Building Typologies & Data-Hall Layout

The legacy data center was a solved problem: a single-story tilt-up shell, a raised floor, a perimeter of CRAH units, and as much white space as the developer could lease. The IT was the point and the building was the wrapper. AI factories reverse that. A GB300 NVL72 rack needs up to 142 kW in NVIDIA's reference architecture and up to 155 kW peak in Lenovo's rating. At those loads the cooling plant, the electrical rooms and the heat-rejection yard grow until they can dominate the developed site, and the data hall, the part everyone still calls "the data center," becomes a minority tenant inside its own mechanical and electrical plant. The building is now the plant; the data hall is what is left over. NVIDIA's DSX reference design plans Vera Rubin NVL72 cabinets at 330 kW TDP for facility power and cooling, while Pegatron's RA4803-72N3 datasheet rates its profiles at 188 kW (Max Q) and 228 kW (Max P). NVIDIA targets H2 2027 for ~600 kW Kyber-class racks; SemiAnalysis reported 2028 in July 2026.

This chapter turns that reversal into layout decisions, in the order they get made. First comes who owns the density risk: greenfield or retrofit, and powered shell or integrated build. Next is single-story or multi-story, then the data-hall layout itself: white space, gray space, adjacencies, plant rooms and the yard. Then comes the space liquid cooling needs (CDU galleries, pipe racks, leak zones, service access), and last the repeatable module and future-flex discipline of bay sizing, knockouts and reserved headroom that separates a building you can ramp from one you have to abandon. The civil and structural basis for these layouts (slab loading, long-span steel, seismic anchoring) is in Chapter 6.2, and the envelope and site civil works are in Chapter 6.3. This chapter owns the shapes.

Before any layout, decide what you are building. Greenfield is a clean sheet on raw land: full control of column grid, slab loading, ceiling height and plant adjacencies, at the price of the longest schedule and the deepest capital commitment. Retrofit reuses an existing shell and starts from its limits. A hall poured for 5–10 kW air-cooled racks can have too little floor loading, no route for liquid distribution, too little electrical headroom and no facility water, the same cooling cliff covered in Chapter 5.1. Without strengthening, those inherited capacities cap the conversion. Retrofit saves schedule and capital where the structure, power and water already suit the target density. Going denser can take engineered strengthening or slab replacement, so price the strengthening, the decanting and the schedule into the comparison (Chapter 6.2 covers the structural options).

Cutting across that is the delivery model: one integrated facility, or a base building (the powered shell) plus a fit-out. The powered-shell model separates the slow building program from the tenant's faster IT selection. A developer delivers the envelope, the structure, utility power to the boundary and primary distribution, the slow, irreversible, capital-heavy substrate. The tenant then installs the IT-specific fit-out (racks, CDUs, the technology-cooling loop, the network) on its own schedule and for its own accelerator generation. The split moves fast-changing risk to whoever is best placed to carry it, and it lets the shell start two years before anyone has decided which accelerator goes inside. JLL's 2026 capacity outlook describes the demand pressure behind the split (Chapter 0.1). The integrated schedule still has to show that the split gets the tenant to service sooner.

The cost of the split is an interface. Someone must own the boundary between base building and fit-out: the busway tap-off points, the facility-water headers and valved stubs, the floor-loading guarantee, the knockouts for future risers. Get it wrong and the shell is energized but cannot be filled. The water stub is in the wrong bay, the busway is rated for last generation, or the floor letter gives only an area average, and nobody has checked the OEM's foot and wheel reactions, the rolling loads, the full floor assembly or the rigging path. A powered shell keeps your choice of IT generation open only if the irreversible substrate was sized to the ramp, not to the lease that signed it.

The default data center is single-story, for good reason. Heavy equipment sits on grade. The slab is accepted only after project-specific checks: geotechnical, thickness and reinforcement, joints, concentrated reactions, punching, rolling loads and anchorage. Pipe runs are horizontal, and a leak goes to a floor drain rather than onto the hall below. Where land is cheap and plentiful, and network reach and the other design constraints also work, single-story remains the right answer. Most US greenfield AI campuses are sprawling single-story halls because power and land, not vertical area, are the constraint.

