Clean compute, where the grid can't reach.

It is what makes the siting possible — and the siting is what makes the capacity obtainable at all.

Generating tokens from a large language model is limited by memory bandwidth, not raw arithmetic — the weights must be read for every token produced. Unified-memory silicon holds the model and reads it fast for very little power. Holding the same capacity in high-bandwidth memory on conventional GPU infrastructure needs roughly 704 kW of facility power: a data hall, three-phase supply, mechanical cooling and a grid connection quoted in years. Ours is a 42 kW load — small enough that a new grid connection is not the starting point for a deployment. That is why we can put capacity where conventional infrastructure cannot go — and why a competitor holding a place in a connection queue cannot shorten it.

Comparator: published NVIDIA DGX H100 specification at a typical enterprise PUE. Our per-node power is an estimate until the silicon ships, so the energy comparison is stated as a range of 1.63×–2.12×; across that range the module’s annual consumption varies by under a tenth of one per cent, so the bill is fixed by the envelope and only the capability it buys moves. We make no claim about the cost of a token — that is set by the model providers.