Data Center
A data center is a facility that houses large numbers of networked computers together with the power, cooling, networking and physical security needed to keep them running continuously. For AI, a data center is where accelerators are packed into racks and linked into clusters to train and serve models, and the building itself, rather than any single chip, has become a major constraint on how quickly and where AI capacity can grow. Luiz André Barroso and Urs Hölzle of Google captured the shift in their 2009 book "The Datacenter as a Computer," which argued that a whole facility should be designed and operated as one warehouse-scale computer.
Racks are the basic building block, and AI has pushed their power density far beyond what older facilities were designed for. Servers mount in standardized cabinets built around the 19-inch rack width, and the Open Compute Project, started by Facebook in 2011, published open hardware designs including wider racks and shared power distribution. A conventional enterprise rack was historically provisioned at a few to roughly ten kilowatts. An eight-GPU AI server such as NVIDIA's DGX H100 is specified at around ten kilowatts by itself, so a rack holding several of them, or a purpose-built rack-scale system, draws many times what a legacy room was cooled and wired for. Higher density means shorter cables and faster links between accelerators, which is why designers accept it, but it concentrates heat and power into small areas.
Power is the gating resource, and it is a supply chain problem before it is a technology problem. A data center needs a grid connection, substations, transformers, switchgear, uninterruptible power supplies and backup generators, and securing a large grid interconnection can take years in many regions. Reliability practice follows the Uptime Institute's tier classification, which rates facilities by redundancy and expected availability. Efficiency is commonly summarized by power usage effectiveness (PUE), a metric promoted by The Green Grid: total facility energy divided by the energy delivered to IT equipment, so that 1.0 would mean no overhead at all. PUE is useful but narrow, since it says nothing about how efficiently the IT equipment itself uses its energy or how clean the electricity is, a topic taken up under energy-consumption.
Cooling is a direct consequence of power, since essentially every watt drawn by a chip leaves as heat. Traditional rooms use air: cold-aisle and hot-aisle containment, computer-room air handlers, and chilled water or evaporative cooling outside. Air becomes inadequate as rack power climbs, so AI facilities increasingly adopt direct-to-chip liquid cooling, where cold plates on processors carry heat into a coolant loop, along with rear-door heat exchangers and, in some deployments, immersion cooling in which servers sit in a dielectric fluid. Evaporative systems can lower electricity use but consume water, so site choice involves local water availability as well as climate, and operators increasingly publish water metrics alongside PUE.
Networking sets how well thousands of accelerators behave as one machine. Training requires frequent collective operations such as all-reduce, so clusters use high-bandwidth, low-latency fabrics: NVLink and NVSwitch inside and across servers, and InfiniBand or specially tuned Ethernet between them. Large fabrics commonly use Clos or fat-tree topologies, a design that descends from Charles Clos's 1953 work on telephone switching, to give many parallel paths between any two nodes. Congestion, link failures and slow stragglers can leave expensive chips idle, so network engineering strongly influences real training efficiency. Inference deployments, by comparison, often fit on smaller groups of machines and care more about latency and geographic proximity to users than about cluster-wide bandwidth.
Misconceptions are common here. A data center is not one fixed kind of building: hyperscale cloud campuses, colocation facilities and enterprise server rooms differ widely in design, and AI-specific halls differ again from general-purpose ones. Retrofitting a legacy site for dense accelerators is often hard because of floor loading, power distribution and cooling limits. Claims about the electricity or water use of "AI data centers" as a whole should be checked for what they include, as figures for one facility, one operator, one model and one country are not interchangeable.
Read the data center as the physical body of cloud AI: a machine whose limits are set less by clever algorithms than by how much power can be delivered, how much heat can be removed, and how fast chips can talk to each other.