The Machine That Builds Itself
Every industrial revolution paired a new machine with the collapse of a cost.
In 1769, the steam engine collapsed the cost of muscle. Work moved from farms, mines, and cottage workshops into factories.
In 1869, the railroad collapsed the cost of distance. Goods moved from local markets to a continental one, and a coast-to-coast journey of months became a week.
In 1913, the assembly line collapsed the cost of scale. Production moved from craftsmen building one at a time to lines building millions.
In 1947, the transistor collapsed the cost of calculation. Information work moved from paper and people into software.
In this decade, the next industrial revolution will collapse the cost of dexterity and judgment in the physical world. Manipulation will move from manual to autonomous.
Machines will move and control themselves to execute any task, instead of traditional automation where a single machine repeats one exact motion millions of times. Autonomous machines will manipulate, assemble, and build products across any variability.
Pair hundreds of them and you get autonomous factories that will usher in a new era of abundance. But how do we get there?
The bottleneck is data, not intelligence
LLMs proved that scaled compute plus data produces general capability. Foundation models inherit the same paradigm but lack the fuel. There is no internet-size dataset of robot actions for models to pre-train on. Compute, hardware cost, and model architecture are all improving faster than the data supply.
So the defining question of physical AI is not "who has the best model?" but "who has the machine that produces the necessary data?"
Data comes from deployments, not labs
Staged data collection and simulation get you to a demo, not to reliability. Simulation accelerates foundation model development and powers evals, but for contact-rich manipulation, it doesn't close the gap.
The highest quality data is a byproduct of real operations: real parts, real orientations, real failures, real recovery. Data collection capacity is downstream of deployment capacity.
You cannot buy your way to the dataset. You earn your way in by running high quality operations someone pays for.
Deployments succeed first in the middle band
Truly low-variability environments already belong to hard automation, and have for decades. Truly open environments — homes and the humanoids aimed at them — have a data variability problem too large to close on any near-term horizon.
Foundation models win in the in-between: high mix, low volume. The ideal environments are where there is enough SKU, orientation, and task variability that fixed automation fails, and enough structure that autonomy rates become meaningful with achievable data volumes.
And "meaningful" has a concrete definition: autonomy matters when one operator supervising N cells beats the economics of N line workers.
The flywheel: deployment position compounds
The bridge to autonomy is human-in-the-loop data collection: teleoperation, leader-follower, and UMI, generating training data from day one while the operation earns its keep.
From there, the flywheel spins: deploy → generate operational data as exhaust → improve autonomy rates → improve unit economics → win more deployments.
The moat is not the dataset snapshot. It's the deployment surface generating data continuously. A foundation model cannot collect a facility's edge cases without being in the facility. Process knowledge, customer relationships, and operational muscle compound alongside the data.
This is the path to autonomy.
The endpoint: the general autonomous factory
Assembly work decomposes into a shared set of task primitives that repeat across nearly every electromechanical product. This is why deployment data compounds: each deployment buys down autonomy on primitives, not just on the product in front of the robot. Capabilities compound along this path, transferring from one device to the next.
Inspection comes with it, and may arrive first: quality checking is perception-native, exactly where learned models are strongest.
The convergence point is a general autonomous factory that can assemble nearly any electromechanical device. High mix, low volume is not a niche to hide in. It's the proof of generality.
To the obvious objection, "why hasn't traditional automation done this already?": classical automation solves a task for one part in one fixture. Learned models attack all primitives across any part, orientation, or context.
Autonomy makes onshoring viable where it matters
Offshore manufacturing wins by devaluing labor: cheaper hands, not better work. Autonomy removes that arbitrage: not by removing people, but by removing the need to buy the cheapest ones.
What remains is lead time, IP risk, and supply chain security, with tariffs in the mix. For critical verticals such as defense, aerospace, medical, and energy infrastructure, those dominate.
In America, the operator doesn't disappear. One person supervises N workcells instead of running one station. We expect autonomy to create tens of thousands of these jobs across millions of deployed robots: fewer people per unit, more people overall, doing work that rewards judgment instead of repetition.
Autonomy onshores the layers of manufacturing where labor costs drove offshoring.
Already-commoditized and verticalized categories will stay offshore, and that's fine. The point is not winning everything. It's that the supply chains America must own become buildable at home.
The autonomous factory is the machine of the next industrial revolution
In 1913, the assembly line took the Model T from over 12 hours of build time to about 90 minutes, and its price from $850 to $260. That's what a collapsed cost looks like.
But every machine in the pattern — the engine, the railroad, the line, the transistor — was built once and then operated. The autonomous factory is the first machine that improves through its own operation: producing goods and the data to get better in the same motion.
The computer automated information work and left physical work rigid: one part, one fixture, one program. Physical AI finishes the job.
Abundance is downstream of the machines that build it.