01The categories of modular data centre

Modular data centre coverage frequently collapses two structurally different product categories into one taxonomy. Doing so misreads the vendor field, the incumbent M&A moves, and the point at which cooling integration appears. The clean split is grey space versus white space. Figures 1a and 1b below show each arrangement directly. DD OBSERVATIONIn the vendor field examined here, grey space appears to account for most merchant modular volume. and deep-dive each category; maps the incumbent transactions onto the split.

Containerised all-in-one deployments (Huawei FusionDC1000A, edge-focused OEMs) bundle both categories into one enclosure at smaller scale. They are structurally closer to white-space modular because they contain the racks, at edge scale (hyperscale operations stay on different unit sizes).

The rest of this essay covers each category on its own terms, then reads the incumbent transactions and the value capture through that lens.

A side-by-side look at the categories

The distinction between grey-space skids and white-space pre-fab lies in what they house, where they are located, and who operates them. Grey-space skids are utility-plant assemblies engineered to keep the facility alive. White-space pre-fab is the compute enclosure engineered to house and cool the revenue-generating IT gear.

The categories are compositional. and deep-dive each; and map the transactions and capex mix onto the split.

Where the categories interlock. The boundary between grey and white space is the construction handover. Grey-space skids terminate at large busduct flanges and mechanical flanged pipe headers that feed into the data hall. Inside the hall, the white-space pre-fab takes over at the busway riser and secondary loop manifold, distributing power overhead down to individual rack PDUs and coolant through in-pod CDUs to cold plates on the compute. The white-space pod captures heat at the chip and chassis level and transfers it into the facility cooling system, which may comprise a primary water loop, heat exchangers, chillers, dry coolers, or cooling towers depending on the campus architecture. The grey-to-white boundary is therefore an electrical, mechanical, and controls interface; a single shared thermal loop would be a different architecture. Every interface at the boundary is a candidate for the interface-ownership question Figure 12b traces.

The lifecycle mismatch is why the categories cannot be structurally merged. Grey-space skids are tied to a facility-scale lifecycle (typically 15-25 years for the transformer, switchgear, and UPS core). White-space pods refresh with the IT cycle, which under AI density transitions can run every 3-5 years. Merging both into one physical assembly forces the shorter cycle onto the longer one (stranding capex when compute refreshes) or the longer onto the shorter (delaying compute refresh behind facility upgrade). Neither is acceptable at hyperscale, which is why the split persists.

02Grey-space modular: the power skid

The pre-integrated power skid is a category where merchant modular volume appears to concentrate. The category exists because assembling the full electrical plant on-site (MV/LV transformation, LV switchgear, UPS, batteries, controls) across dozens of instances on a hyperscale campus is slow, labour-constrained, and error-prone. Doing it once at the factory, then repeating that factory-tested unit dozens of times, changes the deployment model.

Contents. A typical Fibrebond or NordicEPOD power skid combines an MV/LV transformer sized to the skid rating, low-voltage switchgear and panel boards with breakers and metering, a UPS system with associated battery cabinets (lithium-ion increasingly, VRLA legacy; some architectures integrate the two, others keep the cabinets separate), a controls and monitoring panel running the BMS integration and protection relays, dedicated HVAC for the electrical room itself, and fire suppression. All of this is assembled and factory-tested before the skid leaves the plant.

Interfaces. The skid exposes an MV feed in (typically 11-33 kV depending on jurisdiction) and an LV busway out (typically 415 V or 480 V) to white-space pods or IT halls elsewhere on the campus. Controls, monitoring, and life-safety supervision channels connect back to the facility BMS and DCIM. The interface set is common across the product class regardless of vendor.

Category leaders. Eaton Fibrebond (Minden, Louisiana; acquired by Eaton for $1.4B, closed April 2025) is a major US manufacturing footprint for hyperscale-scale grey-space deployment. Eaton NordicEPOD (Oslo, Norway; 49% Eaton stake in NordicEPOD AS with CTS Nordics) covers the European hyperscale market with standard 1.7 MW and 2.0 MW configurations and custom designs up to 3.1 MW per Eaton's current product material. Delta InfraSuite grey-space skids, Flex modular power pods, AVK PowerPods, Layer 1 Technologies PowerPods, WB E-Pod, and ePod Solutions each hold portfolios of merchant grey-space modular product. Vertiv's SmartMod PowerMod variants also cover the grey-space case.

Why AI accelerates this category. A hyperscale AI campus can require tens to hundreds of megawatts of electrical capacity, with the largest campuses extending beyond 500 MW. Delivering that MW capacity reliably at the rack requires MV/LV distribution and backup power infrastructure repeated across the campus. Factory-assembled skids convert that from a bespoke construction problem to a repeatable procurement decision. That is one reason modular manufacturing and integration capability has become an acquisition target for incumbent infrastructure vendors.

Transient load response is a first-class engineering justification. AI training workloads create instantaneous step-loads that drive di/dt transients and resonant frequencies through the LV power path. Factory integration delivers the advantage; site-built systems can also be engineered for these transients, with more field work; site-built systems can be, and often are. It is that the vendor can qualify the transformer, conversion equipment, protection, and battery ride-through against a known and repeatable impedance path before shipment, reducing one class of site-integration uncertainty. This is one of the more consequential engineering justifications for pre-fab power skids and rarely appears in vendor marketing.

Battery chemistry and code compliance. Modern skids increasingly ship with lithium iron phosphate (LFP) cells; NMC lithium-ion is the legacy chemistry. LFP generally has favourable thermal-stability characteristics relative to NMC, which can simplify aspects of fire-safety design in containerised skids; the actual NFPA 855 compliance scope remains system- and jurisdiction-specific, depending on cell format, container configuration, detection, ventilation, and applicable code requirements. LFP is heavier per kWh than NMC, so the trade-off shows up in skid weight and transport-corridor headroom.

03White-space modular: the IT-hall pod

The white-space pod category emerged when rack power draws crossed the density regime where cooling loops and rack rows have to be qualified together at the factory; on-site qualification stays slower and less repeatable. covers the density transition and its adjacency to the modular category more broadly.

