Thesis · Contract manufacturing
What's next for manufacturing?
This thesis proposes a model of how to build a next generation manufacturer.
The thesis- 01CNC is managed in pools of 3–5 machines, 10 at a stretch, never in fleets.
- 02The bottleneck is workholding, and the machinists who do it are retiring.
- 03Physical AI dissolves the bottleneck.
- 04A focused integrator can build the subcontractor that runs 100–1,000 machines per worker.
AI and robotics are accelerating with exponential progress. Neural networks are going to operate manufacturing facilities.
01 · Pools, not fleets
4machines per worker
The bottleneck is workholding
One coupled problem
Software already writes the toolpaths. What it can’t do yet is hold the part. Workholding sets the orientation options, orientation sets axis access and collision risk, and both decide how datums carry from one setup to the next. It’s the hidden tradecraft that keeps a machinist standing at the vise.
The long tail of parts
Illustrative
Then the upkeep that eats the rest of the shift
The people who hold the part are retiring
The workforce bottleneck
Pools exist because every few machines need a machinist standing by, and that supply is shrinking. Machinists are older than the workforce around them, fewer young people come into the trade, and every opening it has this decade is a replacement for someone leaving. The judgment that goes with them, how to hold a part and where to find its datums, mostly leaves with them too.
Median age
46
for US machinists, against 42 across all workers
Aged 55+
32%
of machinists, against 23% of all workers
Under 25
9%
of machinists, against 13% of all workers
Openings a year
30k
to 2035, all replacing people who leave; none from growth
Where the machinists are by age
Where the hours go
One job, one machine
One job from raw stock to shipped part is eleven stages and 44 physical actions. Judgment concentrates in the middle five: setup, registration, prove-out, the run itself and the flip to op 2. That stretch is what ties a person to a handful of machines.
J Judgment: tacit skill, decisions H Handling: routine hands-on work A Automatable with equipment sold today
Physical AI dissolves it
Forecast, Sep 2026 → Sep 2031
Each of the 44 actions, assessed for how far a robot (humanoid or mobile manipulator, with purpose-built end effectors, sensing and cell software) can take it over. Full means no human touch in normal operation, with exceptions escalated to a remote supervisor.
The pattern
Tasks that tolerate imprecision and retries go first: air-blast chip clearing, bin swaps, pick-and-place into fixed stations. Micron-level force feel, deformable objects and open-ended diagnosis go last. Many hard actions are designed out rather than copied: zero-point pallets remove tramming and vise mounting, probing removes edge-finding, sister tooling removes urgent tool swaps.
The bottlenecks by horizon
Most of the cost of a CNC cell comes from the assumption that a person has to stand at the vise. Autonomous cells have no such requirement, and scale up efficiently.
The call
A focused system integrator
Build a manufacturing subcontractor around this thesis, and take it to fleet-level CNC: 100 to 1,000 machines per worker.
Selling cells to job shops waits on the 2–5 year adoption lag. A contract manufacturer that owns the floor, the parts and the data captures each capability the moment it works, and every job it runs trains the next one. The integrator’s focus is the whole stack around workholding, not a robot bolted to a pool.
Signposts that pull dates earlier
What would push them later
Let’s build the factory floor that runs itself.
Opening conversations with partners, customers and investors.