The industry meant to deliver it is a $12T machine that stopped moving.
The market keeps growing: $12T today, $15T by 2030, and the wave that leads to 2050 is larger still. But the productivity behind it is frozen. Over the last 20 years, manufacturing productivity grew 90%. Construction grew barely 10%. In advanced economies, a builder produces less today than 50 years ago.
And the hands are leaving.
Construction employs more people than any other industry on earth, and it is losing them faster than any other. Across the developed world, nearly half the skilled workforce will retire within a decade. The young are not taking their place, fewer and fewer choose hard physical work. This shortage is not a cycle to wait out. It is demographic, and it is permanent. We have spoken with more than 100 construction companies, and every one of them tells us the same story: today, hiring is hard. Tomorrow, it will be impossible.
We have to build more than ever, with fewer hands than ever
The obvious answer is automation. Except, until today, you could not.
The tools have barely changed in a century. Software barely dented this industry, the machines never came at all. The work stayed where it has always been: in human hands.
No one missed a turn. There was no turn to take. Automation and robots conquered factories because a factory is made for them: the same parts, the same sequence, the same controlled environment, a thousand times over. Construction is the opposite. Every project is one of a kind. The parts never quite match the drawing. The plan changes when a delivery is late. The ground is mud one week and concrete the next.
Yesterday, nothing repeated, so nothing could be automated.
Today, nothing has to.
What is changing now has a name: physical AI.
For the first time, machines can perceive the physical world the way software already perceives language and images; they see the scene, recognize the objects, decide what to do, and control their own body to do it, in real time, without being told every move. Robots that did not exist five years ago now walk, grip, and lift, and their cost falls every year. In warehouses, they already reach into bins of unknown objects, pick the right one, and place it, thousands of different products a day, no motion ever programmed. Not because someone wrote better scripts. Because no one wrote a script at all, these machines learned.
So here is our vision for the industry. Two fronts.
1/ The offsite becomes a long tail of distributed, autonomous factories feeding the sites around them.
Prefabrication was the first brick to industrialization. It moved work under a roof and mechanized it, but mechanized is not automated. Behind those walls, the processes are still shaped by hands, the orders still custom, the geometry still shifting. That is the step prefab never took, and the step physical AI makes possible: factories that finally match the variety of the work.
2/ The site stays. The work changes.
It keeps what cannot leave: the countless trades that turn a shell into a building.
The workers who remain will not be replaced but amplified. When hands are scarce, each pair must carry further, one pair of hands, directing the machines, 200x the output.
And robots will absorb the physical work itself. They are becoming capable of it, and the economics hold: a robot on a site does not compete with the worker a builder employs. It competes with the worker no builder can find. The day it earns its place, builders will rent it the way they rent a crane.
The first robots are already earning theirs, one prints the layout on the floor, another lays bricks, another hauls the load. They are proving robots belong on a jobsite, and the market is moving fast behind them. But raising a building takes dozens of trades and thousands of tasks. Specialized robots will keep winning the tasks they were built for. The long tail, everything else a builder's day is made of, needs something more general. The future is an equilibrium of both, and the race is on
Millions of robot apprentices are coming. The race to train them starts now.
All modern AI is trained the same way: a model learns a task from recorded examples of that task. Language models learned to write because the internet held billions of pages of text. Cars learned to drive because fleets filmed billions of miles of road.
Guess what a robot needs to learn how to build.
A library of construction.
For as long as humans have built, the knowledge has moved the only way it ever could: hand to hand, master to apprentice. Japan's Ise Shrine has stood for 1,300 years by never standing still: every 20 years, the whole complex is torn down and rebuilt from scratch, on purpose, so the carpenters' hands never forget. Left alone, even the oldest craft fades in a generation. None of it was ever written down. None of it was ever filmed. No one has ever tried to capture an entire industry. Construction is the largest undigitized body of skill on Earth.
The training floors already exist, millions of them, active at this very moment, each one different from the last. Thousands of high-dimension tasks, from snapping a chalk line to tying rebar overhead, from plumbing a door frame to troweling a slab before it sets, each done millions of different ways by millions of different hands. Moving on ground that changes every day. Holding tools shaped for human hands. Metering force: a tap for the tile, a blow for the stake.
It has never been recorded. So we start by recording it.
We put cameras where the knowledge lives: on the workers themselves, every site, every prefab floor, at scale. Egocentric first, because the work happens through the eyes. Then touch, because building is contact. Then every dimension the models need.
With that data, we train. We partner with frontier labs and robot-makers to turn it into models that finally know how to build.
But watching makes only half an apprentice. The other half is doing.
Once the models are ready, we deploy: machines at work inside real operations and everything they learn there flows back into research that compounds. The machines improve, the next deployment comes faster, until millions of machines are building the world of 2050.
Brick by brick, humans and machines, building the world of 2050.