Structural & Civil Engineering for Dense Liquid-Cooled Halls

The cooling chapters give the HPE GB200 NVL72 a 132 kW nominal rack TDP, and the guide's water-property example puts its 115 kW liquid duty at ~165–236 L/min for a 7–10 K rise. HPE’s V5 QuickSpecs (July 20, 2026) also give its weight: 3,245 lb filled, about 1,472 kg or 1.5 tonnes. For screening, the guide assumes a 3,000–3,300 lb wet rack over about 0.64 m², which works out to about 2,100–2,300 kg/m² or 435–480 psf. The 3,000 lb lower end is an assumption; the 3,300 lb upper end matches HPE’s fully loaded GB300 NVL72. Whether the hall can host the rack is still decided by the actual foot or wheel reactions.

Every other Part 6 decision sits on the structural and civil basis of a dense liquid-cooled hall, and unlike a transformer, a CDU or a fabric switch, a slab cannot be swapped mid-life without decanting the hall. Get the floor-loading basis wrong and you do not have an inefficient building; you have an air-cooled-only building committed to a liquid-cooling roadmap it cannot deliver without decanting and strengthening the structure.

This chapter owns the civil and structural basis for dense halls, set out as decisions with downstream costs. It covers slab-on-grade versus elevated structural slabs, the long-span steel and column grid that make a hall workable, and what liquid adds to a multi-story building: live load, deflection, vibration and the dead weight of coolant and pipe. It also covers the crane and rigging provisions that make equipment moves possible, and the seismic and anchorage strategy for racks that are heavy, plumbed and energized at once. Floor loading for 3,000+ lb racks is owned here as the civil basis the rest of the guide inherits: load states, area averages, support reactions, rolling loads and the limits they impose. The geotechnical and soils interface, and site-level seismic hazard, tie back to Chapter 3.8.

The first and least reversible structural decision is what the racks stand on. There are two answers, and they pull the whole building in different directions.

Slab-on-grade rests the floor directly on prepared subgrade, so the soil and the concrete govern capacity, not beams and girders. Most purpose-built dense AI halls in 2026 are single-story slab-on-grade, because the short load path makes capacity cheap. The notable exceptions go vertical on purpose: Microsoft's Fairwater buildings in Wisconsin and Atlanta are two-story liquid-cooled halls at about 140 kW per rack, stacked so racks can be networked to racks above and below with shorter cable runs.

The slab's capacity is not a class number: the engineer of record designs it for the project's dead, live, concentrated, rolling, rigging, seismic and anchorage load states, so do not treat the ~435–480 psf footprint average as a slab allowance. Design from the site's geotechnical basis and those load states, not from a class-average thickness. Check rack support and rolling reactions, joint layout, reinforcement and serviceability, and let thickness, post-tensioning and reinforcement fall out of that check. The premium over a legacy air-cooled slab is real, and it is a project quantity, priced from the bill of quantities for that design, in that market, at that date. The slab is only as good as the subgrade beneath it, which puts the soils basis in Chapter 3.8 on the critical path.

An elevated structural slab is a suspended concrete deck on a steel or concrete frame. It is what you pour when you go multi-story to conserve land or shorten rack-to-rack cable runs, or when you need a service level below the white space for piping and power. The load path now runs through beams, girders and columns, and each must be sized for the wet-rack live load plus the dead weight of the coolant, pipe and equipment it carries. Capacity is no longer cheap: doubling the design live load on an elevated deck cascades into bigger members, more steel, more foundation and tighter deflection control. Slab-on-grade buys the most density per dollar but spends land. The elevated slab saves land but pays for every pound in structure, and it brings the multi-story liquid problem described below with it.

Structural engineers design floors to a uniform live load in psf or kN/m², but racks deliver support reactions through feet or wheels. The two have to be reconciled; they are not equal. Walk the arithmetic, because the whole civil basis turns on it. A guide-assumed loaded compute rack of ~1,360–1,500 kg on a 0.6 m × ~1.07 m (0.64 m²) footprint imposes about 21–23 kN/m² (~435–480 psf) averaged over its own footprint. Racks do not tile the floor edge to edge, though. Allocating an assumed 1.44 m² to each 1,472 kg rack gives 1,472 × 9.81 ÷ 1,000 ÷ 1.44 ≈ 10 kN/m² (~210 psf) of rack-only bay demand. Aisles, service clearance and CDU/manifold lanes set that denominator, and other equipment and building loads are added on their own paths. So a hall can be specified at a deceptively modest psf and still carry severe local support reactions, and the binding check can be the worst-case concentrated load on a single caster path during rigging even when the tidy hall average passes.

Most floor-loading mistakes come from this gap between footprint load and hall-average load. Design only to the average and you under-provision the slab against the support and rolling reactions of a rack being moved into place. Design the whole hall to the worst-case footprint load and you pay for steel and foundations sized as if racks filled the aisles. Specify both: an equivalent uniform live load for the hall structure, and maximum static foot and wheel reactions plus rolling and dynamic reactions for the rack positions and the rigging paths between them. NVIDIA’s logistics and support data for the rack you are buying feed the structural engineer’s site-specific load basis. The circulating ~440 psf footprint back-calculation cannot replace those data, and the wet, cabled weight governs, not the datasheet empty weight.

