Modular & Prefabricated Construction

By 2026 the binding constraint on an AI build is not chips and not capital — it is time-to-power. A gigawatt of AI capacity is worth on the order of $10–12B/yr in revenue, so landing 200 MW even six months early is worth roughly $1–1.2B (SemiAnalysis / Grid Capacity Intelligence, 2025 — a contested, single-source figure). That number is the entire economic case for prefabrication, and it explains why modular delivery has gone from an edge-and-telco niche to the dominant execution model for new AI capacity. The decision in this chapter is not whether prefab is faster — it almost always is — but how much of the facility you move into the factory, and what that choice costs you in flexibility, density ceiling, and interface risk downstream.

The core idea is mundane and powerful: stick-built construction is a serial process gated by the physical site — you cannot rough-in MEP until the shell is up, you cannot set equipment until the floor is poured, and every trade waits on the one before it in the mud. Prefabrication breaks that chain by running two timelines in parallel: while the site team grades, pours, and erects steel, a factory hundreds of miles away is simultaneously building, wiring, plumbing, and testing the power and cooling modules. The modules arrive substantially complete and commissioned. You have not made any single task faster — you have overlapped tasks that used to be sequential, and you have moved skilled-trades labor from a labor-short, weather-exposed site into a controlled, repeatable production line. The reward is schedule compression of 30–50% and a quality step-change. The price is paid in standardization discipline, transport limits, and the exacting arithmetic of interfaces.

"Modular" is not one thing — it is a spectrum, and where you sit on it is the first and most consequential fork. At one end is the fully integrated all-in-one: a self-contained data center in an enclosure (an ISO container or a purpose-built steel module) with IT space, power, and cooling all inside, arriving as a plug-and-play unit needing only power, water, and fiber at the pad. At the other end is the component-module approach: the building itself is a conventional (often tilt-up or pre-engineered steel) shell, but the high-value subsystems — the power block, the cooling plant, the UPS/battery room, even the MEP galleries — are built as factory skids and PODs and bayed into the shell on site. Most large 2026 AI campuses are not built from all-in-one containers; they are stick-built or pre-engineered shells fed by an aggressive program of component prefabrication — power skids and cooling PODs being the workhorses.

The reason the industry drifted toward component modules at hyperscale is geometric. The all-in-one container inherits the ISO transport envelope — roughly 2.44 m (8 ft) wide, with usable internal width under 2.3 m — which is fine for an edge node or a 100–500 kW micro-site but absurd for a 132 kW NVL72 rack that wants service clearance, a CDU sidecar, and a manifold drop. As one practitioner critique puts it bluntly, ISO containers are a poor data-center solution at AI density precisely because the 8 ft width strangles serviceability. Component skids escape the box: a power skid or cooling POD can be built to a road-legal over-width envelope, trucked as an oversize load, and then bayed together on a pad to form a room of any size. You trade the plug-and-play simplicity of the container for the density and serviceability of a real hall — which, for training-class liquid-cooled white space, is the trade that matters. → Chapter 6.1 (data-hall layout); Chapter 7.13 (the rack as integration unit).

The vocabulary matters because it 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" or "pod") is larger — an enclosed, walk-in volume containing a complete room's worth of function: a UPS-and-battery POD, a cooling-plant POD, 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 a defined set of interfaces.

The economic logic is that these subsystems are where the schedule-critical skilled trades live. Electricians and pipefitters are the scarcest, most expensive, most schedule-determining labor on an AI build, and the skilled-trades shortage is itself a recognized schedule killer (treated as a program risk in Chapter 6.6 and a workforce problem in Chapter 14.11). Building the power and cooling plant on a factory line lets a smaller, stable, trained crew produce many identical skids under cover, with jigs and fixtures, with rework caught at a station rather than at 2 a.m. on a critical-path night. That is the real source of the quality step-change — not that factory workers are better, but that a repeatable line build with in-process QA beats a one-off field build under schedule pressure every time. Vendors have productized this: Vertiv's MegaMod and Schneider's EcoStruxure prefabricated PODs are explicitly marketed as the power-and-cooling plant for NVL72-class halls, sized to 1 MW+ blocks and shipped pre-integrated (Vertiv 360AI reference design; Schneider EcoStruxure, 2024–2026).

