Modular & Prefabricated Construction

Time-to-power is the constraint that construction choices can still move, when your construction path is the one that governs the date. It is not the only constraint. SemiAnalysis argued in June 2026 that the industry has moved from a power-constrained to an accelerator-constrained regime, and Chapter 2.1 shows how to find the path that governs your project. Where construction does govern, speed is worth real money: landing a 200 MW phase six months early brings forward about $1.2–1.3B of gross revenue (Chapter 2.2 works the estimate). That is not the package’s saving: deduct operating costs, demand risk, the factory premium and any utility gate that does not move. Prefabrication has grown from an edge-and-telco niche into an established delivery option for AI capacity. The decision here is whether prefab advances accepted service, how much of the facility to move into the factory, and what that costs you in flexibility, density ceiling and interface risk.

The core idea is simple. Site-built work follows local predecessors: you cannot set equipment before the floor is poured or rough in MEP before the shell is up. Prefabrication adds a second timeline. While the site team grades, pours and erects steel, a factory builds, wires, plumbs and tests the power and cooling modules. They arrive substantially complete with factory test records, but transport, field joints and the assembled installation still need site tests. No single task gets faster. Tasks that were sequential now overlap, and skilled-trades labor moves from a labor-short, weather-exposed site to a controlled production line. Conventional projects overlap too, through procurement and independent work fronts, so compare the two real dependency networks.

SemiAnalysis models a fully modular 50 MW US liquid-cooled hall as about 36% faster than a pure stick-built one, groundbreaking to IT-ready with permitting excluded. Against today’s partly prefabricated baseline, the same analysis gives about 30%. Those are two separate comparisons, not a range. Your package moves the date only if the factory path clears the site and utility gates that govern it. The price is standardization discipline, transport limits and tight interface arithmetic.

"Modular" is a spectrum, and where you sit on it is the first fork. At one end is the fully integrated all-in-one: a self-contained data center in an ISO container or purpose-built steel module, with IT space, power and cooling inside, needing only power, water and fiber at the pad. At the other end is the component-module approach. The building is a conventional shell (often tilt-up or pre-engineered steel), and the high-value subsystems are built as factory skids and PODs and bayed into the shell on site: the power block, the cooling plant, the UPS and battery room, even the MEP galleries. An owner can pair a conventional or pre-engineered shell with factory-built power and cooling packages, and power skids and cooling PODs are the workhorses of that configuration.

The case for component modules at hyperscale is geometric. The all-in-one container inherits the ISO transport envelope, roughly 2.44 m (8 ft) wide with under 2.3 m usable inside. That suits an edge node or a single 100–500 kW containerized module. It does not suit a GB300 NVL72 rack, which is rated up to 142 kW in NVIDIA’s reference architecture (155 kW peak at Lenovo) and wants service clearance, a CDU sidecar and a manifold drop. At AI density the 8 ft container width strangles serviceability. Component skids escape the box. A power skid or cooling POD can be built to a wider permitted-oversize envelope, trucked as an oversize load and bayed together on a pad into a room of any size. You give up the container’s plug-and-play simplicity for the density and serviceability of a real hall, and for NVL72-class liquid-cooled white space that is the trade that matters. → Chapter 6.1 (data-hall layout); Chapter 7.13 (the rack as integration unit).

The vocabulary maps to the contracting and interface model. A skid is a structural steel base frame carrying a complete functional subsystem. A power skid integrates transformer, switchgear and switchboard; a pump skid carries pumps, valves and controls; a CDU skid isolates the technology-cooling loop from facility water. A POD (sometimes "module") is larger: an enclosed, walk-in volume holding a room’s worth of function, such as a UPS-and-battery POD, a cooling-plant POD or an electrical-room POD. The skid is a component; the POD is a room. Both are built, wired, plumbed and ideally tested in the factory, then craned onto a foundation and connected at defined interfaces.

The economic logic is that these subsystems hold the schedule-critical skilled trades. Electricians and pipefitters are the scarcest, most expensive labor on an AI build, and the trades shortage is a recognized schedule risk (a program risk in Chapter 6.6, a workforce problem in Chapter 14.11). A factory line lets a smaller, stable, trained crew produce many identical skids under cover, with jigs and fixtures, and catch rework at a station rather than at 2 a.m. on a critical-path night. The quality gain comes from the repeatable line, not from better workers. In-process QA catches repeated defects earlier than a one-off field build does, provided the factory’s inspection and rework controls are actually enforced.

