The Physical Stack of AI · Datacenters and gigawatt campuses
You can name the three cooling technologies in modern AI datacenters, identify the two principal classes of local pushback (water and grid load), and describe at least one siting rule or controversy that's changing how the industry sites campuses.
You can build the substation and order the GPUs. The next two problems are cooling the building and getting along with the neighbors. By 2026, both have become rate-limiters in their own right.
This chapter is short and concrete: the three cooling architectures (air, liquid, immersion), the water controversies that turned a quiet utility footprint into political dynamite, and the siting controls that are reshaping where campuses get built.
The numbers are site-specific, but the pattern is not subtle. The IEA counts cooling and environmental control as a meaningful share of datacenter electricity use, and accelerated servers concentrate more heat into each rack. That heat must leave through air, liquid, or immersion systems. Microsoft's zero-water cooling design shows the direction of travel: chip-level liquid cooling, closed loops, and designs that reduce evaporative water use while accepting some added energy cost.
Chapter contains 3 lessons.