Why we build infrastructure companies

Over the last fifty years, factories got machines, software and supply chains built around them. Output per hour worked in American manufacturing more than doubled over that time, by the measures of the Bureau of Labor Statistics.

Construction went the other way. Austan Goolsbee and Chad Syverson (2023) find that value added per worker in U.S. construction has fallen over most of the period since 1970.

The crews that put up and repair buildings, and the systems that bring those buildings power, water and a network, are a large part of the physical economy. Most of that work is still planned on paper, priced from memory and checked by eye.

We think this is the largest body of work left for software and machines to change. It is also the work that decides whether a family has heat, water and a working phone tonight.

Automation arrived first where work was concentrated. A car plant repeats the same motion thousands of times a day under one roof, so one owner can pay for a machine and keep it busy.

Work on buildings is spread out. Each job is at a different address, on a house built in a different decade, with problems nobody can see until a wall is open.

The firms that do this work are small. A plumbing company with twelve trucks cannot build its own dispatch software, its own sensors and its own training. A vendor selling to that company cannot sell enough of any one product to make it good.

There was also no data. A factory records every part it makes. A house records nothing about itself, so software had nothing to learn from.

Three things have changed, and together they make a different kind of company possible.

Sensors and radios are cheap. A device that measures flow, pressure, current and temperature, carries several radios and runs for years can now be built from parts that are inexpensive and widely available. Each house can report on itself.

Models can read the physical world. A current vision-language model can watch a phone video of a utility closet and pick out the water heater, the shutoff valve and the corrosion on a fitting. Scoping, inspection and training become software problems.

The people who know the work are retiring. The generation that learned these trades decades ago is leaving them, and much of what it knows was never written down. If that knowledge is going to be kept, it has to be captured while the people who have it are still on the job.

A company can now install its own hardware, read its own data and train its own crews from what both of them record.

We work in five physical systems. Build covers buildings, repairs, renovations, field labor and construction robotics. Energy covers microgrids, batteries, distributed generation and physical power. Water covers capture, purification, recycling, storage and distribution.

Connectivity covers mesh communications and resilient telecom for when conventional networks fail. Compute covers edge computing, physical AI infrastructure and machine coordination.

Each domain's scope is the mission. Each one starts with something small that exists or is in development today, and the figure says which parts come later.

The house is the most common piece of infrastructure in the country, and the smallest place where all five domains meet. It is a building with its own power, water and network, and soon its own computers to coordinate them.

The power company, the water utility and the carriers run their networks to the meter and stop there. From the meter to the outlet, the tap and the router, the owner is responsible, usually without tools, without data and without anyone to call until something breaks.

Those systems are old. The median owner-occupied home in the United States was built around 1980, according to the Census Bureau's American Housing Survey. Its water heater, cooling, wiring and plumbing get replaced one emergency at a time.

Failures are expensive and mostly invisible. The EPA estimates that household leaks waste close to a trillion gallons of water a year across the country.

So the house is where the need is largest and where no utility is responsible. One visit can install devices for three of its systems at once. It is also where 4 Seasons field crews already work in metro Atlanta.

After the house, each domain grows outward to the streets, campuses and cities that run on the same five systems.

Good devices for parts of this already exist. Leak shutoff valves such as Moen Flo and Phyn close the main when they detect a leak. Panel controllers such as SPAN switch circuits at the electrical panel. Failover routers move a home's internet to a cell network when the main line drops.

Each of them is sold on its own, installed by someone else and read in its own app.

We build our own so that one crew installs all three in one visit, one record holds their readings, and a utility or an insurer can settle on a number we measured.

Where a device on the market does the job, we will compare ours against it in real houses and publish the result.

We start where our crews are. The open map data for the 64 km square around Atlanta holds 1,002,904 buildings, 1,340 power lines from 12 to 500 kV, 328 substations, 1,030 cell and communication towers, and 343 water towers, tanks and reservoirs.

That is enough houses to learn from, a grid with time-of-use rates to work with, and storms often enough to test whether a house stays powered and connected through them. A metro this size is also small enough for one dispatch system and one warehouse to cover.

Each domain needs its own hardware, its own regulations and its own partners. A utility, an insurer and a carrier each want a company that speaks its language. So we build one company per domain, each with its own name, product and team.

Underneath, the companies share everything that is expensive to build twice: the field crews, the software, the contract manufacturer, the supply chain and the record of each site. Groundwork, the Compute company, builds the shared software.

That is why one studio builds and holds them. A domain's next company starts with the crews, software, parts and records that the first ones paid for. Each new company should cost less to start than the one before it.

Every job and every device produces a record of how houses work and fail. A crew opens a wall and photographs what is behind it. A Flow Node will record pressure and flow for years. A Current Hub will record which loads ran and when.

Across thousands of houses, those records show which fittings fail, which installs run long, which repairs come back and which readings come before a leak. They make the next price more accurate, the next procedure better and the next visit earlier.

No contractor, device maker or utility sees all of this today. We intend to, because our companies do the work, own the devices and keep the record.

We grow in steps, and each step has a gate. A gate is a measured result we have to reach before the next step starts.

4 Seasons comes first because its crews and software already run. Devices come next, one at a time, then the single visit that installs all three. The later systems in each domain follow their first device, and a second city comes only after Atlanta works.

We will not promise a result before we measure it. Savings, leak rates and install times on our sites are labeled planned or goal until we have field data. We will publish the method behind every number we report.

We will not pay crews for hours on site. 4 Seasons pays for each step that is checked and done, and every job leaves a record the customer can see.

We will not build companies to sell. We hold them for decades, because houses last decades and a record of a house becomes useful over that span.

We will not sell the record of a house to advertisers. It belongs to the house, and the owner decides who sees it.

We will not hide what we have not solved. Our open problems are published, with what we know so far and what we do not.

In ten years, a house in metro Atlanta could have its systems measured, maintained and repaired by one set of crews. Failures would be caught before the owner notices, and prices set before anyone drives over.

Then each domain would grow past its first device. Current would add batteries and local generation, so buildings on one street could share power as a microgrid. Flow would add rainwater capture, purification and recycling. Relay's mesh would carry a neighborhood's messages when the carriers are down. Groundwork would move computing onto the devices and coordinate machines and crews in the field. 4 Seasons would add robots for the most repetitive construction steps. None of these is built today.

After that, the same crews, software and devices go to a second metro, once Atlanta meets its gates.

We expect this to take decades, and we are hiring people who want to stay that long.