Vendors have productized both the plant and the hall, and the products integrate different things. Vertiv’s MegaMod HDX (January 2026) integrates data-hall, power and cooling infrastructure. Its compact version tops out at 13 racks and 1.25 MW, its combo version at 144 racks and 10 MW, with published rack densities of 50 kW to over 100 kW. Schneider’s Prefabricated Modular EcoStruxure Pod (2025) is built to order as white-space infrastructure. It combines liquid cooling, busway, cabling and containment in pods of 1 MW and beyond. Neither headline capacity shows that a layout carries an NVL72 rack. Ask the vendor for the specific layout and its rack-specific electrical and thermal qualification, and take rack power profiles from Chapter 5.1.

Prefabrication does not remove the lead-time problem. It moves it from site labor to a factory-slot constraint. Integrator and module-factory capacity is now a strategic variable, like HV transformer and switchgear lead times (Chapter 2.3 covers long-lead procurement). If everyone in a hot market prefabs their power blocks in the same quarter, the factory line becomes the bottleneck. A factory delay can erase the overlap that justified prefab, and you are then exposed to one factory’s throughput and a longer, more fragile logistics tail.

That tail needs controlled custody. A factory-integrated NVL72 rack must be handled to its OEM logistics sheet, which sets the shipping fluid state, preservation fluid, dry and wet mass, center of gravity, and tilt and shock limits. Loading, unloading and repositioning are the handling points where those limits get tested, so require controlled custody and condition records at each. Nefab, a packaging vendor, says most failures come from handling at those points, not from the long haul (2026). Treat that as the vendor's own assertion: Nefab gives no damage rates. Market-wide delay is a separate matter, and Chapter 2.1 gives the scoped schedule-risk picture. The prefab supply chain has three coupled long poles to manage as one system: (1) factory slot, booked speculatively like a transformer; (2) the haul route, surveyed and permitted before the module’s outer dimensions are frozen; and (3) the upstream components inside the module, which keep their own lead times (a prefab power skid does not arrive faster than the transformer inside it).

The comparison is not "modular good, stick-built bad." It is a multi-axis tradeoff. Modular can win on speed, quality consistency and site-labor reduction. It can tie or win on embodied carbon and waste under matched life-cycle boundaries. It faces a design-specific trade on peak density, customization and very-large-span flexibility, though joined halls are not confined to one transported box. The modeled 36% schedule cut above, and one study of 59 building cases that found about 79% less onsite waste for modular construction, each describe only the scope studied. Your data center gains schedule only if factory work advances accepted service. Carbon and waste comparisons need matched duty, quantities and life-cycle boundaries, including factory and site waste. Factory QA and moving trades into the factory can improve repeatability and relieve site staffing without moving work off the controlling path. The losses are equally real, and most modular regret starts with them. A team that froze its modules at a legacy air-cooled density, then has to land NVL72-class liquid racks for training or inference, finds the frozen geometry cannot take the change. Training and inference differ in fabric sizing, storage path and failure model, not in rack density.

Cost is the most misread axis, because its sign is not settled. SemiAnalysis’s July 2026 model puts a fully modular 50 MW hall about 8% below stick-built on capex per MW. Factory overhead, transport, lifting and the module frame’s steel push the other way, and Schneider’s WP164 models savings from scaling capacity with demand and avoiding overbuilt capacity, which is a different comparison. Installed capex can therefore come out higher or lower with prefab. Ask for installed bids at equal duty and equal scope, with factory integration, transport, lifting, field connections and acceptance all priced in. Value earlier service separately, through incremental contribution and avoided carrying cost at the accepted-service dates. If you import gross AI revenue per gigawatt straight into NPV, you overvalue speed. Pick the wrong scoring metric and you optimize for the wrong building.

Every advantage of prefabrication is collected when the modules mate, and that is where modular programs fail. A stick-built facility has trade and equipment interfaces inside its continuous structure. A modular facility is an assembly of discrete objects that must align mechanically, electrically, hydraulically and in controls, to tolerances set in a factory and verified on a pad. Prefabrication moves integration upstream, but it does not guarantee fewer interfaces: splitting a room, or a pipe or bus run, into transportable modules can add mating joints. In your design, track three things apart: interfaces inside the factory, mating joints in the field, and who is contractually responsible for each. Count the actual joints and owners rather than assuming there are fewer, and hold tolerance stack-up tight at each. A 5 mm misalignment that is trivial in a field build can stop two busbars bolting or two coolant manifolds sealing when each was built to its own datum.

