Modular construction is transforming the fundamental principles of the data center industry. According to a new in-depth report from SemiAnalysis, modular methods have become the default approach for hyperscale cloud providers and AI labs to rapidly build data centers. This approach can compress construction timelines by approximately 36%, which equates to a reduction of seven to nine months, and lower overall capital expenditure by about 8% per megawatt.
AWS is aggressively advancing its modular design, internally code-named "SAMDC," through its project "Houdini." This project breaks down the construction of white space into standardized racks that can be prefabricated in a factory. It reduces the preparation time before server installation from up to 15 weeks to just two to three weeks and cuts over 50,000 man-hours of on-site electrical installation work per module. Meanwhile, Meta is deploying aluminum frame fabric structure "tent" buildings at its Prometheus campus in New Albany, Ohio. Satellite imagery shows that since construction was announced in July 2025, eight structures were completed by April 2026. In contrast, the campus's five permanent buildings previously took two to three years to finish.
SemiAnalysis's modularity tracker now covers over 61 GW of modular capacity and more than 1,000 sites using some form of prefabrication strategy. The firm predicts that by the end of 2028, modular construction will account for over 30% of all online capacity. For cloud service providers, speed directly translates to revenue. Conservatively, delivering each megawatt of IT computing power one month earlier can generate about $500,000 in value for the owner-operator. For a 50 MW facility, this represents an undiscounted benefit of roughly $200 million compared to purely on-site construction.
Labor shortages are a structural bottleneck, and modularization is the solution.
The primary driver of this shift is a shortage of skilled trade labor. Electrical workers account for 30% to 40% of all construction labor hours in data center projects. SemiAnalysis's labor model indicates that, driven by the surge in large-scale mission-critical facility construction, a shortage of electrical workers will become apparent by 2027. This shortage will be especially acute in regions with high construction concentration, such as Texas and Ohio.
This constraint has already been observed in practice. Previously, Crusoe had to raise wages by 30% to attract talent to its Abilene campus, a project that required over 9,000 workers at its peak. Under a traditional on-site construction model, a 50 MW AI data center needs about 12,000 on-site man-hours per megawatt during the mechanical and electrical installation phase, with a single building peaking at around 300 skilled workers. The value of modularization lies in transferring repetitive tasks to be completed in parallel at a factory, while on-site construction proceeds simultaneously. According to SemiAnalysis's calculations, by moving the mechanical and electrical installation scope into the factory, on-site man-hours will decrease by approximately 63% to 4,500 hours per megawatt. The demand for certified electrical workers can be reduced by as much as 85%. Factory methods also avoid constraints from weather and geographical conditions, leading to more stable quality control.
The modular system ranges from prefabricated components to complete facility delivery.
SemiAnalysis categorizes modular products into five levels, from low to high: single components, skids, modules, containers, and prefabricated data center blocks. The first four levels represent subsystem modularization, while the last approaches full facility modularization. At the subsystem level, modular power cabins and modular cooling systems are currently the most concentrated areas in the market.
For example, Flex's subsidiary, Anord Mardix, offers a modular power solution that integrates transformers, switchgear, UPS, and battery systems into a single enclosed unit. This approach can compress the mechanical and electrical installation cycle from about 5.5 months to roughly 2.5 months, saving about 5% in cost per megawatt. On the cooling side, Airedale by Modine ships its rack-mounted CDU as a 2 MW prefabricated unit. On arrival, it only requires connecting two water pipes and one power source. At the full facility modularization level, Vertiv's MegaMod integrates IT racks, power, and cooling into a single module. Its supporting OneCore platform provides standardized 12.5 MW power and cooling cabins that can be combined for larger-scale AI factory deployments. Nvidia also officially launched its DSX reference architecture in March 2026. This standardizes computing, networking, storage, power supply, cooling, and even civil structural design into a single blueprint. CoreWeave has already adopted DSX Air to build a digital twin model of its AI factory.
Who leads modularization? Three integration paths have distinct focuses.