Multi-story is what you build when land is the binding constraint (dense metros, sovereign sites, water-adjacent or grid-adjacent parcels you cannot enlarge), or when stacking shortens electrical and fiber runs enough to matter at gigawatt scale. Network reach counts too: Microsoft says its two-story Fairwater design places racks in three dimensions to cut cable lengths and improve network performance (November 2025). The 2026 generation has made multi-story harder, because what you now stack is heavy and wet. A multi-story liquid-cooled hall must carry wet racks of ~3,000–3,300 lb (GB200/GB300 NVL72) to ~4,000 lb (Vera Rubin NVL72) on an elevated structural slab instead of slab-on-grade. A shell designed to the OCP Open Data Center targets must carry up to 8,000 lb per rack, or 15,000 lb per rack-and-sidecar pair (rev 0.7.0, 2026; design targets for 2027–2028, not measured racks). That drives a far stiffer, costlier structure to control deflection and vibration. The hall must also route the technology-cooling loop and facility water vertically, which adds risers, riser cores and a leak path that points down at the IT below, and it must reject heat from a roof or yard that serves several stacked halls at once. Stacking is paid for in structure, in vertical liquid distribution and in leak-cascade risk that single-story does not have.

Single-story is the default. Multi-story trades structural and liquid-distribution cost for land efficiency and shorter runs, so choose it because land, grid adjacency or network reach forces your hand, not because the renders look impressive. If you stack, the structural basis (elevated-slab live loads, deflection, vibration, coolant weight) is in Chapter 6.2 and the rack-as-civil-load verification is in Chapter 6.7. This chapter only sets the shape that forces those problems.

Two terms govern the layout. White space is the conditioned floor where the IT lives: racks, network and the immediate cooling at the rack. Gray space (or back-of-house) is everything that supports it: electrical rooms (switchgear, UPS, transformers), mechanical plant (CDUs, pumps, heat exchangers), air handling, the BMS room, and staging and storage. In a legacy enterprise hall, white space dominated and gray space was a thin perimeter. The building was mostly IT floor, and the developer leased it by the square foot.

The AI factory inverts that ratio, and the inversion drives every layout decision that follows in 2026. At 132 kW per rack instead of 5 kW (HPE's rating for a GB200 NVL72), the same megawatt fits in about a twenty-sixth of the rack positions. The switchgear, UPS, CDUs and heat rejection serving that megawatt do not shrink with it, because plant area scales with megawatts, not with rack count. Indoor gray space (electrical rooms, UPS, switchgear, CDU galleries, mechanical plant) can then claim as much of the gross floor area as the white space, or more, depending on which plant sits indoors. The heat-rejection yard, substation and generator field take site area on top of that, and the white space shrinks to a dense, valuable core.

Building area and developed site area are two different denominators, and a plan that mixes them under-sizes one of them. NVIDIA's DSX reference design (August 2026) shows the gap: its example puts a 250 MW-class IT load and 1,536 GPU racks in a 972,500 sq ft building, about 22 acres, on a 157-acre site with dry coolers at its east and west edges. That is roughly 630 sq ft of building for every rack. A hall planned on legacy white-to-gray ratios runs out of plant room and electrical-room space long before it runs out of IT floor: a well-proportioned data hall with nowhere to put the megawatts of switchgear and the acres of heat rejection it needs. Plan the gray space first. Size the electrical rooms and cooling plant to the megawatts, then fit the white space into what remains, the opposite of the legacy instinct.

The adjacencies follow from that. Electrical rooms want to sit near the racks they feed, to keep busway and cable runs short and losses low, but they must be fire-compartmented from the hall. The cooling plant and CDU gallery want to sit close to the white space, which keeps the technology-cooling loop short, low in pressure drop and fast to drain. Outdoor heat rejection (dry coolers, cooling towers) belongs in the yard or on the roof, where it competes for setback-limited space with the substation, generators, fuel and BESS (the yard layout is in Chapter 6.3). Chillers can go either way: NVIDIA's DSX example houses them in central utility buildings, while other plants put them outdoors. Carry each footprint into the matching schedule, indoor or site. Every layout must also reserve a laydown and staging area. Dense racks arrive factory-integrated and are rolled in along rigging paths, so a hall with no laydown space and no clear move route cannot be populated on schedule. Construction sequencing and rigging are in Chapter 6.6.

Lay out rows by power per linear foot of row, not by rack count. The OCP Open Data Center specification (rev 0.7.0, effective 27 August 2026) replaced fixed row lengths with kW/ft targets: a 25 kW/ft blended average, which requires 15 kW/ft of air-cooling capability and 25 kW/ft of liquid-cooling capacity, and a 350 kW/ft peak for ML rows, with 125 kW/ft expected from 2027 peak racks. At 125 kW/ft, a 56 in rack-and-sidecar pair would carry about 580 kW. That is this guide's arithmetic, not a Kyber product specification. The same document sets minimum row geometry as accommodation targets: a 54 in hot aisle, an 82 in cold aisle measured rack face to rack face, and room for racks 100 in tall and 71 in deep. It recommends racks front-justified in the cold aisle, on a 24 ft aisle pitch. Draw the column grid around the racks you choose and the space they need for service and replacement, because a fixed pitch alone does not guarantee that a later rack fits. The transport aisle and power-block provisions are in the future-flex section below.