Contents. A Schneider EcoStruxure Pod or a Vertiv SmartMod Max opens to two rows of IT racks in a hot-aisle-containment configuration, an in-pod CDU sized to the pod's IT load with a secondary water loop pre-plumbed and quick-disconnect connectors on the cold-plate side, an overhead busway with per-rack drop-outs, structured cabling infrastructure including fibre trunks and copper ladders, and an integrated leak-detection system with drip pans, conductivity monitors, and rapid-isolation controls.

Interfaces. The pod exposes an LV busway feed in, water supply and return lines to the facility water loop, fibre and controls trunks, and a leak-detection alarm channel. The cooling interfaces (pressure ratings, pipe diameters, quick-disconnect specifications) are typically proprietary to the vendor, which is why multi-vendor pod deployments need a documented cooling-interface schedule at the operator specification stage.

Category leaders. PUBLIC FACTSchneider EcoStruxure Pod, jointly launched with Compass Datacenters on September 9, 2025, is described in Schneider product material at up to 42 racks per pod at 132 kW per rack; actual pod power depends on configured rack count and rack-level density. The pod supports hybrid liquid-air cooling architectures and includes a technical water loop pre-plumbed for direct-to-chip integration. PUBLIC FACTVertiv SmartMod Max supports up to 200 kW; Max DX supports up to 26 racks and Max CW up to 24 racks. Huawei FusionModule2000 and FusionDC1000B / 1000C integrate rack rows, in-row cooling, intelligent power distribution, and integrated leak detection. Delta InfraSuite Prefabricated AI MDC (launched at COMPUTEX 2026) integrates 800 VDC in-row power with GoCool 260 kW LTA and 3 MW LTL cooling. Rittal offers white-space kits with modular CDUs scaling from tens of kW to over 1 MW per unit.

Why cooling appears here and not elsewhere. Cooling integration is a defining feature of white-space pods because the pod contains the racks and their heat load. In grey-space skids, the only cooling is skid HVAC for the electrical room itself. Cross-category matrix compatibility is worth reading directly.

Engineering shifts that drove the white-space pod category.

• Factory-integrated secondary fluid networks. High-density pods increasingly leave the factory pre-plumbed with stainless steel or specialty-thermoplastic headers, flow-balancing valves, and blind-mate dripless quick-disconnects (QDs) ready to hook up to external CDU distribution loops. On-site plumbing scope collapses.
• Overhead infrastructure consolidation. Modern pods route both overhead tap-off busway (415/480 V AC now, 800 VDC increasingly) and fibre trays through unified top-of-rack structural trusses, eliminating field-erected trapeze rigging and reducing the trade coordination load.
• Adaptive liquid-ready containment. Pod vendors are standardising on containment blocks that deploy with air cooling today and accept drop-in in-row CDUs and rear-door heat exchangers without structural rework as density rises. This decouples the pod purchase decision from the cooling architecture commitment.
• Seismic and floor-loading ratings. AI clusters (multi-rack NVLink and Infiniband domains) raise floor-loading and lateral-force ratings materially. Modern pod floors and castor-guide track systems specify for higher dynamic rolling loads per rack footprint than pre-AI containment products supported.

In-pod CDU versus central perimeter CDU. The branches below hold different rack-footprint and operational trade-offs (Figure 3b). In-pod CDU (Vertiv SmartMod Max CW, first-generation Schneider EcoStruxure Pod) consumes 1-2 rack slots and caps at the unit rating; central perimeter CDU (~1-2 MW class and above) leaves the pod as a manifold-only assembly and preserves every rack footprint for compute.

AC busway versus DC bus at the grey-to-white handoff. At 100-132 kW per rack, the current handled by a 415/480 V AC distribution architecture becomes materially larger than in an 800 VDC architecture, increasing conductor cross-section, busway weight, and tap-off burden on overhead pod trusses. Figure 3c shows what a wider 800 VDC deployment would change inside the grey-space skid topology.

Fabric topology dictates pod geometry. Pod dimensions (rack count, aisle spacing, cable-tray height) are set by network fabric topology and signal integrity (mechanical airflow or pipe friction stays a secondary input). Figure 3d shows why pod sizing clusters at 8, 16, and 32 racks.

Volumetric pods versus flat-pack cassette pods. The split runs on logistics:

Cooling-architecture integration summary (Figure 4 covers compatibility matrix). Direct-to-chip liquid cooling integrates through a factory-plumbed secondary loop, a pod-sized CDU (or manifold to a central CDU), cold-plate quick-disconnects, and rack + pod leak detection. Rear-door heat exchangers replace the standard rack door and reject heat into a facility water loop with bolt-on compatibility across every modular class. Immersion is a different form factor with compute submerged in dielectric fluid; modular immersion halls (Submer, Green Revolution Cooling reference designs) constitute their own product class; pod-plus-hall architectures stay DTC-centric.

04Why modular now

The AI buildout is pulling modular procurement into mainstream hyperscale delivery through four converging pressures. Naming them separately helps avoid the trap of treating cooling as the sole driver.

1. Time-to-market compression. Traditional field-built power and IT-hall infrastructure sequences steel erectors, cable technicians, electricians, mechanical pipefitters, and controls integrators across the same slab. Prefabricating shifts that coordination into controlled factory conditions. A hyperscale AI campus requires tens to hundreds of megawatts of electrical capacity, with the largest campuses extending beyond 500 MW. Assembling the electrical plant in parallel at factory scale is materially faster than sequencing dozens of trades on the same site.

2. Factory-tested reliability envelope. Reliable power at the rack requires transformer, switchgear, UPS, battery, and protection coordination all tuned together. When the operator specifies a factory-tested unit, the reliability range is inherited from the vendor qualification programme; on-site commissioning reconstruction stays slower. Underwriting shifts from site-specific unknown to vendor-tested known. This is the core power-side driver, and it is why grey-space skids attract the merchant volume.

3. Skilled labour shortage in field trades. Secondary and tertiary data-centre markets (Iberia, Nordics, US Midwest, tier-2 APAC, sovereign programme sites) face acute shortages of qualified electrical and mechanical trades. Factory-assembled units convert a skilled-labour problem at the site into a manageable installation-and-commissioning problem. The bottleneck affects enterprise and mid-tier operators as much as hyperscalers, and appears to be intensifying through 2027-2030; easing of the constraint stays absent from current signals.

