Building Envelope, Architecture & Site Civil Works

By the time you are detailing the building envelope and the site, the workload archetype (Chapter 1.1), the siting decision (Chapter 3.1), the building typology and hall layout (Chapter 6.1), and the structural basis for the dense liquid-cooled halls (Chapter 6.2) are already fixed. This chapter is where those upstream decisions meet weather, soil, water, and the threat model — and where a surprising number of multi-billion-dollar projects quietly accumulate the defects that haunt operations: a skin that condenses on its inner face, a slab that sits one foot too low in the 100-year floodplain, a yard that cannot accept the next transformer because no one preserved the heavy-haul turning radius. None of these is a marquee decision. All of them are one-way doors.

The organizing tension here is different from the rest of Part 6. A data center is not a building that happens to contain IT; it is a heat engine wrapped in a weather barrier, and the envelope's job is the opposite of an office's. An office envelope keeps heat in during winter and out in summer for the comfort of people. A data center envelope manages a building that is always rejecting tens of megawatts of heat, has almost no human occupancy, and whose interior is held at a tightly controlled — and in the liquid-cooled era, often warm and dry — condition by a mechanical plant that never stops. Design the skin as if it were an office and you get condensation, air leakage, and a plant that burns parasitic power fighting the envelope. Design the site as if it were a warehouse and you discover at delivery that the GSU transformer cannot make the final turn.

Start from the load. A modern AI hall is a high-internal-gain space: the IT load dwarfs anything the climate does through the walls. A 30 MW hall rejects roughly 100 million BTU/hr of heat regardless of whether it is January in Stockholm or July in Phoenix. That single fact inverts the conventional envelope logic. Thermal insulation in the walls and roof is no longer primarily about keeping conditioned air in — the building is fighting to get heat out — it is about three narrower jobs: limiting solar gain that adds parasitic cooling load, preventing the envelope's inner surface from dropping below the interior dew point (condensation), and protecting any air-cooled or hybrid spaces from extreme ambient swings during a mechanical failure.

The dominant moisture and air-control decision is the air barrier and vapor strategy, and it is governed by which way water vapor wants to move. In a hot-humid climate, vapor drives inward toward the cooler, drier conditioned interior; in a cold climate the conditioned interior is the warm, moist side and vapor drives outward. Get the vapor-retarder placement on the wrong side of the assembly and you trap moisture inside the wall, where it condenses, degrades insulation, and corrodes fasteners and steel — a failure that is invisible until it isn't. The liquid-cooling era has, paradoxically, eased one part of this and sharpened another: halls served by direct-to-chip liquid run warmer and at lower air-change rates, so interior humidity excursions are smaller, but the CDU galleries, pipe racks, and any chilled-water surfaces create cold spots where condensation can form on the inside of the building regardless of the skin. The envelope and the mechanical design cannot be drawn by two teams who never speak.

The third envelope job is air-tightness, and it is where most commercial skins quietly fail. Uncontrolled infiltration and exfiltration let humid outdoor air leak into a hall that the plant is trying to hold dry, raising the latent load and the dew-point risk, while letting conditioned air leak out as parasitic loss. A continuous, tested air barrier — verified by a whole-building air-leakage test, not just specified on paper — is the single highest-leverage envelope detail. The fork is blunt: an air-barrier-led envelope (continuous membrane or fluid-applied barrier, sealed penetrations, tested) versus a conventional commercial skin (metal panel or tilt-up with whatever air-sealing the trades happen to achieve). The first costs more per square foot and adds a commissioning step; the second saves that money once and pays it back every year in parasitic cooling load, dew-point incidents, and — in the worst cases — corrosion of the structure it was supposed to protect.

The architectural plan resolves three programs at once: security zoning, physical hardening, and operational flow. They pull against each other, and the building's quality is largely a function of how gracefully the architect reconciles them.

Security zoning is a concentric model — public, semi-public, restricted, and critical — and the architecture must make each boundary a real physical and procedural threshold, not a line on a badge-reader schedule. The classic layering runs from the site perimeter (fence, standoff, vehicle control) inward through a staffed entry and visitor screening, into the back-of-house (loading, storage, gray space), and finally into the white space, which is the innermost zone with the tightest access control, no exterior glazing, and often a man-trap at the threshold. The architectural consequence is that circulation must be designed so that a delivery driver, a contractor, a network engineer, and a security guard each have a path that never crosses a more-restricted zone than their clearance allows. Get the adjacencies wrong and you are forever escorting people through space they should never have entered, or worse, you have a loading dock that opens onto a data hall.

