AI Datacentres
A note to John Gowings · Gowings Bros · Compute, decoded

Compute

From chip to hall, at four scales. The chip does the work. Everything else keeps it cool, fed, and connected.

IThe stack, from die to hall
DIE HBM
GPU · the chip

A single accelerator. Tens of thousands of cores doing matrix maths in parallel.

~80 mm² die · ~700 W USD 25–40k at Blackwell list Economic life 3–4 years
×8
8 GPUs · NVLINK power · network · cooling
SERVER · the node

Eight GPUs in one chassis at terabit speed. The smallest unit a frontier model can train on.

8 GPUs · ~10 kW USD 250–400k per node Liquid-cooled at the chip
×8
RACK · the cabinet

Eight servers in one cabinet, drawing the power of a small suburb. One tile of a much larger cluster.

~64 GPUs · ~80 kW USD 2.5–3.5M per rack Cooling capacity binds the floor
×80+
COOLING PLANT 330 kV GRID FEED
HALL · the building

Many rows of racks. Megawatts of grid feed, megalitres of cooling water, kilometres of fibre, all for the chips.

5,000+ GPUs · 50–200 MW USD 0.5–2 bn all-in The shell does not age
IIWhat the chips do, in two modes
Training all chips, at once
ONE CLUSTER · ONE MODEL · WEEKS TO MONTHS

Tens of thousands of GPUs in one fabric, working as one machine. Network speed inside the building decides whether it converges.

Frontier build One cluster Weeks to months Highest capex per chip
Inference many small jobs
MANY JOBS · MANY SITES · MILLISECONDS EACH

The trained model copied across many small clusters, each serving live queries. Milliseconds, then the chip is free. Distributable across sites.

Production fleet Many sites Milliseconds per query Where the revenue lives
The investor read

An AI data centre is a life-support system for chips. The compute layer ages on a three-to-four year life; the shell, the substation, and the grid right do not. Durable asset: the building and the connection. Wasting asset: what sits on the floor.