Who controls the modularization process directly determines the value capture potential for each party. Three distinct models have emerged in the market. In the operator-led model, the operator designs the specifications and directly procures equipment, leaving integration work to contractors. This model requires strong internal engineering and procurement capabilities and requires the operator to bear equipment lead time and inventory risks. Therefore, it is practically limited to the largest hyperscale operators like AWS and Meta.
In the EPC or system integrator-led model, the operator sets performance requirements, and an engineering, procurement, and construction (EPC) firm or specialized integrator is responsible for procurement, coordination, and construction. Companies like Comfort Systems, Sterling Infrastructure, and Quanta Services' Cupertino Electric fall into this category. Notably, Comfort Systems operates over 3.5 million square feet of factory capacity in Texas and North Carolina through its Environmental Air Systems and TAS Energy subsidiaries. This model is particularly attractive to hyperscale operators because they can retain their own designs, only moving the construction execution into the factory.
In the OEM-led model, exemplified by Vertiv's OneCore, equipment manufacturers integrate their own power, thermal management, cooling, and IT infrastructure into a complete platform for sale. This model has increased Vertiv's content value per megawatt from a historical ~$3.5 million to ~$7 million. However, the trade-off is lead time; delivery cycles for Vertiv's modular solutions currently exceed 12 months.
Different strategies among major operators.
Various operators are not converging on a single modular path. Instead, they are making distinct choices based on their scale, market positioning, and technology stacks. Compass is the colocation operator with the longest history of modularization. An estimated 70% to 85% of the content in each of its buildings is prefabricated in a factory. The framework and roof for a single building can be erected in 18 to 21 days, covering the entire stack of shell, white space, power modules, and medium-voltage switchgear. QTS, on the other hand, builds rapid delivery capabilities by locking in designs early and maintaining about 7 million square feet of warehousing capacity in Kansas to stock long-lead-time equipment. Its latest "Rapids" design has been adopted by two major AI companies.
Aligned Data Centers focuses on adaptability. Its base power and cooling architecture remains standardized, but the data hall configuration can freely switch between air cooling, hybrid cooling, and liquid cooling as rack densities evolve. Its Delta³ air cooling system supports about 50 kW per rack, while its DeltaFlow liquid cooling platform can support over 350 kW. At its Abilene Stargate campus, Crusoe applies both shell modularization and full facility modularization. Prefabricated insulated metal panels reduce the time to seal a single building to under eight weeks. Its self-developed Spark unit, at about 1 MW per unit, ships in a near-complete state and has already been validated at scale through a project in Nevada with Redwood Materials.
Cost and timeline estimates: Supplier promises need adjustment.
SemiAnalysis's underlying calculations reveal a critical context: the speed figures publicly touted by suppliers are often based on a narrower scope, not an end-to-end timeline. Vertiv's SmartRun claims an 85% acceleration, but this applies only to overhead busway and cable management. Vertiv's MegaMod's 50% figure measures the comparison between module deployment and on-site construction. Schneider Electric's 60% figure is applicable only to power and cooling modules.
Based on a full-cycle analysis of a 50 MW liquid-cooled AI data center, SemiAnalysis estimates that the construction window for purely on-site builds is about 18 to 24 months. A fully modular build can compress this to 12 to 18 months. This is about 36% faster than a purely on-site build and about 30% faster than a current baseline that already includes some modular elements. If integrated prefabricated data center blocks or containerized data centers are used, the construction window could even be compressed to under 12 months. In terms of cost, the total cost for a fully modular solution is about $13.5 million per megawatt, lower than the ~$14.6 million per megawatt for a purely on-site build, a difference of about 8%. The savings come mainly from two areas: moving mechanical and electrical installation to the factory saves about $600,000 per megawatt in construction service fees and about $500,000 in installation costs. Shortened construction timelines also lead to savings from cost escalation, contingency reserves, and site management overhead.
It is important to note that modularization also has inherent costs. Compared to a purely on-site build, modular construction adds a layer of module supplier profit, which is the primary source of cost friction. Furthermore, some operators and MEP contractors have reported reliability issues with certain modular solutions. If quality defects emerge, not only are the upfront time savings lost, but valuable hardware assets are also put at risk.
Comments