Modular Data Centers for AI: Manufactured Units and Deployment Speed
A single NVIDIA GB200 NVL72 rack draws 125 to 135 kW in operation, while Xometry puts a typical modular data center module at 50 to 400 kW. At that draw, the advertised 5 to 25 server racks per module shrink to three GB200-class racks at most, and to none in a module of 100 kW or less.
Modular data centers are best suited for AI pilots and smaller GPU clusters under roughly 2 MW, expanding capacity at existing sites, and hybrid-cloud AI deployments. Large hyperscale campuses, highly customized facilities and sites with strict seismic requirements may be better suited to traditional construction.
What Is a Modular Data Center?
A modular data center is a pre-engineered facility assembled in a factory and delivered as standardized modules. Each module arrives with its IT, power, cooling, networking and controls already integrated, reducing on-site installation and commissioning work. For AI deployments, the module's power rating matters most, because AI rack density can exceed a small module's entire rating.
How Many AI Racks Fit in One Module?
Using a GB200-class rack's 130 kW operating draw, we can estimate how many racks fit within Xometry's typical module capacity range. The table below compares module power capacity against rack requirements.
| Module capacity | GB200-class racks at 130 kW operating draw |
|---|---|
| 50 kW | 0 (below one rack's operating draw) |
| 100 kW | 0 (below the 125 kW low end of one rack's operating range) |
| 200 kW | 1 |
| 400 kW | 3 |
A 400 kW module, at the top of the typical range, can support up to three GB200-class racks based on power capacity alone. A 50 kW module cannot power even one such rack, but can support lower-density AI or conventional IT workloads. A GB200-class deployment requiring more than three racks would therefore need multiple modules, making scalability an important consideration alongside deployment speed.
How Much Faster Does a Modular Data Center Deploy?
A single modular data center unit can be delivered in 6–12 weeks, according to Xometry, although full deployment takes longer. Traditional brick-and-mortar facilities typically require 18–30 months of construction, plus 6–18 months for permitting, based on Zettabyte's construction timeline.
Actual deployment timelines depend on site readiness, grid connections and manufacturing capacity. In the United States, permitting delays and community opposition over electricity use, water consumption and noise can further extend schedules.
What Is a Containerized Modular Data Center?
Containerized modular data centers use standard 20-foot or 40-foot shipping-container formats, making them easier to transport to remote or space-constrained sites. Their ability to support high-density AI workloads depends on usable IT power, cooling capacity and rack space, not container size alone.
Where Does Modular Fit an AI Deployment?
Practitioners from Vertiv, CBRE and HPE identify five deployment scenarios in Data Center Knowledge's Winter 2023 report on when modular data centers work.
| Deployment scenario | Relevance to AI/GPU infrastructure |
|---|---|
| Small, gradual deployments (under 2 MW) | Start with an AI pilot or small GPU cluster and add modules as demand grows. |
| Expanding existing facilities | Add GPU capacity to sites that lack sufficient rack-level power or cooling for high-density AI workloads. |
| Large deployments using standardized modules | Scale through repeated modules, provided total power, cooling and site infrastructure can support the deployment. |
| Edge data centers | Deploy smaller AI clusters close to users or data sources where low latency matters. |
| Hybrid cloud and AI/HPC | Add dedicated AI compute capacity alongside existing cloud or on-premises infrastructure. |
Three scenarios are particularly relevant to enterprise AI: smaller GPU clusters, expansion of existing facilities and hybrid-cloud deployments. For existing sites, modular infrastructure can provide dedicated high-density power and cooling without rebuilding the original facility, provided sufficient upstream electrical capacity is available.
Where Is Modular the Wrong Choice?
The same practitioners identify three situations where modular data centers may be less suitable.
- Heavy customization. Modular construction relies on standardized designs to reduce cost and deployment time. Highly customized power, cooling or layout requirements can reduce those advantages.
- Very large and hyperscale facilities. At approximately 50 MW and above, traditional construction may offer greater flexibility and economies of scale, although there is no fixed threshold. Hyperscale facilities typically operate at 40–100 MW for a single tenant, with multi-building campuses reaching several hundred MW.
- Sites with strict seismic requirements. Standardized modules may require additional structural engineering or customization to meet local seismic codes.
The rack calculations show why modular deployment becomes more complex at hyperscale. A 50 MW AI deployment at 130 kW per rack requires roughly 385 GB200-class racks. At three racks per 400 kW module, that means at least 129 modules. At this scale, manufacturing capacity, site power and cooling infrastructure become major deployment constraints, making traditional brick-and-mortar construction an important alternative to evaluate.
Is a Modular Data Center Cheaper Than a Traditional Build?
Modular data centers can reduce construction costs through standardized components, factory assembly and lower on-site labor requirements, while traditional brick-and-mortar facilities may incur higher costs from custom designs. Salas O'Brien estimates $13.5 million/MW for modular versus $14.6 million/MW for traditional construction, roughly 8% savings. Zettabyte's recent deployment experience in Asia indicates approximately 9% lower construction costs for modular facilities. Actual savings vary by project, particularly location, power and cooling requirements for high-density AI workloads.
Does Modular Fit a Planned AI Deployment?
Choosing between modular and traditional brick-and-mortar data centers comes down to three factors: capacity, site requirements and deployment economics.
- Calculate the required capacity. Divide each module's usable IT power by the planned rack's operating draw to estimate how many racks it can support. Then check whether enough modules can be delivered and commissioned within the project timeline.
- Evaluate scale and site requirements. Modular works well for smaller AI clusters, expanding existing facilities and hybrid-cloud deployments. Highly customized facilities, large hyperscale campuses and sites with strict seismic requirements may favor traditional construction.
- Compare deployment time and cost. Factory-built modules can shorten deployment through parallel manufacturing and site preparation, but single-module delivery estimates do not represent full facility commissioning. Standardization may also reduce construction costs, although comparable cost-per-kW data is needed to quantify the savings.
FAQ
Can Modular Data Centers Support Liquid-Cooled GPUs?
Yes. Modular data centers can support liquid-cooled GPUs when equipped with the necessary cooling infrastructure. Vertiv's MegaMod HDX integrates direct-to-chip liquid cooling for high-density AI workloads. Compatibility depends on cooling distribution units (CDUs), heat-rejection capacity and electrical power, not just the module's kW rating.