Hardening is a threat-model decision, and the fork is how far up the threat ladder you design. At the low end: standard commercial construction with a perimeter fence and CCTV. In the middle: forced-entry/ballistic-resistant (FE/BR) envelopes at critical thresholds, vehicle-barrier standoff (bollards, berms, ha-has) sized to keep a vehicle-borne threat outside a blast standoff distance, and blast-resistant glazing or no glazing at all on the critical core. At the high end — sovereign, defense, and hyperscale-sensitive workloads — full blast hardening of the critical envelope, progressive-collapse-resistant structure, and EMP/TEMPEST considerations. Each step up multiplies cost and constrains the plan (standoff eats developable land; blast walls eat usable area), so the decision belongs in the design basis, set against a named threat model, not discovered during value engineering when it is too late to add standoff that the site no longer has room for.

A data center is a tall, electrically dense, mission-critical object — exactly the kind of structure lightning finds. The lightning protection system (LPS) is the physical apparatus on the envelope and structure: air terminations (rods/mesh/catenary) that present the preferred strike point, down-conductors that carry the current to ground, and the structural-earthing interface that dissipates it. This chapter owns the physical LPS on the building; the electrical bonding, earthing regime, ground grid, and surge-protective-device coordination are canonical in Chapter 4.11. The two must be designed as one system — an LPS that terminates into a poorly bonded ground grid, or down-conductors not bonded to the building steel and the signal-reference grid, creates exactly the ground-potential-rise transients that the SPD coordination in 4.11 is there to suppress.

The governing standards fork by jurisdiction. The international/IEC world uses IEC 62305, whose Part 2 (updated 2024) requires a quantitative, probability-based risk assessment that calculates site-specific risk and outputs a mandatory Lightning Protection Level (LPL I–IV). The US world uses NFPA 780, a prescriptive installation framework that specifies component geometry without a calculation-driven protection level. Both use the rolling-sphere method to lay out air terminations: a sphere of a level-dependent radius (20 m for LPL I, 30 m for II, 45 m for III, 60 m for IV under IEC 62305-3) is rolled across the building, and anywhere it touches needs protection. The design consequence: a mission-critical data center is almost always designed to the most stringent level (LPL I, 20 m sphere), which means a dense mesh of air terminals and frequent down-conductors — and that geometry must be reconciled with the roof-mounted mechanical equipment, the PV if any, and the air-handling penetrations before the roof is detailed.

The high-leverage decision is whether to use the structural steel as the down-conductor and earthing path (a natural-component LPS, where the building's own frame and rebar carry the current to a perimeter ground ring) or to run a dedicated external conductor system. The natural-component approach is cheaper, more robust, and aesthetically cleaner — but it demands electrical continuity through the structure that must be designed and verified at the structural stage, bonding rebar and steel connections that a structural engineer would not otherwise care about. Decide this late and you are retrofitting bonding straps onto a frame that was never detailed for continuity. Decide it early, in coordination with the structural basis (Chapter 6.2) and the earthing basis (Chapter 4.11), and the building's own steel becomes the cheapest, best Faraday cage you will ever get.

The site civil works are the least visible and most schedule-determining scope on the project. They are also where the campus's relationship to water, weather, and the heaviest objects it will ever receive is fixed permanently into the ground.

Grading and stormwater are coupled and, on large AI campuses, often binding. A data-center site is a vast impervious sheet — building roofs, parking, generator yards, switchyards, and heavy-haul roads typically push impervious coverage to 70-90%, which means nearly all the rain that falls becomes runoff that must be detained and released at a controlled rate. Regulators size detention for the 100-year storm; one large site documented over 4 million gallons (500,000+ ft³) of required post-construction storage. The civil fork is surface detention ponds versus underground detention: surface ponds are cheap but consume developable acres the campus would rather use for the next building or a BESS yard; underground systems preserve the footprint but cost far more per cubic foot. Stormwater is therefore a site-plan variable, not an afterthought — detention requirements directly reshape buildable area and how many halls the parcel can hold. Every project disturbing an acre or more also triggers an NPDES Construction General Permit and a stormwater pollution-prevention plan, a permitting path that lives on the critical path alongside the air and power permits (Chapter 3.9).