4. Standardised reproducibility (design one, build many). Hyperscale campuses require identical pod and skid layouts across regions to simplify DCIM integration, structural loading, fire suppression zones, and maintenance operations. Reference designs propagate: what a hyperscaler standardises becomes accessible to the mid-tier and enterprise field through vendor product portfolios. This is the mechanism that makes hyperscale-grade modular economically reachable outside hyperscale operators ( ).

The white-space cooling adjacency. At 100 kW-plus per rack, direct-to-chip liquid cooling becomes practical and, at the high end, necessary. Racks and cooling then have to be qualified together, which is why the white-space pod category grew. 100 kW/rack functions as a practical marker for the density regime; a hard engineering threshold would be a different claim; the relationship is continuous.

Interconnection remains the multi-year critical path. Modular does not compress utility interconnection queues. ERCOT is redesigning its large-load interconnection process against a pipeline of more than 438 GW of tracked large-load requests, nearly 89% from data centres, with the Batch Zero process introduced in 2026 to evaluate qualified projects at or above 75 MW by study maturity and commitment criteria. In constrained markets, ground-break to revenue-ready can still run 24-48 months or longer even when modular hardware is fabricated in months.

05The modular build timeline

Modular compresses construction time. It does not compress energisation or utility interconnection. Ground-break to first MW of commissioned electrical capacity is a shorter interval than ground-break to revenue-generating IT load. The three intervals respond to different levers.

05bBoundary friction: where the composed campus fails at the flange

The composed campus fails during construction and commissioning at specific boundaries where vendor scope explicitly ends. Naming them together helps the operator specification stage; each gets a longer treatment downstream.

1. The civil boundary, where concentrated rolling weight meets slab-loading limits. A pod populated with fully-equipped liquid-cooled AI racks can push rack-point and concentrated floor loads far beyond legacy raised-floor designs. Depending on the rack and cooling configuration, individual rack footprints can approach the tonne-plus range, materially higher than air-cooled predecessors. Legacy 24-inch raised floors may require reinforcement or replacement; heavy-duty raised-floor systems vary substantially in load rating and support configuration. Modern deployments increasingly go directly on reinforced slab-on-grade with embedded steel channels, or on dedicated heavy structural skids. Retrofit deployments treat floor rebuild as a first-day scope item. Deep-dive.

2. The hydronic boundary, where the primary facility loop hands off to the secondary technology cooling system. Grey-space skids terminate at the primary loop side, in carbon-steel pipe holding raw chilled water or condenser water. White-space pods hold the technology cooling system (TCS) in stainless steel, typically deionised water or ~25% propylene glycol, with treated biocides and 50-micron secondary filtration, feeding into nickel and copper microchannel cold plates on the GPUs. Chemical passivation of the secondary loop, particulate certification, and helium leak testing before pod energisation are procurement requirements. Deep-dive.

3. The electrical boundary, where AC busway bulk gives way to DC busway weight and the safety profile changes. A 415 or 480 V three-phase busway at 100-132 kW per rack requires copper cross-sections and tap-off cables that stress overhead pod trusses on weight and bend radius. The shift toward 800 VDC (bipolar ±400 V or 0-800 V configurations, per OCP LVDC and Mt Diablo work) at the grey-to-white handoff reduces current for a given power transfer, shrinks conductor weight, and changes the arc-flash, touch-potential, and protection coordination scope. Skid switchboard termination and busway head-end tap-off are separate warranty stops. Deep-dive and the DC-DC Transition series.

4. The logistical boundary, where road-haul constraints bound modular unit sizing. Volumetric pods wider than roughly 3.65 m or taller than roughly 4.1 m trigger wide-load escorts, permits, and route surveys, which is what pushes larger deployments toward flat-pack cassette architectures. The permit wall is why some hyperscaler powered-shell campuses standardise on flat-pack cassette pods that erect on-site in roughly 72 hours, at the cost of more on-site scope than a volumetric pod would hold. Deep-dive and .

These four boundaries also determine which vendor category attracts which acquisition capital, which covers next.

06The incumbents are moving up the integration layer

The 2024-2026 window saw five major transactions along the integration layer. Each pursued the buildout through a different transaction structure, and each transaction maps directly onto the grey-space and white-space categorisation.

Eaton + NordicEPOD (grey-space). PUBLIC FACTIn May 2024, Eaton took a 49% strategic minority stake in NordicEPOD AS, a Norwegian JV with CTS Nordics that designs and assembles standardised power modules for European data centres. The 2024 NordicEPOD configurations described by Eaton supported up to 2 MW of electrical power per unit. Eaton's current NordicEPOD product material describes standard 1.7 MW and 2.0 MW configurations, with custom designs up to 3.1 MW. Eaton and NordicEPOD announced the official opening of the Oslo factory on September 26, 2024. Grey-space category, European hyperscale market.

Eaton + Fibrebond (grey-space). PUBLIC FACTEaton announced the Fibrebond acquisition on March 11, 2025, and closed on April 1, 2025, for $1.4 billion in announced purchase consideration. Fibrebond generated approximately $378 million in revenue and approximately $110 million in adjusted EBITDA for the twelve months ending February 2025 per Eaton's acquisition disclosure. Grey-space category, US hyperscale manufacturing footprint. Between NordicEPOD and Fibrebond, Eaton assembled a substantial grey-space manufacturing footprint across EMEA and North America.

Schneider + Compass Datacenters (white-space). PUBLIC FACTOn September 9, 2025, Schneider Electric and Compass Datacenters jointly announced a prefabricated modular EcoStruxure Pod for white-space fit-out. Schneider describes configurations up to 42 racks at 132 kW per rack, with hybrid liquid-air cooling and a technical water loop pre-plumbed for direct-to-chip integration; actual pod power depends on configured rack count and rack-level density. Schneider began manufacturing the pods for Compass in fall 2025. White-space category, delivered at hyperscale volume through the operator partnership.

Vertiv SmartMod (both categories). PUBLIC FACTVertiv has expanded the longer-running SmartMod family with higher-capacity configurations. SmartMod Max supports up to 200 kW; Max DX supports up to 26 racks and Max CW up to 24 racks. SmartMod Combo extends the architecture to larger multi-module deployments. SmartMod Max CW was originally introduced in 2023, so the 2024-2025 window represented product-line expansion; initial launch dates fall earlier in the catalogue. Vertiv covers both categories through variants.

