Skip to content

Insight, note 01

Silicon is no longer the binding constraint. Power is.

Why power, not silicon supply, now sets how much AI capacity a data centre can run, and where headroom comes from without new plant.

Note 01Capacity

ATHLAZ,

A site can hold more accelerators than it can energise. The rack is available, the floor is available, and the accelerators can be bought. The power envelope, the switchgear and the grid connection cannot, at least not on the timescale the deployment plan assumes.

That changes what an efficiency gain is worth. When power is the constraint, a watt released is not a line on a utility bill. It is capacity that can be scheduled, revenue that can be booked and a deployment that can happen this year rather than after the next connection window.

It also changes where to look for the gain. Adaptive Current Modulation treats the power an accelerator draws as something software can shape. Headroom is then released from hardware already installed and already paid for. Nothing is built, nothing is dug, and control is handed straight back on removal.

The measured result behind that statement is stated plainly on the proof page: what was run, on what hardware, for how long, and what was held equal on both arms.

Up to 21%Less GPU die power, measuredMeasured on NVIDIA H100 NVL under an equal cap on both arms.
HoursTime to deployNo new plant, no new cabling.

Measured on production NVIDIA H100 NVL under an equal power cap on managed and baseline arms. Your fleet baseline confirms the figure for your site.

Rather discuss it directly

Bring your own power envelope to the conversation.

All notes