Rack Civil Integration: Mass, Floor-Loading & Seismic Anchoring

Most of the AI-density story is told in kilowatts and litres per minute. This one is told in kilograms and kilonewtons. The HPE GB200 NVL72 rack draws 132 kW nominal and has a worked-example flow of ~140–220 L/min for its 115 kW liquid duty, and it also weighs. Filled with 18 compute trays and 9 NVLink-switch trays in a single 48U cabinet, the rack comes to 3,245 lb (1,472 kg) wet. HPE's September 2026 QuickSpecs (V6) list that mass unchanged from V5, and this guide spreads it over an assumed ~0.64 m² footprint. The rack, its floor-standing CDU and its PDU are three separate loads on three separate load paths.

That is a structural-civil load case, not IT furniture on a floor, and how you carry it (slab, floor system, anchorage, move route) is one of the hardest decisions in the building to reverse. You can re-plumb a cooling loop and re-pull a busway. A slab cast 50 mm too thin, or reinforced for the wrong support and rolling reactions, can only be strengthened, at real cost and disruption.

This chapter treats the rack as a load on the structure. The rack as an IT integration unit (cabling, RU geometry, factory build, burn-in) belongs to Chapter 7.13. The design basis comes from the hall's structural engineering (Chapter 6.2) and from the geotechnical and seismic basis set at siting (Chapter 3.8). What follows are the decisions a project faces when steel meets concrete meets a 1.5-tonne GB200-class rack with OEM-defined caster or leveling-foot reactions: how to verify floor loading, whether to keep a raised floor, how to anchor against an earthquake and a tip-over, and how to get the rack from the loading dock to its position without cracking the slab or the schedule.

The rack arrives with a logistics sheet, and that sheet, not a class-average kilowatt or psf figure, is the structural input. It gives installed-dry and filled mass, footprint, foot and caster geometry, the reaction at each foot, center of gravity and anchorage loads. Chapter 6.2 derives the separate bay, support and rolling checks. Here the engineer of record matches the delivered configuration to that signed load schedule before the first rack rolls.

A rack that passes the bay average can still crack a tile, buckle a pedestal or overload a trench cover on the way in. Match the whole floor assembly's test contact area, wheel diameter, passes, span and working-load basis to the move; a uniform or ultimate rating does not establish rolling capacity. Record each rack position and its dock-to-row route together, so the installation crew receives an accepted path, not a mass-over-area shortcut.

For two decades the raised access floor was the default: a 600-900 mm plenum delivered cold air, routed power whips, and carried liquid and data under the white space. AI density challenges that model in two ways. First, direct-to-chip liquid cooling cuts the air duty the plenum was built for, but does not remove it. NVIDIA's DGX Vera Rubin guide (October 2026), for example, keeps the power-shelf supplies air-cooled. Size the residual air path for your platform, under the floor or overhead (Chapter 5.4).

Second, the mass stresses a pedestal-and-tile system. It must be rated for a plumbed GB200-class rack of about 1.5 tonnes at a high point load, and seismic codes then require lateral bracing of that elevated structure. A rack set on a structural slab avoids both the tile and pedestal load paths.

So some designers put dense liquid-cooled halls on slab-on-grade, or on a heavily reinforced structural floor poured to the rack basis, and run services overhead, in trenches or cable troughs, or on a dedicated service level instead of a continuous deep plenum. The slab carries heavier equipment with simpler anchoring, because you anchor into structural concrete instead of bracing an elevated platform, and it removes the pedestal-movement and tile-deflection failure modes. The trade is flexibility: a slab gives up under-floor service flexibility and easy reconfiguration, and it commits you to a service-distribution strategy up front. It is a genuine fork, and it interacts with the hall typology in Chapter 6.1.

The choice is not absolute. Many halls keep a shallow raised floor for cable management and leak containment while dropping it as the primary cooling path, and some operators reinforce the access-floor grid (heavier pedestals, stringers, structural load spreading) to land dense racks on it. Decide deliberately. Keep a raised floor in a liquid-cooled hall and you owe a point-load verification per rack and a containment plan for the coolant inventory now above an electrified plenum. Go to a slab and you owe an overhead, trench or service-level distribution design, and you accept reduced reconfigurability. What you cannot do is inherit a legacy raised floor and assume it will carry the new racks: its component and rolling capacities are unverified.