Vertiv + PurgeRite (cooling services adjacency). PUBLIC FACTOn November 4, 2025, Vertiv announced its intent to acquire PurgeRite, and completed the acquisition on December 4, 2025 for approximately $1 billion in cash plus up to $250 million in additional consideration based on 2026 performance metrics. PurgeRite provides specialised fluid management services (mechanical flushing, purging, filtration) for liquid cooling systems. PurgeRite is a lifecycle-services business; the modular product line stays with SmartMod and Motivair. The acquisition runs alongside the SmartMod product family and extends Vertiv into the ongoing thermal-management services layer that operating liquid-cooled facilities require. Not a modular acquisition in itself; it complements the white-space pod product family with services capability.

Ecolab + CoolIT (white-space thermal). PUBLIC FACTEcolab completed its acquisition of CoolIT Systems for approximately $4.75 billion on July 2, 2026. Unlike the power-side transactions, this is a direct bet on the liquid-cooling interface itself: CDUs, cold plates, and associated thermal infrastructure connect at the pod-to-facility water interface that Figure 4 identifies. Ecolab described the CoolIT acquisition as the platform for an AI-cooling business targeting ~$4 billion by 2030 (per Ecolab IR release). This is the clearest recent transaction showing the white-space thermal interface has become an acquisition target of its own, and it belongs alongside the 2024-2025 power-side transactions; the forward consolidation frontier is a different timeline.

What each transaction targets. Reading the five transactions together sharpens the pattern. Eaton (Fibrebond + NordicEPOD) solidifies the grey-space MW engine of factory-built power modules at hyperscale volume. Schneider (EcoStruxure Pod with Compass) captures the prefabricated white-space integration layer through a hyperscale operator partnership. Vertiv (SmartMod + PurgeRite) covers both categories on the product side and layers on the operational bottleneck of fluid flushing, filtration, and commissioning services on the services side. Ecolab (CoolIT) buys the white-space fluid interface and its consumables and services stream directly. The five transactions cover the modular product value chain across grey and white space, add operational-services capability, and stake out the thermal interface as a direct target.

Hyperscaler internal programmes. Microsoft has publicly described its Fairwater AI datacentre architecture. Meta contributes open infrastructure specifications through the Open Compute Project. Google publishes detailed TPU system and pod architectures. These programmes give hyperscaler architecture teams significant specification power over the modular products they procure from incumbent vendors.

07The manufacturing footprint

Where modular is built matters as much as who builds it. Manufacturing location determines transport constraints, tariff exposure, labour cost, factory lead time, and the physical dimensions the vendor can ship in one piece.

Tariff and trade dynamics. US manufacturing incentives, domestic-content requirements, and customer procurement preferences can affect the delivered economics of modular electrical equipment sourced for domestic AI campuses. Specific IRA credits apply to defined equipment categories and qualification mechanisms; AI infrastructure generically does not qualify. In Europe, CBAM's definitive regime began in 2026 and covers specified sectors including iron and steel, aluminium, cement, fertilisers, electricity, and hydrogen. Where imported modular construction inputs fall within covered steel or aluminium categories, embedded-carbon costs can affect the delivered economics of the module into EU projects.

Regional labour cost differential. Modular assembly labour is a material share of total module cost. Regional labour costs are one input into modular assembly economics. The resulting trade-off across the vendor field is between larger modules that amortise factory labour and logistics over more MW, and smaller or more standardised modules that support higher production throughput.

Transport-corridor design as engineering constraint. Manufacturing location bounds what dimensions the vendor can ship in one piece. North American transport corridors impose route-specific height, width, length, and weight limits, with larger modules often requiring specialised permits, escorts, and route surveys. European road corridors face similarly route-specific constraints, often favouring smaller shippable modules where access roads, bridges, or urban geometry limit transport footprint.

Upstream supply-chain constraints modular procurement cannot solve. Large power transformers (LPTs) at 2.5-4 MVA class face lead times of roughly 80-160 weeks driven by grain-oriented electrical steel (GOES) shortages (reaching back into iron-ore and cold-rolling capacity in China, Japan, Korea, and India), core-winding bottlenecks, and test-bay queues at the transformer OEMs. Low-voltage side, high-interrupt breakers (4,000-5,000 A frame sizes with 100 kA interrupt ratings) run 40-60 weeks and longer. A pre-integrated power skid still contains a full-size MV/LV transformer and full-rated LV switchgear. Packaging those components onto a skid in Louisiana or Norway does not bypass the upstream physics: modularity compresses on-site scheduling risk, not macro BOM procurement lead times. Domestic-content requirements (US IRA-adjacent procurement preferences, EU CBAM effects on imported components, sovereign-programme localisation rules) restructure the delivered economics further.

Maintenance and field serviceability (mean-time-to-repair). Factory integration improves initial QA and deployment speed but introduces field-service complexity. In a traditional stick-built electrical room, a failed 3,500 kVA dry-type transformer or a 4,000 A breaker at Year 7 or Year 10 rolls out along a wide maintenance corridor with overhead monorail hoist access. Working clearances around switchgear and battery compartments in a stick-built room are typically comfortable above NFPA 70E and NEC minimums. In a tightly packed modular enclosure, working spaces are engineered to close to the bare minimum for those same NFPA 70E and NEC clearances, and replacing the same component often requires unbolting outer weatherproof panels, staging exterior mobile cranes, and taking an entire modular bay or adjacent bus duct offline for the duration. For Tier III or IV operations, tightly coupled modular geometry can trade upfront build velocity against 15-plus years of operational maintenance friction. Operators procuring modular should evaluate the vendor's regional service footprint, spare-parts availability, hot-swap architecture, and maintenance-clearance drawings as first-class procurement inputs, alongside capex and delivery timeline.

Structural floor loading and the civil interface. A high-density AI pod populated with liquid-cooled iron can drive per-rack footprint loads beyond the rating of some legacy raised-access floors, requiring reinforcement or a slab-based installation. White-space pods increasingly go on reinforced slab-on-grade with embedded steel tracks, or on dedicated heavy structural skids. Retrofit deployments often need floor rebuild before pod placement, which is a hidden cost in the retrofit-vs-greenfield decision. Volumetric pods wider than roughly 3.65 m or taller than roughly 4.1 m typically trigger wide-load escorts, route surveys, and specialised transport that flat-pack cassettes can avoid at the cost of more on-site erection scope.

