Frontier industrials

The AI story is turning into an industrial story, and most people are still reading it as a software story. That gap is where this decade’s value sits.

Start with the models. Frontier capability is being commoditised from above. Open weights arrive months behind the frontier. The price per token has collapsed. A capability that cost a fortune to run last year runs close to free this year. This is what a maturing technology does. The value stops sitting in the thing itself and moves to what you build on it.

So the question worth asking is where the value goes next. It goes to where intelligence meets the physical economy. NVIDIA calls this physical AI, the translation of trained intelligence off the screen and into machines. The phrase is theirs, and the direction it names is real. Robots, autonomous vehicles, agentic factories, embodied systems. The frontier labs trained intelligence on text. The next decade spends it on atoms.

The capex is already industrial

You can see the rotation in the capital before you see it in the products. The AI build-out reads as a software event and lands as an industrial one.

Data centres pull power. The International Energy Agency puts data-centre electricity use at around 415 terawatt-hours in 2024, about 1.5% of world demand, and projects roughly 945 terawatt-hours by 2030. That later figure is close to the entire electricity consumption of Japan. In the United States, the IEA expects data-centre demand to rise about 130% by 2030.

Power pulls grid and generation. A gigawatt of new load needs a gigawatt of firm supply, wires to carry it, and a connection point. That demand has walked the hyperscalers into the oldest heavy industry there is. Microsoft signed a twenty-year deal to restart Three Mile Island Unit 1. Amazon contracted nuclear output from Talen’s Susquehanna plant. Google backed Kairos Power to build small modular reactors. Every major hyperscaler has now signed at least one nuclear deal. Software companies are buying reactors.

Robots and vehicles pull semiconductors and manufacturing. An internal-combustion car carries about 500 to 600 dollars of chips. An electric vehicle carries two to three times that, per industry estimates, and the number climbs as autonomy and power electronics deepen. A humanoid robot sits further along the same curve: sensors, power modules, compute, actuators, all of it fabricated in plants that take years to build. Embodied AI is a bill of materials before it is a demo.

Read the chain as one motion. Compute demand becomes a power problem, the power problem becomes a grid and nuclear problem, and the machines that spend the intelligence become a semiconductor and manufacturing problem. The AI capex story was never a screen story. It runs through the physical stack from the first dollar.

The open position is the analysis layer

None of this is secret. The phrase physical AI is on NVIDIA keynote slides and a16z has staked adjacent ground under American Dynamism. The thesis is loose in the wild. Anyone can recite it.

The position that is still open sits one level down. Nobody owns the synthesis and analysis layer where frontier technology meets the industrial economy. The herd has the slogan. The read of it is unclaimed, the read that reasons across the whole stack and holds the nine domains together. Speed matters on the position. It never matters on the phrase.

There is a second reason the analysis layer is where the value settles, and it is the part most commentary gets wrong. The big winners from cheap robots will mostly not be robot makers.

That pattern is old. The winners of cheap electric motors were not the motor makers. The motor stopped being a product and became a component. The value moved to whoever put it to work: the factory that automated a line, the appliance maker, the lift company. Compute followed the same path. The cloud commoditised raw computing and the value moved to the firms that deployed it into a workflow. Cheap robots will follow the motor and the server, and their price falls the same way.

The mechanism is Wright’s law. For technologies made in high volume with short cycle times, cost falls a fixed percentage with every doubling of cumulative production. Solar modules have fallen about 20% per doubling for decades. Lithium-ion cells went from over 1,000 dollars per kilowatt-hour in 2010 to under 150 today, a learning rate near 18%. Robots that ship in volume will ride the same curve down. As the machine gets cheap, the money stops sitting in the machine.

Where deployment actually lands

So the winners are the deployers, and deployment happens inside ordinary industrial businesses. A mine that runs autonomous haulage. A logistics operator that automates a warehouse. A machine-tool shop that puts a cheap arm on every station. These are not frontier companies. They are the unglamorous middle of the industrial economy, and they are where a falling robot price converts into margin.

This is the part a pure frontier lens misses. Point the whole camera at the robot makers and you photograph the least valuable link in the chain. The value lands in the boring domains, inside businesses that will never appear on a keynote slide. Cover only the frontier and you miss the place the thesis pays off.

Frontier industrials is the name for reading both at once. The frontier that makes the machines, and the industrial base that deploys them. AI is eating the industrial economy, and the decade’s value migrates to where intelligence touches the physical world. That is the bet. The analysis layer that reads it across the whole stack is the position that is still there to take.

The weekly version of this argument runs in Industrial Sector Insights.