Here is the consequence that surprises first-time modular buyers: prefabrication does not remove the lead-time problem, it relocates it. You have traded a site-labor constraint for a factory-slot constraint. Integrator and module-factory capacity is now a strategic variable in exactly the way HV transformer and switchgear lead times already are (the long-lead procurement story in Chapter 2.3). If everyone in a hot market decides to prefab their power blocks in the same quarter, the module factory's line becomes the bottleneck, and your 16-month modular schedule slips back toward the 24-month stick-built one — except now you are also exposed to a single factory's throughput and a longer, more fragile logistics tail.

That logistics tail is fragile. A finished power POD or a factory-integrated NVL72 rack is a 1.5–3 tonne, multi-thousand-connection, liquid-filled object, and the data show most damage occurs not on the long haul but at loading and handoff — the crane lift, the truck-to-pad transfer — where shock and tilt events crack solder joints and unseat connectors (Nefab logistics analysis, 2026). An estimated 30–50% of large 2026 data-center projects face delays, and module logistics is now a top contributor. The prefab supply chain therefore has three coupled long poles you must manage as one system: (1) factory slot — book it 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 still inherit their own lead times (a prefab power skid does not arrive faster than the transformer inside it).

The honest comparison is not "modular good, stick-built bad." It is a multi-axis tradeoff where modular wins decisively on speed, quality consistency, and site-labor reduction, ties or wins on embodied carbon and waste, and loses on peak density, customization, and very-large-span flexibility. The schedule and quality wins are real and quantified: 30–50% faster, factory QA, up to ~90% less site waste, and the skilled-trades load moved off the critical path. The losses are equally real and are where most modular regret originates — a team that prefabbed an inference-grade hall and then needed to land training-class liquid racks discovers the module's frozen geometry cannot absorb the change.

Cost is the most misread axis. Modular is frequently sold as "cheaper," but on a pure $/MW basis a heavily prefabricated build is often at parity or slightly more expensive than stick-built — you are paying for factory overhead, transport, lifting, and the structural steel of the module frame itself. The savings are not in the unit cost of the concrete-and-copper; they are in the time value of money (revenue pulled forward, financing carried for fewer months) and in de-risked schedule and quality. If you score modular on $/MW alone you will reject it; if you score it on NPV with speed-to-power valued at $10–12B/GW/yr, it usually wins. Choose your scoring metric wrong and you optimize for the wrong building.

Every advantage of prefabrication is collected at the moment the modules mate — and that moment is where modular programs fail. A stick-built facility has effectively one continuous structure with no hard interface; 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. The governing rule: the more you prefab, the fewer but more critical your interfaces become, and the more tightly you must manage tolerance stack-up. A 5 mm misalignment that is trivial in a field build can prevent two busbars from bolting or two coolant manifolds from sealing when each was built independently to its own datum.

The mature practice is an explicit interface control document (ICD) per mating plane — defining 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. Ownership of each interface is assigned to a single party; "the gap between scopes" is the single richest source of finger-pointing and schedule slip on a modular job. The corollary discipline is commissioning-in-factory: you pull as much of the Cx scope upstream as the interfaces allow. A witnessed factory acceptance test (FAT) and point-to-point verification on each skid, ideally a partial integrated systems test of a module with its controls, means the on-site Cx window collapses to interface verification and the integrated whole-facility test — not a from-scratch commissioning of every subsystem in the mud. Factory L11 rack integration already compresses floor install-and-commission from ~3 months to ~2 weeks; the same logic applied to the plant is where the schedule win is actually banked. → site Cx sequencing in Chapter 6.6.

The deepest shift prefabrication enables is philosophical: it turns the data center from a construction project into a manufactured product. A stick-built facility is designed once and built once; a modular reference design is engineered once and replicated across sites, generations, and continents. This is the same move that the rack made when integration shifted to the factory at L11 (Chapter 7.13) — the unit of repeatability moved up a level, from the server to the rack, and now from the room to the module. The payoff compounds: every replicated module amortizes the engineering, hardens the design through repetition, shrinks the commissioning learning curve, and lets an operator 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, e.g. Omniverse DSX) are precisely this productization taken to the campus scale — a validated, repeatable blueprint rather than a one-off design.

Standardization is also where the EN 50600 / ISO-IEC 22237 international modular standard family earns its keep: it codifies availability and protection classes at the module level, which is exactly the abstraction a product-ized, multi-jurisdiction rollout needs (EN 50600 / ISO-IEC 22237). The consequence to internalize is a tradeoff between repeatability and 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 very repeatability that justified the product. The discipline is to define a tightly-bounded set of configuration options and refuse one-off redesigns; the failure mode is "productized" modules that are quietly re-engineered for every site until they are stick-built with extra steps.