Trace the installed load to the ground. One rack weighs 2,000 kg × 9.81 m/s² ÷ 1,000 = 19.62 kN, about 20 kN. Its footprint average is 19.62 kN ÷ (0.60 m × 1.20 m), about 27 kN/m². The bay carries (6 × 2,000 + 600 + 240) kg × 9.81 m/s² ÷ 1,000 ÷ 24 m², about 5.2 kN/m². The most-loaded support carries 0.40 × 19.62 kN, about 7.8 kN. The bay demand is below 8.0 kN/m² and the support demand below 9.0 kN. Neither comparison uses footprint pressure as a floor rating.

The installed load runs from the feet through the slab and subgrade, or through deck, beams, columns and foundations in the elevated case. The allocated bay load and the local reactions pass these two supplied capacities. Joints, punching, settlement, anchorage and the movement of connected services still need their own matching checks. Take this equipment arrangement forward to detailed structural release; do not buy stronger members just because the footprint average exceeds the bay allowance. That would pay for rack-level concentration over the aisles as well.

The move is a separate load state. Its maximum wheel reaction is (2,000 + 150) kg × 9.81 m/s² ÷ 1,000 × 0.40 × 1.20, about 10 kN; carry the unrounded demand into the route comparison in Chapter 6.7. Check the joints on the route, not just the slab panels: hard casters crossing an unfilled sawcut joint spall its edges. ACI 302.1R and 360R guidance is to fill those joints full depth with a semi-rigid epoxy or polyurea filler (about Shore A80 or harder) and to dowel or armor construction joints. The filler goes in best after most slab shrinkage, so a fast-track hall that sets racks weeks after the pour needs temporary route protection and a planned refill.

In the installed-position flip case, a 2,500 kg rack drives the support to about 9.8 kN and fails the 9.0 kN capacity, even though the bay stays below 8.0 kN/m². The support crossover is 9.0 × 1,000 ÷ (0.40 × 9.81), about 2,300 kg; check it without display rounding.

ACI 360R-10 sets out slab load and support conditions; its public preview gives the method’s scope, not these capacities. This chapter owns the load derivation. Chapter 1.7 supplies the equipment range, and Chapter 6.7 releases the actual rack and route.

A hall is more than a floor. It is a volume that equipment has to move through for two decades, and the column grid and roof span decide whether that volume is workable. Dense AI halls trend toward long-span steel with wide column bays for one blunt operational reason. NVL72-class racks ship pre-integrated at 3,000–4,000 lb and must be rigged into position, CDUs and manifolds need clear runs, and overhead cable tray and busway want unobstructed paths. Every column in the white space is an obstacle that a forklift, a rigging skate or a future row reconfiguration has to route around. A tight column grid saves steel up front and costs flexibility and move time for the life of the building. That capex-now, opex-forever trade almost always favors the wider bay in a building expected to absorb three GPU generations.

Roof and overhead structure carry a load case that is easy to miss. In liquid halls a great deal of the mechanical distribution (pipe racks, manifolds, and sometimes overhead cable tray and busway) hangs from the structure above the white space. Take those suspended loads per linear metre from the tray, busway and pipe schedules, not from a per-rack rule of thumb. That is suspended dead load the long-span steel must carry on top of its own weight and any crane provision. It also interacts with deflection limits, because sagging structure pulls on rigid pipe joints. Running liquid and power overhead (common where racks sit on slab) or in a service level below (common in multi-story) is therefore a structural-loading decision and a maintenance-access decision at once; the hall-layout side is in Chapter 6.1.

Going vertical to conserve land, or to shorten scale-up and scale-out cable runs, is increasingly common in power- and land-constrained metros. Liquid cooling makes it materially harder than a stacked air-cooled building ever was. Four structural problems compound floor on floor:

• Live load on every deck. On slab-on-grade the wet-rack live load passes straight to soil. On each suspended floor, beams, girders and columns carry it, and the columns accumulate the load of every floor above. The lowest columns and the foundation carry the whole stack, so they grow with every floor added above them: a load pyramid that drives oversized lower members and deep foundations.
• The dead weight of coolant and pipe. Air weighs nothing structurally; water does not. The rack's own coolant is already inside the loaded-with-fluid mass the OEM publishes. Take the contained volume from the OEM, or derive it from the cold plates, manifolds and hoses rather than from a flow rate in L/min, and do not add an assumed inventory on top of a wet mass that already includes it. Beyond the racks, the facility distribution (risers, mains, CDUs and full pipe) adds permanent dead load on every deck that an air-cooled building never carried. It is present 100% of the time, so it adds more long-term deflection than transient live load does.
• Deflection and the rigid-pipe conflict. Elevated decks must hold tight deflection limits, such as a project-selected L/360 gravity criterion, with differential movement checked separately at rigid pipe supports. A floor that sags under load stresses the pressurized coolant joints running across it. A leak from a deflection-driven joint failure is a structural problem that becomes an electrical-safety problem one floor down, so deflection control is a coupled mechanical and structural requirement.
• Vibration. CDU pumps, large pipe flows and rotating plant introduce vibration that a stiff slab-on-grade absorbs but a framed floor can amplify or transmit. Apply the relevant ISO 20816 part only to machinery within its scope; pipe and support vibration and floor serviceability need their own agreed limits. The structure must be stiff enough that resonance does not loosen connections or fatigue supports over years of continuous operation.

