📝 Originally published (in Japanese) at forge.workstyle.tech.
In this article, we'll explain what a data center is and how containerization changes the game.
We'll also explore the relationship between GPUs and liquid cooling, as well as the various usage options.
By the end, you should be able to explain how containerized data centers work to others.
No prior knowledge of GPUs or data centers is required.
What Exactly is a Data Center?
A server is a computer that stores and processes data. It's used to run websites, train AI models, and perform AI inference. Inference is the process of feeding data into a trained AI model to get answers or results.
Servers consume electricity and generate heat when operating. To safely run many servers, you need power supplies, cooling systems, and racks to house the equipment. A data center is a facility that houses all these components and provides the environment to operate servers.
What Exactly Is a Containerized Data Center?
A containerized data center refers to a facility where all the necessary data center equipment is housed inside an intermodal shipping container typically used for maritime or rail transport. It consolidates power distribution units, cooling infrastructure, server racks, servers, and an anteroom into a single container, allowing it to function as a fully operational data center.
Instead of constructing a dedicated building from scratch, think of it as packaging all the essential infrastructure into a container and shipping it directly to where it's needed. Because it is built inside a standard shipping container, one of its standout features is how easily it can be relocated to a different site after deployment. Keep in mind, however, that you will still need to verify the specific site requirements and prerequisites at the new location.
How is a Containerized Data Center Different from a Building-Type Data Center?
In a building-type data center, a dedicated building is constructed to house servers, and the necessary infrastructure is set up within it. In contrast, a containerized data center eliminates the need for a building by modifying shipping containers to accommodate the required equipment, which is then installed and operated. This approach offers the advantage of faster deployment.
Another key difference is the ability to design power and cooling systems according to the specific scale of usage. Containerized data centers can be optimized to allocate most of the power directly to servers. In building-type data centers, additional costs are incurred for non-server infrastructure and air conditioning.
Commonly cited benefits include faster return on investment, lower operational costs, and quicker installation. However, actual results may vary depending on the operator and specific conditions. It’s not guaranteed that a containerized data center will always reduce costs.
Why Are Containerized Data Centers Getting So Much Attention Right Now?
To put it simply, it comes down to time.
Traditional brick-and-mortar data centers take a long time to get up and running. First, you have to scout land, secure power access, obtain permits, and then actually construct the building. For massive AI hubs, reaching a 1 GW (gigawatt) scale has historically taken anywhere from 1 to 3.6 years from groundbreaking (according to data compiled by Epoch AI in November 2025). Given the sheer pace of AI growth today, spending two or three years just building a place to house GPUs means falling behind.
On top of that, the hardware going into these facilities has fundamentally changed. GPU servers generate vastly more heat than traditional servers. What's more, they need to be brought online at an unprecedented density and on aggressive timelines. If you can't find an existing empty building right away, containerized solutions—where the entire infrastructure is shipped pre-packaged in modular enclosures—become a very practical choice. That is how I see it.
Containers aren't the only approach for rapid deployment, either. At its Ohio facility, Meta houses GPUs inside weather-resistant tents. These fabric-covered, weatherproof structures are what Meta calls "rapid deployment structures." Mark Zuckerberg explained that instead of spending four years constructing concrete buildings, they developed an approach to speed up deployment by setting up weather-resistant tents (as reported in July 2025).
Don't wait for the building—get compute capacity up and running first. The growing focus on containerized data centers is part of this exact trend.
Why Does It Work So Well with GPUs?
GPUs are computational devices originally developed for image processing. Today, they are also widely used for AI training and inference, which require performing massive amounts of calculations in parallel. When GPU devices are deployed at high density, they generate significant heat, often exceeding the cooling capacity of standard air-cooling systems.
This is where liquid cooling comes in. Liquid cooling is a thermal management method that uses water, instead of air, to carry heat away from equipment. With this approach, the cooling infrastructure inside a container can be designed to precisely match the thermal output of GPU servers.
