Containerized data centers: the 2026 buyer’s guide

What is a containerized data center?

A containerized data center is a self-contained computing facility built inside a standard ISO shipping container. Instead of constructing a building and then fitting it out, the manufacturer integrates servers, cooling, power distribution and networking in a factory, tests the system as a whole and ships it to site by road or sea, like any other container.

The format is not new. Modular data centers have served disaster recovery, defense and remote industrial sites for years. What has changed is the payload. A rack of current AI accelerators can draw several times the power of a traditional enterprise rack, and the heat that comes with it is difficult to remove with air alone. Containers designed around that density, often with liquid or immersion cooling, have turned the format from a stopgap into a credible way to deploy AI compute.

In practice, buyers meet three broad families: edge micro data centers for telecom and industrial sites, general-purpose IT modules, and high-density AI containers built around GPU clusters. This guide focuses on the third.

Containerized vs. conventional builds

The two approaches answer different questions. A conventional facility suits large, long-lived campuses where scale on day one justifies the effort. A container suits projects where speed, location or phasing matter more.

Neither model wins everywhere, and many operators combine them: containers to bring capacity online quickly, at the edge or on a new site, and conventional halls for long-term core capacity.

What’s inside an AI container

To make the components concrete, we use the BAYWAW AI Container as a worked example: a 1.2 MW, immersion-cooled AI data center in a 40ft ISO High Cube container, built, integrated and tested in the UAE.

Compute

The container houses 624 NVIDIA B300 GPUs in 78 Hypertec servers, eight GPUs per server. Each GPU delivers up to 13.5 PFLOPS of peak FP4 compute and carries 288 GB of HBM3e memory, for about 179 TB of GPU memory across the container. For large models, memory capacity matters as much as raw compute, so ask for both figures on every quote.

Cooling

The servers are immersed in six GRC NX single-phase immersion tanks, arranged three on each side of a central service aisle, with 13 servers and 104 GPUs per tank. A coolant distribution manifold feeds dielectric liquid across the server blades, which sit on liquid-cooled plates; a return manifold carries the warmed liquid back to the coolant distribution units (CDUs), and roof-mounted coolers reject the heat outside. Coolant loops and CDUs are redundant. The result is a PUE of 1.05. Our page on single-phase immersion cooling explains the principle in more detail.

Power

Total facility power is 1.2 MW, supplied through two power modules: one active at full capacity, one on standby, with N+2 redundancy and failover in under 10 milliseconds. The design is energy-agnostic, so the modules can be fed from the grid or from an optional solar farm with battery energy storage (BESS). We explain that architecture on our solar-powered data center page.

Network

A rail-optimized, non-blocking GPU fabric gives every server eight 400G east-west links, one per NIC, plus two 400G north-south links. Each tank has its own back end of eight leaf and four spine switches. A front-end spine of six Cisco Nexus 9364 switches, redundant edge routers, a pair of next-generation firewalls and an out-of-band management network complete the design. The fabric is validated end to end before shipment, and every network component is listed in the technical specifications.

Site requirements

A container removes most of the construction, not all of the preparation. Site works usually run in parallel with the factory build, which is where much of the time saving comes from.

• Power. A connection sized for the full facility load, 1.2 MW in the case of the AI Container, from the grid, on-site generation or a solar-plus-storage system. Utility lead times vary widely by country, so start this conversation first.
• Foundation or pad. A level, load-bearing surface rated for a loaded 40ft High Cube, whose maximum gross weight is 30,480 kg, with access for the delivery truck and lifting equipment.
• Connectivity. Fiber for north-south traffic to your users, your cloud and your other sites, ideally over diverse routes.
• Clearances. Space around the unit for service access and for airflow around the roof-mounted heat-rejection units.
• Permits. Planning, electrical and environmental approvals depend on the jurisdiction. They are typically simpler than for a new building, but they are never zero, so involve local authorities early.

The 8-point buyer checklist

Use these eight questions to compare offers on substance rather than on brochures.

1. Density and GPU generation. Which accelerators, how many, and how much memory per GPU? A system built around the current generation, such as the NVIDIA B300 with 288 GB of HBM3e, stays relevant longer than one sized for the last.
2. Cooling method and PUE. Air, direct-to-chip or immersion? Ask for the PUE and how it is measured. Industry-average PUE has hovered around 1.5–1.6 for years, according to the Uptime Institute’s annual survey; the BAYWAW AI Container is rated at 1.05.
3. Power redundancy. How many power modules, what redundancy level and what failover time? A single point of failure in power is a single point of failure for the whole cluster.
4. Network fabric. Is the GPU back end non-blocking and rail-optimized? How many 400G links per server, and is the front end sized for your data ingress?
5. Testing before shipment. Is the complete system (compute, cooling, power and network) integrated and validated before it leaves the factory, or only its parts?
6. Delivery time. What is the realistic time from order to operation, and what does it depend on? Get the site-readiness assumptions in writing.
7. Managed services and SLA. Who monitors the system, who replaces a failed component, and how fast is the first response? Compare scope, not only price. Our managed services include 24/7/365 support with an initial response in under 15 minutes.
8. Scalability. Can you add identical units as demand grows, and will they share the same management, network and power architecture?

Thinking about total cost

Price per GPU is the easiest number to compare, and the least complete. Three other factors often weigh as much over the life of the system.

Time-to-compute

Every month a cluster is not running is a month of delayed training, inference capacity or research. Conventional data center projects are typically measured in years from planning to commissioning; a pre-integrated container can be operational in as little as four months.

Energy overhead

PUE compares the total energy a facility draws with the energy that reaches its IT equipment; everything above 1.0 is overhead. A 1.2 MW facility running continuously consumes about 10.5 GWh a year (1.2 MW × 8,760 hours). At that scale, the gap between an average PUE and an immersion-cooled design is a meaningful share of the energy bill, and the source of that energy, grid tariff or on-site solar, matters as much as the quantity.

Phasing capital

Because capacity arrives in self-contained units, investment can follow demand: deploy one container, prove the workload, then add the next. Capital stays aligned with utilization instead of being committed upfront to a building that fills over years.

Conclusion

A containerized data center is no longer a compromise. For AI, it is often the fastest way to put current-generation GPUs to work on your own site, under your own control. The right choice is the one that fits your workload, your site and your operating model, and that has been tested end to end before it ships.