells you what isn't the asset.

Three years ago, "we run inference on the device instead of shipping your data to a server" was a product claim you could build a company on. This week it is a Microchip reference design running on a dev board with a temperature sensor and a light sensor stuck to it. Earlier this year Siemens put condition monitoring inside the sensor itself. Last week I wrote about a group at Queen Mary that hit 100-micrometer tactile resolution by deleting the model entirely and letting a color-changing polymer do the sensing.

The intelligence layer is deflating toward zero, and it is deflating faster than I can build on it. Any part of my offering whose value rests on "we do the math on the data" is standing on a floor that is dropping.

There is a second problem with the model layer, and it's worse than the price. A piece I read this week made an argument I have not seen stated plainly by anyone selling this stuff: a deployed edge model performs best on its first day and gets worse from there. Cameras face different light as the seasons turn. Machines wear, and their vibration signatures wander away from the training data. People use the system differently than whoever collected the samples assumed they would.

And the fix is a trap. To refresh the model you either ship field data back to a central server — which deletes the privacy and latency story that justified putting it at the edge in the first place — or you accept a slow, silent decline. Trying to learn on the device runs into catastrophic forgetting, which is a research problem with venture money behind it and no shipping answer.

Nobody publishes a decay rate. Not the vendors, and, I'll say it plainly, not me either. I have two years of history on the sump pump and a smart building deployment with real logs, and I have never gone back and compared how the first ninety days behaved against the last ninety on the same equipment. I'm going to. If it held steady, that is a credential nobody in small building monitoring is currently offering. If it drifted, that is the more useful post.

What Actually Appreciates

Last week I wrote about Kilroy Realty running about four thousand water sensors across eighty buildings. The metric they publish is not detection rate. It's whether the engineer closed the work order. The sensor was assumed. The paper trail was the product.

I took that as an owner-side lesson about documentation. This week the M&A market said the identical thing in dollars, from the other end of the transaction.

What a buyer pays 8–12x for is a signed, transferable, recurring relationship — and a technology stack that makes the relationship stick after the founder walks out the door. Not a clever algorithm. A contract with history behind it.

Which is, conveniently, the only durable thing I actually make. The classifier is a free download. The two-year log with a customer's name on it is not, and cannot be, because it is made of time.

What I'm Doing About It

Two things this week, and I'll report back on both.

First, I'm rewriting the pitch for contractors and taking it to three commercial mechanical shops. Commercial is the early-stage side of this consolidation — residential is already mid-cycle, but the commercial pool is still largely unconsolidated, which means those owners are getting the calls right now and have the multiple in their heads. The offer is one sentence: I put your existing break-fix customers on recurring agreements, I supply the monitoring and the documentation that justifies the agreement, and you own the customer and the revenue. Lead with the multiple, not the sensor.

Second, I'm running the decay comparison on my own data. First ninety days against the most recent ninety, same site, same thresholds, both deployments. The question isn't only whether anything degraded. It's whether I would have noticed if it had.

I've spent a year getting better at explaining what a sensor can see. The more useful year is probably the one I spend getting better at explaining what a two-year record is worth to the person who owns the customer.