What is a modular data center? A guide for AI workloads
A modular data center is built in factory-made blocks and assembled on site. This guide covers the types, sizes, cooling, speed claims and trade-offs for AI.
The short version
- A modular data center is built from factory-made blocks that hold power, cooling and sometimes the IT equipment, and that are assembled on site and added one after another.
- For AI, current modular products range from about 1.77 MW to 10 MW per unit and are designed for liquid-cooled racks such as NVIDIA GB300 NVL72.
- Modular design shortens the construction part of a project, but it does not remove the wait for a grid connection or on-site power.
A modular data center is a data center built from factory-made blocks, called modules, that are tested before shipping and joined together on site. Instead of building one large hall over many months, the builder installs repeatable units and adds more as demand and power allow. The idea is old, but AI has made it relevant again because buyers want capacity faster than traditional construction delivers it.
What counts as modular
There is no single standard, and vendors use the word for several different products. Four forms are common.
Prefabricated power modules hold the electrical gear: switchgear, transformers, batteries. Schneider Electric says it is the largest maker of prefabricated power in the world, citing a May 2026 Omdia report; that is a company claim that we have not checked against the report.
Prefabricated IT pods hold the rack row, with the power and cooling distribution for high-density AI compute inside the white space. Schneider says pods can cut installation time from months to days.
Complete modular data centers combine the building, power, cooling and computing space in one unit. Crusoe’s Spark Liquid and Armada’s Orion are examples. Armada’s Orion has a six-module core that supports up to 2,880 GPUs across 40 racks, with each module containing compute, networking, power and cooling.
Modular buildings are larger steel-frame structures assembled from modules. Eliovp’s ModFlex is described this way. Some vendors package modules in standard shipping containers; others, ModFlex among them, say they do not.
Eliovp’s ModFlex covers the building, power and cooling in one deployment. Its steel-frame modules are designed per project, capacity can be added module by module, and cooling is air, direct liquid or hybrid. Eliovp has not published a deployment time. See Eliovp’s ModFlex page and its post on NVIDIA DSX POD.
Typical sizes for AI
Size per unit has grown fast. DatacenterDynamics reports that Armada’s new Orion is 10 MW per unit, almost ten times its Leviathan model launched last year at 1.77 MW. Crusoe’s Spark Liquid supports up to 3.5 MW per unit and, according to Crusoe, delivers five times the compute density of earlier Spark generations. Production joins Crusoe’s Spark line in Brighton, Colorado, in 2027.
At the small end, BlueFlare’s Alberta project, a modular site paired with on-site gas generation, is 9.9 MW in total, according to DatacenterDynamics. These sizes are far below the gigawatt campuses getting attention. Cleanview counts 58 GW of US data center capacity under construction, so a 3.5 MW module is a small fraction of one site. Modular products reach that scale by repeating units.
Why AI buyers choose modular
Three reasons come up in vendor and trade coverage.
Speed. Building the units in a factory lets site preparation and manufacturing run at the same time. Crusoe’s Bryan Schramm says it can bring capacity online “in as little as a few months from a ready site.” Schneider claims up to 60 percent less deployment time for modular designs. Both are vendor claims, and the phrase “from a ready site” matters: it assumes power and land are in hand.
Predictability. Schramm says factory building makes delivery dates more reliable, which matters when customers reserve capacity ahead of time. Schneider describes operators signing multi-year agreements for prefabricated modules, one example being a contract that might call for ten modules a month for three years.
Staging. Capacity is added in blocks. An operator can start with a few modules and add more as customers sign, instead of building and powering a full hall before the first rack runs. Crusoe’s Schramm also notes that a factory-built unit can use power that is available now, from the grid, on-site generation, small modular reactors or solar with batteries, in smaller blocks close to demand.