A 1.5-tonne GB200-class rack with a high centre of gravity, plumbed to a coolant loop and cabled to a fabric, is exactly what seismic codes exist to restrain. Whether the code requires anchorage depends on the adopted code and ASCE 7 edition, the seismic design category, the component's classification and Ip, and the Chapter 13 exemptions. Where the code exempts the rack, the owner may still demand tip-over restraint, and the OEM may demand it as well. NVIDIA's DGX Vera Rubin user guide, updated October 2026, tells operators to anchor the rack to "an unmovable support" so it cannot tip while a server or other equipment is pulled out of it. The practical decisions are to what standard and with what scheme.

Two regimes dominate. In the US and IBC jurisdictions, ASCE 7-22 Chapter 13 governs nonstructural-component anchorage. The authority having jurisdiction assigns the facility's risk category, and data centers are not universally Risk Category IV. The importance factor and any special certification or anchorage requirements follow the adopted code, the component's function, hazardous contents, life-safety consequence and continued-operation criteria. Apply Ip = 1.5 only where those criteria require it, then calculate each component's force and qualification basis. Telecom-derived equipment may be qualified to Telcordia GR-63-CORE (NEBS) where the owner or governing criteria invoke it. GR-63 grades its shake-table test by earthquake zone, and Zone 4 is the most severe; which zone and acceptance criteria apply is set project by project.

In ASCE 7-22 the design force F_p depends on the component's operating weight, the site spectral acceleration SDS, Ip, the height amplification Hf, the structure's ductility reduction Rμ, and the component factors CAR and Rpo, within prescribed upper and lower bounds. Do not apply that variable set to an earlier edition if the project has adopted one.

The recurring anchorage failures are non-seismic-qualified anchors and anchors set too close to a slab edge, where concrete-breakout capacity is project-specific under ACI 318 Chapter 17 and depends on embedment, edge distance, spacing, cracking, reinforcement, load direction and load combinations. Design post-installed anchors to ACI 318 Chapter 17 using products qualified for cracked concrete and seismic loading under an evaluation report the authority having jurisdiction accepts (ICC-ES or IAPMO UES; AC193 for mechanical, AC308 for adhesive). Verify edge distance, spacing and group effects, and specify special inspection of the install.

This is where the slab-versus-raised-floor choice pays off. Anchoring into structural concrete is direct. A rack on an access floor must reach the supporting structure either through independent anchors to the slab or through a floor assembly engineered for the rack's seismic shear and overturning forces (FEMA P-2082-1, commentary C13.5.7). Attaching it to a removable panel that was not engineered for that load is not enough. Chapter 6.2 sets the structural basis and Chapter 3.8 the site seismic basis.

The most overlooked civil load exists for only an afternoon: getting the rack from the dock to its position. A rack has several masses: crated, unpacked on its casters, installed dry and filled. Find the heaviest of those states that will actually travel the route (several may), apply the lift or roller system's dynamic factor to it, and check that load against what the route can carry. Under a pallet jack the load in motion is routinely higher than the final static load. Automated movers need the same check: Grenzebach's ORW Automatic Rack Mover V1 on the OCP Marketplace lists no load capacity, so get the mover's mass, wheel layout and dynamic factor from the vendor. If the route crosses a raised floor, temporary ramp, trench cover or lower-spec slab section, that is where the floor cracks. Treat the rigging plan as a civil document.

The plan must resolve, end to end, the dock-to-position load path and every floor section it crosses, checked against the moving load, with steel road plates or load spreading where needed. It must also settle door, corridor and turning clearances, and here height is the usual problem, not width. HPE's GB200 NVL72 measures 600 mm wide by 1,068 mm deep and stands 2,495 mm tall (QuickSpecs V6, September 2026), and Supermicro's GB200 NVL72 rack is 2,236 mm tall. Upright, either is taller than a 7 ft (2,134 mm) clear opening.

Measure the actual clear door height and width, thresholds, swept turns and any tilt the OEM allows, and add whatever crate, pallet and handling gear travels with the rack. Ask which of those the OEM dimensions already include. Design any door-frame or wall removal in advance rather than improvising it. Add floor protection (Masonite, steel plate, protective matting) so concentrated wheel reactions do not mar or crack the finished surface, and plan the sequencing against construction and other equipment sets. None of this is exotic, but all of it is high-stakes for a plumbed GB200-class rack of about 1.5 tonnes, and it is tightly coupled to construction sequencing and phased turnover in Chapter 6.6 and to the heavy-rigging EHS program in Chapter 6.9.

