The bottleneck is energy

Intelligence is becoming abundant, but the power to run it is not. The constraint is moving from chips to electrons.

The conversation about AI fixates on models and chips, which is reasonable, because those are the parts that improve in ways you can demo. But constraints have a habit of relocating. As the cost of producing intelligence falls, the binding limit is migrating underneath the silicon, to the thing the silicon consumes: power.

Every step change in capability has so far meant a step change in compute, and every step change in compute means more electricity, more cooling, more grid. The frontier of intelligence is starting to press against the frontier of energy, and energy does not follow a software release cycle. You cannot ship a new substation overnight. Transmission lines take a decade. The physical world sets the pace, and the physical world is slow.

This reframes a market that many investors still file under software. The cost curve of intelligence will increasingly be set by the cost curve of power, and the regions, firms, and technologies that can deliver clean, cheap, reliable electricity at scale will quietly become strategic in a way they have not been for a century.

We find the second-order opportunities most interesting: the software that squeezes more useful work out of each watt, the systems that route compute to where power is cheapest, the financing structures that get new generation built. Abundant intelligence is not free. Someone has to power it, and that someone is going to be paid.