The compute ceiling
Energy, fabrication, and data are the physical walls around SI. Where they actually sit, and what they decide.
Intelligence is not the only input
Reasoning quality now scales with test-time compute, which scales with power and with the physical ability to build accelerators. Those are industrial variables, and industry moves on a different clock than software.
The three walls
Power. A large training campus draws what a small city draws. Grid interconnection is the slowest step in the chain, and it is not solved by hiring more researchers.
Fabrication. Advanced packaging and memory bandwidth are the real constraints, not the headline process node. Capacity here is booked years ahead.
Data. The remaining margin is not more text. It is proprietary process data, sensor data, and the expensive business of generating verified trajectories of correct reasoning.
Why it matters for the timeline
If capability is limited by any of these, then a takeoff is bounded by construction schedules. That is genuinely good news for coordination: it gives the world a window.
It is also the reason the geography of the transition matters. Whoever controls cheap power and advanced packaging controls the pace at which anyone can attempt the frontier.
The ceiling is industrial before it is intellectual. That changes the politics of the whole transition.
The counter-argument
Efficiency gains can beat the ceiling. Each generation does more per joule, and better algorithms reduce the compute required for a given capability. The historical record in computing is a long series of walls that turned out to be slopes.
Which is why the honest position is this: the compute ceiling is real, it is the best available reason the transition might be slower than feared, and it has never yet held for long.
