More compute per gigawatt

The limit on AI is moving from chips to power.

The world generates about 4 terawatts of electricity on average. Plans for AI data centres grow faster than grids can be built. Doubling the grid in a few years is not a plan.

Where the energy goes

Inside an AI chip, most of the energy is not spent on math. A 32-bit addition costs about 0.1 pJ. Fetching the same 32 bits from off-chip memory costs hundreds of times more. The energy goes into moving data between logic and memory.

The shortest path between the two is no path at all: build one directly on top of the other. Silicon can't do that, because forming a transistor takes heat that destroys what is underneath. Assembly at room temperature can.

That is the bet behind Corbel: more compute out of every watt, by changing how chips are built.