Air-cooled halls had a simple spatial logic: hot aisle, cold aisle, containment and a plenum. Direct-to-chip liquid adds rack-side liquid distribution on top of that air layout and does not replace it. In HPE's GB200 NVL72, 115 kW goes to liquid and 17 kW per rack stays on air, so the room must still be cooled by air for that remainder, with airflow routes sized to it. Chilled-water halls already ran pipe, but liquid to the rack brings plumbing next to the IT and makes space for it a layout decision. Four items grow or move toward the IT, and each has a service-access consequence.

CDU galleries. The coolant distribution unit separates the clean, controlled technology-cooling loop (the secondary loop, to the cold plates) from the facility water loop (the primary). It also provides the pumping, heat exchange, filtration and dew-point control for the rack-side fluid; its internal design and sizing are in Chapter 5.6. Spatially, CDUs are large, heavy, serviceable machines that need a home. They sit either in-row (in the rack line, close-coupled to a few racks and easy to drain, but using white-space floor) or in a dedicated CDU gallery along the hall perimeter or in adjacent gray space (centralized and serviceable without entering the IT rows, but with longer secondary-loop runs). The choice trades white-space density against service access and loop length, and it must be made before the column grid is fixed, because the gallery's footprint and the manifold routing depend on it.

Pipe racks and risers. Liquid distribution needs a physical route: overhead pipe racks carrying supply and return headers down the hall, drop legs to each rack-row manifold, and, in multi-story buildings, vertical riser cores between floors. Air-cooled halls spent this volume on a deep plenum; liquid halls spend it overhead and in dedicated chases. If you under-size the pipe-rack space, you cannot add capacity without re-routing live plumbing. If you forget the riser cores in a multi-story shell, the upper floors are stranded. The hydraulic design of these loops is in Part 5; this chapter only insists that the space be reserved at layout time.

Leak zoning. Liquid inside the white space puts a leak path next to the IT. Good layout treats the hall as a set of leak zones, bounded regions with their own leak detection, bunded or curbed containment and isolation valves. A failed quick-disconnect or manifold breach then wets one zone and is isolated, rather than wetting the whole hall or, in a multi-story building, cascading onto the floor below. Zoning is a spatial discipline: it dictates where isolation valves sit, where containment curbing runs and how the drainage is graded. It is cheap to design in and very expensive to retrofit after a hall is energized.

Service access. Every item above must be reachable without shutting the hall down. Concurrent maintainability, the Uptime Institute's Tier III and IV site-infrastructure property, means the required capacity components and distribution paths can be removed for planned maintenance while the critical load stays served. It depends on layout before it depends on the single-line diagram. Working inside an operating IT row does not by itself break it. What matters is whether you can isolate a component, reach it, remove and replace it, and keep the remaining duty units serving the load. If reaching a CDU or manifold valve means interrupting served IT, the hall is not concurrently maintainable, whatever the single-line diagram claims. Reserve the access aisles, valve-pit clearances and CDU pull-out space at layout time, or pay for them in downtime later. Commissioning these loops and proving the leak response is in Chapter 13.5.

At gigawatt scale nobody designs a hall; they design a module and pour it many times. The unit of repetition is a standardized block (a data hall of fixed size, a power-and-cooling skid, a pod) with frozen interfaces, replicated across a campus to a master plan. Vendor reference designs show the idea. Vertiv's 2024 360AI reference block with NVIDIA is a roughly 7 MW data hall at 132 kW/rack with prefabricated power and cooling modules, sized so the electrical, mechanical and structural interfaces repeat identically; current-generation blocks size to GB300 densities and above. The master plan exists to make the module tileable. Modular-construction trade-offs (factory build against stick-built, transport limits, commissioning in the factory) are in Chapter 6.4.

Campus scale shows the stakes: Crusoe's Abilene campus is planned at 1.2 GW across eight buildings, with two operating and six still under construction in Crusoe's June 2026 update. Epoch AI's satellite tracker runs ahead of that count: it estimated about 421 MW of IT in late July 2026, with buildings three and four probably running and the last four roofed but still in fit-out.