So a multi-story liquid hall trades land savings for a structure that is heavier, stiffer, more tightly deflection-controlled and more expensive per square foot than its slab-on-grade equivalent. In a land- or power-bound metro that can be the right trade, but it is a trade, and the structural premium belongs in the siting math, not discovered after the frame is designed.

Pre-integrated racks and multi-ton CDUs do not levitate into place. The hall has to be designed for how heavy equipment arrives, moves and gets replaced, and that is a structural provision. Three load cases matter, and they are routinely under-scoped.

Rigging paths and dynamic loads. A 3,000–4,000 lb rack rolled on skates concentrates its full weight on a narrow caster path that moves across the floor. A rack being lifted imposes a dynamic load at the lift point, an impact factor above its static weight. The slab and any elevated deck along the rigging route, from the loading dock through the corridors into the hall, must be rated for these moving concentrated loads, which are often higher than the resting load at the final position. A hall whose resting positions pass but whose dock-to-row path crosses an under-rated deck cannot receive the equipment it was built to host.

Crane and hoist provisions. Overhead monorails, jib cranes or gantry provisions for setting racks and swapping CDUs put suspended dynamic loads on the roof structure, so design them in from the start. Retrofitting a crane rail into an occupied dense hall is slow, disruptive and costly. Knockout panels, oversized doors and clear height for equipment ingress are cheap at design time and slow, disruptive and costly to add later. Reserving the rigging corridor, the crane load path and the ingress route applies the same density-ramp discipline that governs floor loading: the move-in path is far dearer to retrofit than to provision, so you provision it. Construction sequencing, and the equipment-set phase that exercises all of this, are in Chapter 6.6.

In a seismic zone an AI rack is an unforgiving structural object: it is heavy, plumbed with pressurized coolant and energized at 415 VAC or 800 VDC, all at once. An earthquake that merely tips a server rack is an annoyance. One that ruptures a coolant line above an energized power bus, or shears a quick-disconnect under a moving rack, is a flood-into-electrical event. So seismic protection does more than protect life and equipment; it keeps the liquid and the electricity apart when the ground moves. The building code already requires anchorage of heavy installed equipment above the lowest seismic design categories (ASCE 7 Chapter 13 in the US); beyond that, lenders and insurers price seismic risk through probable-maximum-loss studies, which links this decision directly to Chapter 2.6.

The design ladder runs from cheapest and stiffest to most protective and costly. Anchorage ties racks and equipment rigidly to the floor so they ride with the building. It is adequate in low-to-moderate hazard, but it passes the floor's full acceleration into the racks and, critically, into the rigid coolant piping connected to them. On an upper floor that acceleration is amplified above ground motion: ASCE 7-22 Chapter 13 scales component force with height in the structure, and the older ASCE 7-16 form reached three times ground acceleration at the roof. Size anchorage and flexible crossings per floor, not per building; the anchorage derivation is in Chapter 6.7. Stiffened or braced structure raises the building's own seismic capacity. Base isolation (equipment-level ball-bearing or elastomeric platforms, or isolating the whole building at its foundation) decouples the equipment from ground motion, permitting a project-selected displacement, for example an assumed 8 inches, while holding vertical stability. It is the high-hazard answer, and it changes the plumbing problem entirely. An isolated platform that moves through its full rated travel relative to the building needs flexible coolant connections, slack power whips and seismic loops in every pipe crossing the isolation plane, or the isolation shears the very services it was meant to protect.

Every load path above ends in the ground, and the ground is the part of the structure you did not get to design. A slab-on-grade is only as strong as its subgrade bearing capacity. An elevated frame's deep foundations are sized to the soil's resistance and settlement behavior, and a base-isolation system's performance assumes a defined site-response spectrum. So the structural basis cannot be set apart from the geotechnical basis; they are one decision seen from two ends. Poor bearing soils, a high water table, expansive clays or a stiff seismic site class can force a different and more expensive foundation system, or push a planned slab-on-grade onto piles and pile caps, materially changing the civil cost and schedule.

The geotechnical investigation is a critical-path input that must come before the structural design. You cannot finalize slab thickness, reinforcement, foundation type or seismic detailing until the soils report and the site-specific seismic hazard are in hand. Skipping or compressing the geotech to save schedule is a classic way to discover, with the frame half-designed, that the floor-loading basis the whole density roadmap depends on rests on soil that cannot carry it. Land, soils, seismic hazard and flood diligence are covered in full in Chapter 3.8; that diligence is a secondary screen, but it gates this entire chapter.

Choose the floor and restraint scheme from the bay demand, local reactions and movement checks together. A cheaper slab or a tighter column grid is worth its saving only if the rack you are installing can stand, move and reconnect safely; otherwise the saving turns into decanting, reinforcement and lost service.