Where demand is heading
Schneider Electric cites a JLL forecast that the global data center industry will grow at a 14 percent compound annual rate through 2030 and add 100 GW of power capacity, which JLL values at $1.2 trillion in real estate assets. We have not checked the JLL report. If the forecast is even roughly right, factories that build modules in repeatable batches have an obvious role, and Schneider says it has increased its prefabrication factory footprint by 3.5 times and signed multi-year supply contracts with operators. Those are company statements, but they show where the supply side is heading: buyers are booking module output in advance, which means lead times at the factory are part of the schedule you have to check.
Cooling and rack density
AI racks run hot, and the design of the module depends on the rack. NVIDIA’s DGX GB200 guide puts a full rack at approximately 120 kW, liquid-cooled on the CPUs and GPUs, with networking and storage still cooled by air. Modular designs for these racks are built around liquid: Crusoe’s Spark Liquid targets GB300 NVL72 and Vera Rubin NVL72 racks, and says it follows NVIDIA’s DSX POD, a blueprint for modular infrastructure for the Vera Rubin platform. NVIDIA designs and validates the reference design, and partners build and service it, Crusoe says.
If the rack is lower density, an air-cooled or hybrid module may be enough. This is why cooling choice should be made per project, not assumed by the product category.
What modular does not fix
Power. Crusoe’s Schramm points out that large grid interconnects can take years. A module can be built faster than a utility can deliver power to it. Developers still need a grid interconnection, or on-site generation with its own permits. In Texas this week, regulators issued a major air permit for a planned $24 billion gas-powered data center, and in Virginia a county paused board votes on new applications. Modular units face the same local approvals as any other data center.
Site work. Foundations, utilities, fire approvals and commissioning remain on-site tasks. Factory build helps with schedule only if site work runs in parallel and the integration is planned.
Heat and climate. A unit rated for one climate may be derated in another. FingerMotion’s BlueFlare Snap concept is aimed at extreme heat in the UAE, but it is early-stage and has no published capacity.
Lock-in. A module designed around one rack generation can be hard to reuse for the next. Ask which platforms the design has been engineered for and what changes when the rack power rises.
Questions to ask a vendor
Ask for the rated IT capacity at your site’s design temperature, not the nameplate figure. Ask which cooling method the module uses and which rack models it was designed for. Ask what the schedule covers: factory build, shipping, site integration and commissioning are different durations, and “months” means little unless it names which ones. Ask who supplies the power and what happens if the grid connection slips. And ask whether capacity can be added in stages without interrupting the running modules.
Questions readers ask
What is a modular data center?
It is a data center built from standardized blocks that are made in a factory, tested, shipped and joined on site. A block can hold power equipment, cooling, or the computing racks themselves, depending on the design.
How fast can a modular data center be deployed?
Vendors give ranges, not guarantees. Crusoe says it can bring capacity online in as little as a few months from a ready site, and Armada says its units can be delivered within months. Schneider Electric claims modular designs can cut deployment time by up to 60 percent. All of these are vendor statements, and none includes the time to get grid power.
Are modular data centers the same as shipping-container data centers?
No. Containers are one form. Others are steel-frame modular buildings, prefabricated power modules and pre-built IT pods. Eliovp's ModFlex, for example, uses steel-frame modules designed per project and says they are not standard shipping containers.
Can a modular data center cool high-density AI racks?
Yes, if it is designed for liquid cooling. Crusoe's Spark Liquid unit is built around direct liquid cooling for NVIDIA GB300 NVL72 and Vera Rubin NVL72 racks. Air-cooled modules are limited to lower rack densities.
Does going modular solve the power problem?
Only partly. A modular unit still needs a grid connection or on-site generation. Its advantage is that capacity can be added in smaller blocks as power becomes available.
Sources
- Modular AI Data Centers Gain Ground as Crusoe Debuts Unit, Data Center Knowledge
- Armada launches 10MW modular data center, Orion, DatacenterDynamics
- How modular data centers fast-track AI infrastructure growth, Schneider Electric
- FingerMotion unveils modular data center concept designed to operate in extreme climates, DatacenterDynamics
- NVIDIA DGX GB200 user guide: hardware, NVIDIA
- Data center construction slows in Q3 2026 after record Q2 (Cleanview data), DatacenterDynamics