Standardization is a sharp trade-off. The upside compounds. The first module is slow and expensive to design and commission, but each later one reuses the same grid, interfaces, commissioning scripts and supply-chain order, so schedule and quality improve as the campus ramps. The downside compounds too: a mistake in the standard unit scales with the unit. An under-sized CDU gallery, a riser core one bay short, or a floor-loading spec that just misses the next wet rack becomes a campus-defining liability when it is replicated forty times. The master plan therefore has to fix the seams between modules as well as the module: shared yard and substation capacity, inter-module fiber and busway, phased construction that lets one module be commissioned while the next is still steel, and expansion bays held in reserve for a generation not yet announced. Siting and yard context for the campus is in Chapter 3.1 and Chapter 6.3.

The defining condition of 2026 is that the density ramp is steeper than any building's depreciation. A hall scoped for today's GB300 NVL72 at up to 142 kW (NVIDIA's reference architecture; Lenovo's peak is 155 kW) will, within its structural life, be asked to host racks at 2x to 4x that draw. Vera Rubin NVL72 is the intervening fit-out, planned at 330 kW per cabinet in NVIDIA's DSX reference design, and the ~600 kW Kyber-class rack remains a roadmap planning point (NVIDIA targets H2 2027; SemiAnalysis reported 2028 in July 2026). Do not install 600 kW of fit-out on day one: those racks are not shipping. Install only the fit-out your committed IT needs, and reserve the structural and utility interfaces for the future rack you set as the design basis. That is future-flex: the irreversible substrate accommodates the ramp while the reversible fit-out stays matched to the current generation.

There is now a published basis for that ramp. The OCP Open Data Center specification, rev 0.7.0 (AMD, Google, Meta and NVIDIA; effective 27 August 2026), sets design targets for 2027–2028 systems. Adopting them is a choice, not a code requirement. They include a single rack of up to 8,000 lb on a 28 × 48 in footprint, and 15,000 lb for an ML rack plus its 7,000 lb power sidecar on 56 × 48 in, with handling equipment counted in the moving load. They also include a recommended 12 ft transport aisle; a minimum clear height of 18 ft, floor to beam bottom, with two busways per row (layouts with four busways and extra tray tiers may need more); at least four power blocks per data hall; and clusters of 12 MW minimum and 24 MW preferred. Row density and aisle geometry are in the layout section above. Spread over the rack footprint, the single rack is about 857 lb/ft² (8,000 lb over 9.33 ft²) and the pair about 804 lb/ft² (15,000 lb over 18.67 ft²). These are footprint pressures, not uniform floor-design loads; floor loading is in Chapter 6.2 and rack-as-civil-load verification is in Chapter 6.7. If you adopt these targets, write them into the reservation register as the design basis. Later changes to the slab, the bays, the clear height or the move route can mean emptying a live hall to strengthen or replace them.

Concretely, future-flex is a short list of provisions that are cheap now and costly later. Bay sizing generous enough to absorb deeper, heavier racks and wider service aisles without re-gridding. Floor loading signed for the future wet-rack mass and reactions you set as the design basis, because strengthening a slab under operating racks needs a disruptive decant and cannot be priced from kilowatts. Clear height for the busway tiers, tray and plenum the next generation needs (the OCP 18 ft applies to two busways). Knockouts: pre-formed openings, sleeves and reserved chases in walls and slabs for future busway, water headers and fiber, so the next capacity step is a tap-off rather than a demolition. Reserved space for the CDU gallery, pipe racks, risers and additional electrical rooms, held empty or lightly used against a density step-up. And a documented reservation register recording, bay by bay, what headroom is committed and what is held, so the next team does not unknowingly fill the space the ramp needed.

The economics favor reserving, but price the two halves separately. Knockouts, sleeves and reserved chases are interface reservations that cost almost nothing at design. Over-sized bays and a stronger slab are structural scope that the engineer has to quantify per bay against the local baseline. Price the structural and service-route changes together with the cooling retrofit scoped in Chapter 5.1, the stranded capacity and the downtime from opening an occupied hall. Reserve the headroom you cannot retrofit; defer the spend you can. It is the same reversible-versus-irreversible logic derived in Chapter 5.1, applied to the floor plan.

Choose the building shape from separate indoor and site area schedules, then reserve the CDU gallery, risers, yard and rack-replacement paths against the limits of the equipment you have chosen. Spending the same space twice gives a full-looking plan and a hall that cannot take its next load.