The mature practice is an explicit interface control document (ICD) per mating plane. It defines the exact location, tolerance, torque and acceptance test for every bolt circle, busbar landing, pipe flange, UQD coupling, conduit stub and control termination where one module meets the next or the shell. Assign each interface to a single party; the gap between scopes is the richest source of finger-pointing and schedule slip on a modular job. The corollary is commissioning in the factory: pull as much Cx scope upstream as the interfaces allow. A witnessed factory acceptance test (FAT) and point-to-point verification on each skid, ideally with a partial integrated systems test of a module and its controls, let the on-site Cx window focus on transport-sensitive checks, field interfaces, remaining discipline tests and the integrated whole-facility test. It no longer has to commission every subsystem from scratch in the mud. Factory L11 rack integration already compresses floor install-and-commission from months to weeks, and the same logic applied to the plant is where the schedule win is banked. → site Cx sequencing in Chapter 6.6.

Ask whether the package is certified as an assembly, not only as listed parts. As of September 2026, UL 2140 certifies a whole prefabricated power-distribution assembly and follows up at the factory. The first certification went to specified Schneider skid models. UL 4501 covers eligible cold plates, manifolds and quick disconnects. It checks flow, durability, interoperability, coolant compatibility, and pressure and leak integrity. It covers cooling subassemblies, not a whole cooling installation (requirements in Chapter 5.4). Both are Outlines of Investigation. The submittal should name the exact certified model, its configuration and ratings, and its installation instructions. Check field work against those instructions and any authorized changes; changes beyond them need an evaluation disposition, not automatic loss of the mark. Field-joint verification and integrated site acceptance still apply.

Apply the predecessor method. Module installation: max(18, 20) + 4 = week 24; accepted service: max(24, 22) + 2 = week 26. Site build: 18 + 10 = week 28; accepted service: max(28, 22) + 2 = week 30. Choose the module for four weeks of earlier usable capacity. Those four weeks on site still cover lifting, field-joint inspection, protection tests and control-point verification.

Flip case: if utility is ready at week 30, both options finish at max(installation finish, 30) + 2 = week 32. The utility crossover is week 28. At or beyond it, reject the schedule premium unless separately priced quality or crew benefits repay it. Delivery slipping to week 24 also pushes module installation to week 28 and erases the advantage. Book the factory and haul route against those crossovers before you freeze the box.

Chapter 2.1 sets out the integrated schedule and risk method; 2.4 assigns custody and remedies; 6.6 releases the site package. Value accepted service through contribution and avoided carrying cost, not gross cluster revenue.

The deepest shift prefabrication enables is a change of category: the data center goes from a construction project to a manufactured product. A site-built facility can reuse a standard design, but each one is still executed in place. A modular reference design is engineered once and replicated across sites, generations and continents. The rack made the same move when integration shifted to the factory at L11 (Chapter 7.13). The unit of repeatability went from the server to the rack, and now goes from the room to the module. The payoff compounds. Each replicated module amortizes the engineering, hardens the design through repetition and shortens the commissioning learning curve, and an operator can roll out near-identical capacity in many metros with a known cost, schedule and quality envelope. NVIDIA’s and the hyperscalers’ "AI factory" reference designs, including digital-twin-before-build such as Omniverse DSX, take that productization to campus scale: a validated, repeatable blueprint rather than a one-off design.

Standardization is also where EN 50600 and ISO/IEC 22237 earn their keep. Allocate the project’s facility-level availability and protection requirements to module interfaces, then verify them in the integrated data center. The tradeoff to internalize is repeatability against fit. A productized module is cheaper, faster and more predictable the more sites accept it unchanged, and every site-specific deviation (a climate-driven cooling change, a local code variance, a customer’s bespoke density) erodes the repeatability that justified the product. Define a tightly bounded set of configuration options and refuse one-off redesigns. The failure mode is "productized" modules re-engineered for every site until they are stick-built with extra steps.

Choose the factory scope that advances accepted service through a reserved slot, a passable haul route and tested site joints. Where the common utility gate erases the gain, keep the cheaper alternative that meets the same duty, and spend the premium only on a benefit you have demonstrated separately.