GPU and AI colocation: costs, density, and how to choose
AI and GPU workloads are fundamentally redefining datacenter requirements. Where a classic rack draws a few kilowatts, a GPU rack can demand several dozen. The bottleneck is no longer floor space, it is power delivery and thermal management.
Estimate my cost →Density, the new binding constraint #
A GPU cluster concentrates very high power in little space. Most data centers built for classic loads cannot power or cool a high-density rack. The first filter is not the price per square meter, but the power available per rack and the cooling capacity.
Air or liquid? #
From around 20 to 30 kW per rack, traditional air cooling reaches its limit. Direct-to-chip liquid cooling becomes essential. Not every site is equipped for it, which makes it a selection criterion in its own right.
Cost is driven by energy #
For GPU infrastructure, electricity and cooling weigh far more than the space rent. The right questions are: how much power per rack, what type of cooling, what energy price, and how much power is genuinely available on site.
Linking GPUs: DAC, AOC or optics #
From 400 Gbit/s up, the cable type directly affects performance. A passive-copper DAC cable is cheap and very low-power, and fits intra-rack links under three metres. AOC takes over up to a few dozen metres between neighbouring racks. Beyond that, you switch to dedicated optical transceivers, which quickly adds to both the energy and the financial bill.
FAQ #
A few cards, sometimes. A dense cluster, rarely, for lack of adequate power and cooling. You need a site built for high density.
Written on 1 September 2026.
From reading to comparing: relevant data centers
Estimate my cost → Compare data centers