08The vendor field, categorised

Reading the vendor field is easier when categorised by the layer of the modular stack each player addresses. The revised categorisation makes the grey-space and white-space split explicit.

09The campus composes multiple vendor products

The essay's central point for procurement: a hyperscale AI campus is composed of products from several vendors across the categories in Figure 1. Figure 10 illustrates one four-vendor composition; the point is that compositional procurement across categories describes the campus better than a specific vendor count.

The interface question that Figure 10 raises. When Eaton delivers a power skid, Schneider delivers a white-space pod, and Compass delivers the shell, no single vendor has warranted the interface between them. Insurance-stack thinking, commissioning discipline, and interface qualification pick up from here. See for the detailed treatment.

10Prefab product categories with capex and timeline ranges

The three main modular product categories with typical capex and timeline ranges. Denominators (MW electrical vs MW IT vs MW facility) are made explicit because they measure different quantities.

Capex and timeline ranges are AUTHOR SCENARIOS synthesised from publicly disclosed vendor configurations, published operator commentary, and anonymised DD engagements 2024-2026. Denominators and scope are stated explicitly in each cell to prevent apples-to-oranges comparison. Actual project economics vary substantially by scale, cooling configuration, jurisdiction, and interconnection queue position.

11A worked example. Illustrative 100 MW AI campus

AUTHOR SCENARIOTo make the compositional procurement thesis concrete, consider one illustrative 100 MW facility / approximately 80 MW IT-load AI campus in a constrained ISO. The facility total covers utility interconnection and site infrastructure; the 80 MW is what reaches the IT racks after conversion, cooling, and support-load overhead. The 80 MW IT figure represents an illustrative 80% facility-to-IT allocation for this worked example. The split covers UPS losses (and does not represent a PUE claim), distribution losses, cooling plant, pumps and fans, controls, and miscellaneous facility load. The bill of materials below is a synthesised range built from vendor pricing, published operator commentary, and anonymised DD engagements 2024-2026. It is illustrative; empirical measurement at this granularity requires different source data.

Ranges reflect variance in cooling architecture selected, geographic jurisdiction, interconnection deposit terms, and vendor sourcing strategy. If the IT load were entirely configured around GB300 NVL72 racks at the published maximum rack power, roughly 560 racks and 40,000 GPUs would fit inside the 80 MW budget. This is an illustration only. A real campus mixes compute, storage, networking, and non-GPU loads, so the GPU count is lower and rack density varies materially by workload.

Observations from the illustrative campus.

1. Modular vendor spend concentrates in the grey-space, white-space and cooling layers. Grey-space power skid + white-space pod + cooling infrastructure ($530M-$1.05B combined range) accounts for a large share of the modular infrastructure procurement spend.
2. Land + utility + substation offers no modular compression. This portion is entirely site-specific; modularity does not touch it.
3. Commissioning is roughly 5-6% at the midpoint. It spans a wider band across the low and high ends of the scenario, and is material either way for the skilled-labour pipeline.
4. Interface qualification stays outside the discrete-line-item treatment. Vendor-to-vendor interface qualification gets rolled into owner engineering and integrator soft costs; a separate procurement line stays unusual.
5. IT compute dominates total campus capex. Modular procurement optimises the infrastructure range surrounding the IT; the IT compute itself stays unoptimised by this layer.
6. Timeline runs 12-18 months to first MW under favourable conditions. Full 100 MW extends to 24-36 months or longer depending on interconnection queue position.

12Residual value, financing, and the secondary market

Modular units hold economic properties that stick-built shells do not. The comparison below runs across the procurement and financing dimensions where the two build types diverge.

The residual-value story is one reason some financing structures for modular procurement look more like equipment leasing and less like traditional data-centre project finance. It is also why some operators explicitly maintain vendor relationships across upgrade cycles; discrete-event procurement loses the install-base continuity.

13The residual constraints of modular procurement

Modular procurement introduces its own risks that operators need to price:

Transportation constraints. Bridge heights, road weight limits, tunnel clearances, and rail loading gauges bound how large any single shippable unit can be. Fibrebond's Louisiana manufacturing footprint supports certain module dimensions well; regional differences in road, rail, and bridge infrastructure feed back into how vendors size their standard configurations.

On-site staging complexity. Modular assembly is materially less complex than a fully stick-built campus, where dozens of trades sequence on the same slab; plug-and-play assumptions overshoot the current state. For example, an illustrative campus receiving dozens of skids plus multiple white-space pods plus multiple hall sections still needs staging area, crane access, coordination between modular vendor installers, and interface management. The compression comes from moving fabrication upstream, not from eliminating the assembly discipline; operators that read modular as plug-and-play typically encounter the interface layer at commissioning.

Regulatory, AHJ, and jurisdictional testing friction. Factory qualification (hipot testing, breaker injection, helium leak testing, arc-flash coordination) generates witness-test documentation that vendors present as evidence of site-ready equipment. Local Authorities Having Jurisdiction (AHJs) and utility inspectors frequently do not accept factory-stamped witness tests at face value. Depending on jurisdiction, they may require complete field re-testing, third-party field evaluations (UL/ETL field evaluations or equivalent regional accreditation), and local trade or union inspection before energisation. Common friction points in practice are (1) field labelling and NRTL listing on modified assemblies inside the skid, which can trigger an expensive third-party Field Evaluation if the whole-skid label is missing, and (2) grounding and protection coordination, where ground grids, earth-fault protection curves, and short-circuit withstand cannot be finally certified until the skid is bonded to the specific site MV utility characteristics and soil resistivity. The result is that some of the factory-commissioning gain gets consumed at the site by regulatory friction the vendor cannot solve for the operator. Operators should include AHJ acceptance testing scope in the procurement specification and budget the calendar time separately from the vendor commissioning window.

