The Building: Civil, Structural, Fire/Life-Safety & Construction Execution

The building is no longer a box that wraps the IT — in an AI factory the mechanical and electrical plant is the building, and the data hall is a shrinking tenant inside it, so the typology you pick (greenfield vs powered shell, single- vs multi-story, the white-space-to-gray-space split) is a wager on a density ramp you have not yet seen.

A dense liquid-cooled hall is a structural-engineering problem disguised as a mechanical one: the slab, columns, and anchorage you choose at design freeze silently cap the rack density you can ever land — and a slab is the single most expensive thing in the building to get wrong.

The envelope and the site are the only parts of the facility you cannot retrofit cheaply — they are a one-time bet on climate, security, and a 20-year load profile, made before the first GPU is ordered and unforgiving of every assumption you got wrong.

Prefabrication moves the build from the muddy critical path to a parallel factory line — buying you 6–12 months of speed-to-power, but only on the workloads you can standardize, and only where the road is wide enough to ship the module you designed.

Fire and life-safety is not a compliance afterthought bolted onto a finished hall — it is an insurability gate that silently vetoes cooling fluids, battery chemistries, and density choices made three chapters earlier, and the operator who discovers this at the FM Global review instead of at scoping pays in re-engineering and lost time-to-power.

An AI data center is not finished when it is built — it is finished when it is commissioned, and the only schedule that matters is the one that delivers energized, accepted, GPU-ready megawatts in turnover-sized blocks before the depreciation clock and the revenue window pass you by.

A loaded liquid-cooled AI rack is a moving 3-tonne point load that the slab, the floor system, the anchors, and the move route were probably not designed for — and unlike a cooling retrofit, a structural mistake here is poured in concrete.

Acoustics and emissions are not finishing trades you tune at commissioning — they are permit thresholds set years earlier in Chapters 3.9 and 3.11, and the only cheap variables left to meet them are the ones you size into the site plan before steel is cut: setback distance, equipment placement, and the decision to plumb SCR and attenuators into the generation island from day one.

EHS in an AI data center is not a margin baked into the design and forgotten — it is a living program that must follow the same hazards the rest of this guide engineers, because the very decisions that raise density and goodput (800 VDC, liquid loops, on-site gas, live concurrent maintenance) are the decisions that put a worker one mistake from a fatality.