The Physical Stack of AI · Datacenters and gigawatt campuses
You can describe what's inside a hyperscale AI campus — racks, substations, cooling loops, fiber — and explain why an AI datacenter looks different from a generic cloud datacenter.
A modern AI datacenter is closer to a small power plant with computers attached than to the server rooms most people picture. Across this chapter we'll break it down: the anatomy (what's in the building), the trilemma (power, cooling, land — pick all three or you don't get a campus), and how to read an announcement without being misled by the PR.
The headline difference from a 2018-era cloud datacenter is density. A modern AI rack is no longer just a cabinet of independent servers; it is a tightly coupled compute product with power and cooling designed into the rack. NVIDIA's GB200 NVL72 OCP design supports 72 Blackwell GPUs in one NVLink domain and calls for 120 kW of rack cooling capacity. The IEA's Energy and AI report treats accelerated servers, cooling, networking, and grid connection as the new demand stack. Multiply that rack pattern across thousands of cabinets and you see why the conversation has moved from "how many servers" to "how many megawatts."
Chapter contains 3 lessons.