infrastructure//data center

A data center is a building engineered to keep thousands of servers powered, cooled and connected without interruption, and it is the physical plant behind every cloud region, every training run and every API a company calls. From the outside it is a warehouse; inside it is two industrial systems wrapped around the computers: one that brings electrical power in and one that takes the same power out again as heat.

A data center is a building engineered to keep thousands of servers powered, cooled and connected without interruption, and it is the physical plant behind every cloud region, every training run and every API a company calls. From the outside it is a warehouse; inside it is two industrial systems wrapped around the computers: one that brings electrical power in and one that takes the same power out again as heat.

The IT load is what the servers, storage and network actually draw, and almost every other number in the building follows from it. A 50 MW campus means about 50 MW must arrive from the grid through transformers and switchgear (power chain), and nearly all of it leaves as heat that pumps, chillers and cooling towers must reject, at a cost of their own (the ratio of total facility power to IT power, the PUE, is around 1.1 to 1.6 depending on climate and design). Cooling has moved from air blown through raised floors to liquid piped to the chips, because a rack of current AI accelerators draws over 100 kW, far more than air can carry away from that volume (power density).

Redundancy at every level, even for the backup.

A server that loses power for a few milliseconds reboots, so the design goal is that no single failure (a utility feeder, a UPS, a generator that does not start, a breaker) ever reaches the chips. Each layer is duplicated or given a spare, and the way that is done is stated as a redundancy level.

The members of this family each answer one question. Who owns the building and how much of it you rent is colocation, from a cage in a shared hall to a whole hyperscale campus. How much power fits in each square metre or rack is power density, the number that decides whether a hall can take AI hardware at all. How the electricity reaches a chip without a gap is the power chain. How the computers are stacked (server, rack, row, pod) and joined (interconnect) sets what workloads the building can serve.

The scale is the point. A data center is closer to a process plant than to an office: the same concerns of availability, maintenance without shutdown, single points of failure and common-mode failures apply, and the same engineering culture (one-line diagrams, commissioning, load tests) runs it.

Power is now the binding constraint. A new campus is sited where the grid can deliver hundreds of megawatts, and the wait for a utility connection often outlasts the construction, which is why the economics of compute (cloud computing, compute clusters) increasingly start from megawatts rather than from servers.