The New Buildout [/ðə ˈnu ˈbuildout/] n - For most of their history, data centers grew by quantity: more of the same racks, drawing 5–15 kW each, cooled by air, fed by the chain described in The Engines of the Hall. The arrival of large-scale AI broke that pattern. A single AI rack may now draw more power than a whole row once did, and that one fact ripples through every part of the building. These notes track the moves that ripple outward from it.

The Shifts Worth Watching

• Power is the new bottleneck. [/ˈpɒwɛr ˈɪs ˈthɛ ˈnɛw ˈbɒttlɛnɛkk/] n The scarce resource is no longer space or even chips, but megawatts at the fence and the means of getting more. This drives On-site power generation and a scramble for sites near generation.
• Cooling goes liquid. [/ˈkʌːlɪng ˈgɒɛs ˈlɪqʌɪd/] n Air cannot draw heat from a 100+ kW rack, so Direct-to-chip liquid cooling (and immersion) is passing from exotic to default.
• Voltage goes up inside the rack. [/ˈvɒltægɛ ˈgɒɛs ˈʌp ˈɪnsɪdɛ ˈthɛ ˈrækk/] n To carry tens of kilowatts to a single rack without enormous copper, distribution is moving toward 800-volt DC distribution.
• The network turns to light. [/ˈthɛ ˈnɛtwɒrk ˈtʌrns ˈtɒ ˈlɪght/] n Silicon photonics, Co-packaged optics, and Optical interconnects in AI clusters reduce the energy spent moving data between accelerators.
• Everything co-designs. [/ˈɛvɛrithɪng ˈkɒdɛsɪgns/] n Compute, network, power, and cooling can no longer be planned in isolation; see AI factory power and cooling.
• Location becomes extreme. [/ˈlɒkætɪɒn ˈbɛkɒmɛs ˈɛxtrɛmɛ/] n The long-term pressure is not merely about better technology on Earth, but about whether some workloads should leave the planet entirely. See Data Centers on the Moon.

Why These Pages Concern Themselves With It

Each of these is, beneath the surface, a thermodynamics-and-electrons problem. A faster optical link saves watts; saved watts ease the cooling load; an eased cooling load improves Power usage effectiveness; a better PUE stretches the fixed number of megawatts the utility granted you. They are one system.

See also the messier reality of operating all this in Field Notes and Problems We Could Only Work Around. For how the largest operators have pushed efficiency, see Efficient datacenter case studies. For small, custom devices that fill monitoring gaps, see Bespoke Devices for the Datacenter.