Interconnection queue. The utility interconnection queue for large data centre loads runs 24-48 months or longer in constrained ISOs. ERCOT and PJM ( ) are the North American cases. In Europe the queue map is more fragmented: ENTSO-E coordinates cross-border planning, but national TSOs (National Grid ESO in the UK, TenneT in the Netherlands and Germany, Amprion in Germany, Elia in Belgium, EirGrid in Ireland) each run distinct large-load connection processes with distinct queue times. Ireland's data-centre grid moratorium and Germany's Stromnetzausbau constrain volume differently again. The Nordic grid plus Nord Pool operates in yet another regime. For a modular operator sourcing across jurisdictions, the interconnection queue map is a first-order procurement input. A modular campus that assembles in 12 months and waits multi-year for interconnection has not saved go-live time. Sequencing the interconnection application ahead of the modular hardware order is what lets operators run the workstreams in parallel.

Retrofit versus greenfield. Modular procurement is dominantly a greenfield play, but the retrofit-vs-greenfield decision framework in the DC-DC Transition series covers cases where power skids and white-space pods drop into an existing shell that was previously stick-built. The retrofit-modular workflow trades module dimensions and interface constraints against the utility interconnection that the retrofit site already has. That trade works primarily where interconnection is scarce and the shell can accept the module dimensions without expensive rework.

Vendor lock-in on module spec. Once a campus is built around a specific vendor's skid or pod dimensions, cooling interface, and control system, expanding capacity typically means procuring from the same vendor. Operators should explicitly underwrite supplier longevity, service coverage, and replacement-path risk where a modular design creates vendor dependence. Dual-sourcing mitigates at the cost of interface complexity.

Commissioning skills. Prefab reduces on-site commissioning hours but does not eliminate them. On-site electrical, mechanical, cooling, controls, and integration commissioning still require skilled commissioning personnel. The DC-DC Transition series essay 12/13 covers the retraining requirement for 800 VDC work; the same constraint applies to liquid-cooling commissioning in modular halls.

14Who captures the value

Modularisation shifts value from individual equipment supply toward whoever owns the interfaces, qualification programme, and repeatable reference design. Figure 12 traces where in the modular product stack the vendor value gets captured and why that makes integration capabilities strategically valuable.

A short list of physical disconnect points produces most warranty and commissioning disputes. The mechanical boundary is the dielectric isolation flange between facility primary piping (typically carbon steel) and pod secondary piping (stainless 304 or 316); the flange handles thermal expansion, galvanic isolation, and pressure-rating hand-offs at the pod ingress. The electrical boundary is the busway head-end tap-off versus the skid switchboard termination; who owns the connector and its coordination study is a first-day question. The control boundary is the protocol translation between in-pod leak-detection and controls (typically BACnet or Modbus) and the campus EPMS/DCIM; this is where alarm mapping, priority handling, and life-safety integration either work or fail. Each of these is a physical hand-off where vendor scope explicitly ends and operator scope begins.

15Illustrative 2030 procurement mix and what to watch

Rather than pseudo-precise share percentages, the useful framing is which trajectories accelerate or decelerate the value-capture stack in Figure 12.

AUTHOR SCENARIODirectional analysis, not a market forecast. Each row states one plausible trajectory and the specific factors that could bend it.

The pattern the trajectories describe: integrated incumbents that own the interface layer can capture more of the content-per-rack expansion across both grey and white space. Operator-integrated platforms hold volume-commit leverage. Hyperscaler internal programmes set specification, which pulls the incumbents toward standardised reference designs. The value capture pattern in Figure 12 is what gives the segmentation its direction.

16The operator model post-modularisation

Modularisation restructures what the data centre operator role covers. The traditional operator model runs across design, procure, build and operate. The modular operator model runs across a different four: specify, assemble, commission, and operate. The shift is from a design-and-build role to a specification-and-integration role.

Colocation operators. For pure colocation platforms (Compass, Vantage, DataBank, DigitalBridge), the shift makes them skilled interface-buyers; deep design shops stay with other operators. The Compass and Schneider partnership around the EcoStruxure Pod is the clearest shipping example.

Hyperscaler-owned campuses. Microsoft, Meta, and Google run in-house architecture programmes that retain specification power. In the modular operator model, hyperscaler architecture teams can define the reference facility and specify the modular products assembled into it.

Margin implications. The modular operator model can be margin-accretive if the operator retains specification power and captures the interface-qualification value. It can be margin-dilutive if specification power is held by a customer (hyperscaler) and the operator is compressed to a pure assembly role. Operators most exposed to hyperscaler specification power face the strongest margin compression as modular procurement matures.

Skills pipeline. The modular operator model requires different skills than the traditional model. Design and drafting functions shrink. Commissioning and interface qualification grow. Operators that under-invested in commissioning capacity during the pre-AI era typically find themselves rebuilding that capacity now.

17The market-operator layer

The two-category thesis interacts with four operator tiers, each of which uses modular products differently and captures different value from them. The tiers are worth naming because vendor commentary frequently conflates hyperscaler procurement with the broader market.

Hyperscalers (Microsoft Azure, AWS, Meta, Google, ByteDance, Alibaba) run internal architecture programmes and consume modular product as assembled components against their own reference designs. Their specification power restructures what vendors produce. Microsoft Fairwater architecture, Meta OCP contributions, and Google TPU pod designs each pull incumbent vendors toward hyperscaler-audited configurations; vendor-native reference designs lose share here.

Colocation platforms (Compass, Vantage, DataBank, DigitalBridge, Equinix, Digital Realty) operate between hyperscaler specification and vendor manufacturing. Some, like Compass with Schneider, co-engineer modular products. Others procure through their own construction programmes. The colocation platform tier is the one most structurally affected by the operator-model shift covered.

The colocation multi-tenant paradox. Grey-space skids are universally adopted across operator tiers because power is commoditised in kVA and MW. Colocation platforms serving multiple tenants struggle to deploy rigid white-space pre-fab pods because each hyperscale or enterprise tenant brings distinct rack dimensions (48U versus 52U, standard versus wide), varied cold-plate supply temperatures (ASHRAE W27 to W45), and proprietary cabling architectures. The observable pattern is that grey-space modular is adopted everywhere, while white-space modular is dominated by single-tenant hyperscaler builds and operator partnerships (Schneider with Compass) where the operator can standardise across pods. Multi-tenant colocation operators often deploy grey-space modular fully and stick-build or semi-modular the white space to preserve tenant flexibility.

Mid-tier enterprise and industrial operators (Fortune 500 with on-prem AI, financial services, healthcare, defence contractors) increasingly access hyperscale-grade modular products through the same reference designs that hyperscalers use. The reference-design portability that hyperscalers created makes enterprise-scale modular procurement possible without a bespoke design programme. This is a plausible 2026-2030 growth pattern that hyperscaler-focused commentary tends to miss.