Compare the same wheel with every surface it crosses. Chapter 6.2 supplies an unrounded reaction of 10.12392 kN, shown as about 10 kN. The route minimum is min(12, 8.0) kN = 8.0 kN, so the original route fails by about 2.1 kN at the trench cover. The passing 5.2 kN/m² bay and 7.8 kN resting support do not help a wheel balanced over that weaker cover.

Select the engineered replacement and verify its bearing seats, fasteners and underlying structure before reopening the corridor. At 12 kN, the replacement has about 1.9 kN of wheel margin. A sheet of floor-protection board preserves the finish but does not raise the cover's structural rating. Keep the move on hold until the engineer accepts the temporary works and the rigging lead confirms the inspected route. The consequence is one planned route intervention before delivery, instead of a cracked cover, a tipped rack and a blocked hall during the set.

Flip case: if the replacement is rated only 10.0 kN, it still fails the unrounded 10.12392 kN demand, although both read "10 kN" on a rounded worksheet. The crossover is the actual maximum wheel demand for this configuration. Reroute or increase verified capacity. If the handling mass, wheel distribution or dynamic factor changes, return to Chapter 6.2 before comparing the route again. The release record carries the OEM configuration, the engineer's load schedule, the rigging method and each inspected obstruction. Chapter 6.6 owns the delivery window and Chapter 6.9 the exclusion zone and lift authorization.

Sort the decisions by the cost of changing your mind. The civil substrate is the hardest layer in the building to change, and the density ramp is what it must anticipate.

Costly to change (size to the ramp now): slab thickness and reinforcement, the structural grid and bay spacing, the floor-system choice (slab or raised), embedded anchorage provisions, and the geotechnical and foundation basis. These are poured or framed once, and strengthening them later takes specialist works and outage planning. If the hall may host Vera Rubin, size the slab, anchorage and move route for NVIDIA's roughly 4,000 lb (about 1.8 t) Vera Rubin NVL72, a figure published in March 2026 without a statement of whether it is filled. That mass screens at about 580 psf on an assumed GB200-like 0.64 m² footprint, a guide footprint average and not a vendor dimension or floor rating (Chapter 6.2). AMD's double-wide Helios rack (Open Rack Wide), now in production, weighs nearly 7,000 lb (AMD, CES 2026). Its width spreads the load, so its footprint pressure stays close to a GB200 rack's; move routes, lifts and ramps still carry all of it (Chapter 6.2).

Kyber has no published mass. A kilowatt figure is not a mass, so the 330 kW Vera Rubin facility basis and Kyber's ~600 kW planning point say nothing about the load. For a hall that may host Kyber, carry an owner-declared mass, footprint and load-state contingency. For either generation, hold final rack-position acceptance until the OEM logistics sheet exists.

The Open Compute Project's Open Data Center specification (rev 0.7.0, effective August 2026; Chapter 6.1) gives owners a longer-horizon basis. It sets design targets for 2027–2028 hardware. The heaviest single rack is 8,000 lb on a 28 × 48 in footprint. An 8,000 lb ML rack plus its 7,000 lb AC/DC sidecar comes to 15,000 lb on 56 × 48 in. Handling equipment counts toward the moving load. Over its own 9.33 ft² footprint the single rack presses about 860 psf, and the pair about 800 psf over 18.67 ft². Those are footprint pressures, not floor-design uniform loads, so compare them only with the footprint screens of about 470 psf for GB200 and 580 psf for Vera Rubin. They are targets, not the weights of shipping racks. If you adopt them as an owner basis, work out the real support, anchorage and route load cases with Chapter 6.2.

Reversible (defer, keep cheap): the per-rack spreader-plate detail, the specific anchor product within an evaluated family, exact rack positions within a verified grid, and floor-protection logistics for a given move. Decide these rack by rack and move by move. The play is the same as the cooling cliff: reserve the structural headroom that is costly to retrofit, and defer the rack-level detail you can change. Treat the heavier slab and anchorage scheme as a priced owner contingency, weighed against the cost and disruption of strengthening later. Chapter 1.1 frames the density ramp, and Chapter 12.2 prices whether over-building pays.

Release a rack only when its position and every dock-to-row segment pass for the same handling state in the load schedule. Replace or bypass the weak cover before delivery; an accepted bay cannot catch a rack that tips on the way to it.