Utilities, roads, and heavy-haul access are where civil meets the long-lead equipment supply chain. The single heaviest, most schedule-critical object the site will receive is the HV/GSU power transformer — often 150-400+ tonnes, with lead times around 128 weeks (Chapter 2.3) — and the road geometry, turning radii, bridge/culvert load ratings, and pad access must be designed for it from day one. Get the heavy-haul route wrong and a transformer that took two-plus years to build sits at a weigh station while you re-engineer a turn. The same logic extends to the modular and prefabricated era: if the project is buying factory-built power and cooling modules (Chapter 6.4), the roads, laydown yards, and crane pads must accept the largest module the transport plan specifies. Civil access is a constraint on the procurement strategy, not a downstream accommodation of it.

The yard is where the campus master plan either holds together or fights itself for 20 years. It is a packing problem with hard constraints: the substation and switchyard (Chapter 4.3), the backup generators and their fuel storage (Chapter 4.9), the chillers, cooling towers, and dry-coolers (Chapter 5.7), the BESS, and increasingly behind-the-meter gas generation each demand area, setbacks, and adjacencies that conflict.

The setbacks are not arbitrary; they are driven by fire, code, acoustics, and emissions, and several are owned in detail by other chapters. Fuel storage carries fire-code separation distances and spill-containment requirements (Chapter 4.9, Chapter 6.5). BESS has become the most consequential new setback driver: lithium-ion thermal-runaway propagation has pushed codes (NFPA 855 and equivalents) toward larger separation distances and dedicated deflagration/exhaust provisions, and the BESS yard can no longer be tucked against the building (Chapter 6.5). Generators and their acoustic enclosures and emissions stacks must be placed for exhaust dispersion and community-noise compliance (Chapter 6.8, Chapter 3.11). Evaporative cooling towers need separation from air intakes (Legionella and recirculation) and from property lines for plume and drift. The yard layout is, in effect, the physical reconciliation of half a dozen other chapters' constraints into one buildable plan.

The recurring decision is compact-but-coupled versus spread-but-resilient. Packing the yard tight saves land and shortens cable/pipe runs (and at $11.3M/MW, every acre and every meter of MV cable matters), but it concentrates risk: a BESS fire near the generators near the fuel near the substation is a single-event cascade. Spreading the yard buys fire breaks, maintenance access, and concurrent-maintainability — at the cost of land, longer runs, and a bigger civil scope. The frontier-AI default has shifted toward more separation as BESS and behind-the-meter gas have grown, but the right answer is site-specific and is the kind of decision that, once the pads are poured and the underground utilities are run, is effectively permanent.

The envelope and site basis is ultimately a bet on the local climate and hazard regime, and that bet is encoded in one upstream decision that cascades into nearly every structural and envelope detail: the Risk Category. Under ASCE 7-22 (and analogous provisions elsewhere), a mission-critical data center is typically designed to Risk Category IV — the same tier as hospitals and emergency-response facilities — which raises the design wind speed, the seismic importance factor, and the flood-design elevation relative to ordinary commercial Risk Category II construction, with roughly 1.5x the effective design forces. This is not a detail you can revisit: it sizes the structural steel, the foundation, the envelope's wind-pressure rating, and the flood elevation of the slab, all of which are poured into the ground before the building is operational.

The hazard-specific hardening then follows the regional threat. In hurricane and high-wind regions, the envelope needs impact-rated assemblies, missile-resistant glazing or none, and roof systems rated against uplift and the wind speeds the updated ASCE maps now demand. In seismic regions, the structure, base isolation, and — critically — the anchoring of every heavy plumbed rack and yard equipment item is governed by ASCE 7 Chapter 13 (the rack-civil integration is owned in Chapter 6.7; the geotechnical/seismic basis in Chapter 3.8). In flood-prone sites, the decision is to elevate the critical slab above the design flood elevation (often 100-year-plus a freeboard margin), dry-floodproof the envelope to that line, and keep all critical electrical and mechanical equipment above it — a decision made against FEMA NFHL or equivalent mapping at diligence (Chapter 3.8). In wildfire, hail, and extreme-heat regions, the envelope needs defensible space, ember-resistant detailing, impact-rated roofing, and louver/intake protection. These regional deltas are consolidated as a quick-reference crosswalk in Appendix G; the point here is that they are envelope and site decisions, made once, that the building lives or dies by for two decades.