Sovereign AI programmes (India's IndiaAI Mission, UAE's G42, Saudi Arabia's HUMAIN and Alat, Singapore's National AI programme) commission modular data centres at scale to support domestic AI development. Several of these programmes are commissioning or evaluating modular approaches where speed, repeatability, and localised manufacturing are important. The sovereign volume is a growth vector beyond the traditional hyperscaler-plus-colocation market and is one reason the manufacturing footprint evolves toward more localised production over 2027-2030.

Standards fragmentation is now a first-class procurement input. Open Compute Project (OCP) remains the North American and Western European reference. Open Data Center Committee (ODCC) is the Chinese domestic reference, with specifications increasingly divergent from OCP. Nordic bodies (CTS Nordics, Nordic Data Center Association) codify a distinct Nordic hyperscale reference. Vendors that can produce multiple variants at scale handle this fragmentation; it is a real constraint for smaller specialists.

18Commissioning, warranty, and skills

The modular procurement model concentrates specification and factory testing but expands commissioning and interface qualification. Four dimensions of that expansion.

Interface qualification across the composed campus. When Eaton delivers a power skid, Schneider delivers a white-space pod, and Compass delivers the shell, no single vendor has warranted the interface between them. The operator handles interface qualification risk unless it explicitly contracts otherwise. Insurance-stack thinking matters here. What does the equipment warranty cover, what does the general-contractor warranty cover, what does the commissioning-agent liability cover, and what remains uninsured? Most operators encounter this gap on their first fully-modular deployment.

Factory QA and digital twin. Modular's factory-test advantage is a digital-twin story. When the vendor factory tests a unit, the digital twin from that test can travel with the unit and inform on-site commissioning. Schneider, Vertiv, Delta, and Rittal each publish digital-twin capabilities for their factory QA programmes. The operator benefit is that commissioning validates against a digital baseline; design intent alone stays weaker as a validation reference. Adoption is uneven across operators, and the operators that build this into procurement contracts run cleaner commissioning cycles.

OT/SCADA cybersecurity. Modular products ship with vendor-managed controls, firmware, and remote-diagnostics capabilities. This creates OT/SCADA cybersecurity exposure that stick-built facilities may not. IEC 62443 compliance, factory-managed patch delivery, and secure-boot on skid controllers are becoming procurement requirements at the hyperscaler tier and are propagating downward. This remains a real gap in most enterprise-tier procurement specifications.

Skills pipeline. The modular operator model needs different skills. Interface qualification, cooling-loop commissioning, DC protection coordination testing, factory-to-site digital-twin reconciliation. These skills differ from traditional facility engineering. Operators that under-invested in commissioning capacity through the 2010s are rebuilding it now. The commissioning skills bottleneck is a genuine 2027-2030 constraint. Industry associations (Uptime Institute, DCPro, AFCOM, BICSI) have been slow to update apprenticeship and certification frameworks. Some operators run in-house academies to close the gap; the ones that do treat it as a capex line; a training-expense treatment stays undercapitalised.

Hydro-commissioning and water quality. The commissioning boundary between grey-space chilled-water loops (raw CHW or condenser water) and white-space technology cooling systems (TCS, typically a 25% propylene glycol solution) is where liquid-cooling deployments most commonly stall. Chemical passivation of the secondary loop, particulate filtration (50-micron and tighter), and helium leak testing before pod energisation are procurement requirements; commissioning nice-to-have treatment stays under-specified. ASHRAE TC 9.9 sets the reference framework; specialist fluid-management services (which is what Vertiv acquired in PurgeRite) handle the flushing, chemistry, and testing scope that operators would otherwise have to build in-house.

The 800 VDC transition compounds the commissioning problem. The DC-DC Transition series essay 12/13 covers the retraining requirement for 800 VDC work in detail. In the modular context, the practical consequence is that a factory-tested 800 VDC skid still needs commissioning personnel who can validate DC protection coordination, arc-fault detection, and battery integration at the site. This is a specific hiring problem for operators moving to 800 VDC native architectures over 2027-2030.

19The consolidation frontier

The 2024-2026 incumbent transactions concentrated on grey-space power skids, prefabricated white-space pods, and (with Ecolab/CoolIT) the white-space thermal interface. Several specialist categories occupy adjacent positions to those layers and could plausibly attract the next round of integration-oriented transactions.

Coolant distribution unit (CDU) manufacturers (white-space adjacency). Beyond CoolIT ( ), Motivair, USystems, and Boyd Corporation are among the specialist CDU vendors operating at this interface. As DTC becomes native to prefabricated pods, the CDU occupies the pod-to-facility water interface directly.

Fluid-management and leak-detection specialists (white-space adjacency). The 2025 Vertiv PurgeRite acquisition was a substantial move into the liquid-cooling services layer. Adjacent categories, such as leak-detection cable manufacturers, dielectric-coolant chemistry providers, and cold-plate specialist fabricators, occupy similar structural positions.

High-power busway and DC protection (grey-space adjacency). As 800 VDC native architectures move from reference designs to procurement (see the DC-DC Transition series), the busway and DC protection layers become more valuable. Specialist SSCB manufacturers, DC arc-flash sensor vendors, and high-power busway fabricators could attract acquisition interest from the modular incumbents.

Prefabricated shell and campus specialists. Compass Datacenters, Vantage Data Centers, and DataBank each hold significant portfolios of campus-scale factory-built shell capability. The Schneider partnership with Compass demonstrates one model for how an infrastructure incumbent can access shell scale without owning the operator.

Immersion cooling and dielectric fluid vendors. Submer, Green Revolution Cooling, and adjacent immersion specialists operate outside the pod-plus-hall composition architecture, which limits their attractiveness to the current incumbent set. If a subset of AI workloads standardises on immersion, this changes.

20What this means for operator procurement

Two practical implications hold the two-category thesis into procurement:

Read the category correctly, underwrite interface ownership. Grey-space power skids and white-space IT pods are structurally different products. Product-level vendor comparison loses analytical purchase when the campus is composed across both categories. The vendors capturing the standardisation returns are the ones owning interface qualification; ask which vendor owns which interface, which interfaces the pod or skid spec locks in, and where replacement-path risk concentrates.

