The next bottleneck is beyond the screen.
Intelligence consumes power. Experiments consume materials. Output comes from factories. Cheap cognition may increase the value of physical execution.
Original X thread: 2026.09.22 · Website adaptation: 2026.10.04
Contents & sources
AI does not escape atoms. It bids for them. Better models lead to questions about power, transmission, factories, robots, laboratories and logistics. Follow the route from software performance to physical production.
Where does intelligence go?
If capital flowed into screen-based services, our next direction is intelligence flowing into energy, machines, experiments and production. This is a direction to investigate, not an inevitable forecast. Measure physical output.
Capital → compute → intelligence → experiments → machines → output → revenue → more capital. Physical execution adds distinct costs and failure points to the cognitive loop.
Discovery is not deployment
AI can search a design space. Laboratories must synthesize materials, factories must achieve yield, and customers must pay. Permitting, reliability tests, supply chains and capital expenditure do not move at token speed.
A robot that succeeds in simulation needs separate validation in a real environment. Atoms punish errors.
Which complement is scarce?
Electricity, grid access, land, robots, validated data, laboratories, manufacturing yield and distribution can complement intelligence. Ask which is the bottleneck, who controls it, and whether that owner has pricing power.
A bottleneck does not make every related company a good investment. Check what the price already assumes, whether supply can expand, and whether AI can increase the capacity of existing infrastructure.
Physical metrics over demo quality
Watch watts, real-world robotic success, cycle time, manufacturing yield and return on capital. Faster discovery may coexist with slow commercialisation.
If AI improves utilisation enough to reduce scarcity, the expected rents of bottleneck owners can weaken. Track both the constraint and the technology that can relieve it.
As cognition gets cheaper, ask again what it costs to change the physical world.
An editorial adaptation of Kamui’s original X thread, retaining its argument and countertests. It is not a recommendation to buy or sell a particular asset.