Sequence the workstreams in parallel with the hardware. Utility interconnection application starts before the modular hardware order. Site preparation runs in parallel with factory assembly. Commissioning skills pipeline belongs at Stage-1 hiring. The modular template compresses the build but does not compress the sequencing around it.

21How modular evolves 2027-2030

AUTHOR SCENARIOThe 2024-2026 incumbent moves establish the grey-space and white-space split. The next three-year window looks likely to compound the same pattern along axes.

Standards convergence in grey space. Open Compute Project reference designs (ORV3, MHPS) and their equivalents at ODCC (China) and Nordic standardisation bodies have been mostly reference-only through 2026. Hyperscaler procurement volume is starting to pull incumbent products toward those reference specs. The likely destination is fewer vendor-proprietary interfaces and more standardised, hyperscaler-audited configurations. This concentrates procurement toward vendors that can meet the reference at scale. The 800 VDC transition (see the DC-DC Transition series) is the deeper shift under this dynamic.

DTC native in white space. Direct-to-chip liquid cooling has been optional through 2024-2026; the DTC-native white-space pod becomes the default configuration for AI training campuses over 2027-2030. Immersion mainstreams in specific workloads (some inference clusters, some HPC/scientific compute); DTC stays dominant across the board. Lifecycle services (PurgeRite-type fluid management) scale with the installed base of liquid-cooled halls.

Consolidation frontier moves to adjacent categories. The 2024-2026 incumbent transactions run from the pre-integrated skid and prefabricated pod (Eaton, Schneider, Vertiv) through into the thermal-interface layer with CoolIT/Ecolab in July 2026. Leak-detection specialists, dielectric-coolant chemistry providers, cold-plate specialist fabricators, SSCB manufacturers, and campus-shell operators are structurally similar targets. The pattern of 2024-2026 repeats along adjacent categories.

Regional buildout shifts. North American hyperscale plus European (Nordic + Ireland) programmes dominated 2024-2026 modular procurement. Over 2027-2030 the volume weight moves. India and Gulf Cooperation Council sovereign AI programmes commission at scale; APAC hyperscaler volume expands; CBAM starts to drive the delivered economics of imported prefab into EMEA. The manufacturing footprint evolves accordingly, with more localised production in India and MENA joining the existing US, Nordic, and Chinese anchors.

Operator model completes its transition. The specification-and-integration model becomes the default; the emerging-alternative framing fades. Colocation operators complete the transition to skilled interface buyers; hyperscalers extend specification power across more of the stack; mid-tier enterprise operators access hyperscale-grade modular products through the same reference designs that hyperscalers use. The skills pipeline for interface qualification and commissioning becomes a first-class capex line for operators that lag on it.

What could bend these trajectories. A sharp hyperscaler capex pause (not the base case) delays every axis by 12-24 months. A geopolitical constraint on cross-border modular movement (India-China, US-China) fragments the vendor field along regional lines. An unexpected cooling breakthrough (single-phase immersion at hyperscale, cost-parity two-phase) restructures the DTC-vs-alternatives dynamic. A materials cost shock (transformer steel, lithium battery pricing) reprices grey-space skids. Each of these is a real risk; none is the base case as of 2026.

Frequently asked

What are the categories of modular data centre?

Grey-space modular is the power skid: MV/LV transformer, LV switchgear, UPS, batteries, controls, factory-assembled into one tested unit that delivers reliable predictable power to racks elsewhere on the campus. White-space modular is the IT-hall pod: rack rows, hot-aisle containment, in-pod cooling (CDU + secondary loop), busway, cabling, and integrated leak detection. Cooling appears in white-space pods because the pod houses the racks. DD OBSERVATIONIn the vendor field examined here, grey-space power skids appear to account for most merchant modular volume.

What is inside a grey-space modular power skid?

A typical Eaton Fibrebond or NordicEPOD power skid contains an MV/LV transformer, LV switchgear and panel boards, a UPS module with an integrated battery cabinet, controls and monitoring, skid HVAC for the electrical room only, and fire suppression. It exposes an MV feed on the utility side and an LV busway on the facility side. The skid delivers reliable predictable power to racks housed elsewhere on the campus.

What is inside a white-space modular pod?

A Schneider EcoStruxure Pod or Vertiv SmartMod Max opens to two rows of IT racks in a hot-aisle containment configuration, an in-pod CDU sized to the pod's IT load with a secondary water loop pre-plumbed and quick-disconnect connectors on the cold-plate side, an overhead busway with per-rack drop-outs, structured cabling, and integrated leak-detection. Cooling appears in this product class because the pod contains the racks.

What were the major modular data centre M&A moves in 2024-2026?

Eaton took a 49% strategic minority stake in NordicEPOD (May 2024, grey-space) and acquired Fibrebond for $1.4 billion (closed April 2025, grey-space). Schneider Electric and Compass Datacenters jointly launched a prefabricated EcoStruxure Pod (September 2025, white-space). Vertiv expanded its SmartMod family and separately acquired PurgeRite for approximately $1 billion cash plus up to $250 million in additional consideration (announced November 4, 2025; completed December 4, 2025). Ecolab closed its acquisition of CoolIT Systems on July 2, 2026 for approximately $4.75 billion, capturing the white-space thermal interface.

Does modular procurement reduce data centre time to revenue?

Modular procurement compresses construction time from serial site work into parallel factory work. Energisation, commissioning, and utility interconnection follow their own timelines and remain independent of factory fabrication cadence. In constrained markets, ground-break to revenue-ready can still run 24-48 months or longer even when modular hardware is fabricated in months. Utility interconnection is frequently the critical path; module fabrication finishes earlier.

Why does cooling appear in white-space pods but not in grey-space skids?

Grey-space power skids contain electrical infrastructure only. The only cooling inside a power skid is HVAC for the electrical room to keep transformers, UPS, and switchgear at operating temperature; the racks themselves are housed in a separate white-space unit. White-space pods contain the racks, so the pod needs an in-pod CDU, secondary water loop, and cold-plate quick-disconnects to serve the rack-level heat load. At 100+ kW per rack, direct-to-chip liquid cooling becomes practical and eventually necessary, which is why the AI density transition drove the growth of the white-space pod category.