Let’s trace the water flow step by step. Water, chilled using well water or another source, passes through a cooling tower and a CDU (Coolant Distribution Unit). The CDU is a device that distributes cooling water to where it is needed. From there, the water travels through pipes to the cold plates. Cold plates are cooling plates mounted directly against the GPUs to dissipate heat. The water absorbs the heat from the GPUs, warms up, and returns to the loop.
What kind of processes are they used for?
Processes that utilize GPUs include AI training and inference. For example, this encompasses the training and inference of Large Language Models (LLMs—large AI models that process text), image recognition and generation, AI that handles multiple types of data, and Agent AI. Agent AI refers to AI that proceeds with tasks based on a specific objective.
Outside of AI, GPUs are also used for simulations of fluid dynamics, genomes, materials, and electromagnetic fields. Simulation is the process of setting conditions on a computer to calculate phenomena and changes. The type and amount of calculations required vary depending on the application.
What are the possible use cases?
There are several ways to utilize a containerized GPU data center. Your choice depends on how much infrastructure you want to manage yourself.
- Renting GPUs is referred to as "hosting." Instead of preparing your own GPU servers, you use them as a service.
- Placing your own GPU servers in a facility is known as "colocation." You provide the servers yourself, but use the physical space and facilities.
- There is also the option to have containers installed on your own property. In this case, your company must provide the land, power, network, and water source.
- Another option is to build your own private data center within the premises of a service provider's campus.
The right model for you depends on whether you want to rent specific GPUs, use your own hardware, or whether you can provide the physical location for installation.
How are global leaders using container-type data centers?
We also investigated how major overseas IT and AI companies are using container-type data centers (based on publicly available information as of October 2026). To summarize the findings upfront, massive bases for training AI are still being built in large buildings. True container-type data centers are being used in situations where portability and rapid construction are valuable.
The origin of this approach dates back surprisingly far, with Google using containers in its first self-built data center, which was launched in the fall of 2005. They lined up 45 containers, each holding up to 1,160 servers, and consuming 250kW of power per container. The contents of this were made public in 2009.
Cloud service providers are using them as data centers that can be taken to the field. Microsoft's "Azure Modular Datacenter" (announced in 2020) is a data center packed into a 40-foot container. It can operate even in areas without internet connectivity and can be connected to satellite lines, making it suitable for defense and national security applications. AWS's "AWS Modular Data Center" (2023) is also housed in a sturdy shipping container and can be transported by ship, rail, truck, or military transport aircraft, targeting US government customers.
On the other hand, regarding the large-scale AI bases of OpenAI, xAI, and Anthropic, no publicly available information was found on using containers to house servers. xAI's "Colossus" uses a building that was formerly a factory in Memphis. Before being connected to the power grid, it relied on temporary gas turbines for electricity, and also uses large batteries. Even if containers housing generators or batteries are placed there, they are not the data center itself, but rather power equipment. When reading news, distinguishing between these can help avoid confusion.
There are also new developments. "Modular-type" data centers, which are assembled in factories and transported for AI use, are emerging. Crusoe announced in March 2026 that it would build a factory for its modular AI data center, "Crusoe Spark". It can be scaled from several hundred kW to several hundred MW and can be delivered in as little as three months. Currently, it uses an air-cooling system, with a liquid-cooling version planned for release in the second half of 2026.
In summary, the reasons why container-type or modular-type data centers are chosen are that they can be built quickly, are portable, and can start small to accommodate power supply conditions. They are being used in various situations, distinct from massive learning bases.
When is it suitable, and what should we be cautious of?
Since it is designed as a transport container, its ease of relocation is a feature that also aids in disaster preparedness. BCP (Business Continuity Planning) refers to plans and measures to ensure business continuity in the event of disruptions caused by disasters or other incidents. Container-based solutions are considered when you want the option to move equipment to necessary locations.
This form is also suitable for edge computing, which processes data closer to where it is generated or used. The feasibility depends on the availability of suitable locations for container placement, as well as the necessary power supply, communication infrastructure, and water sources.
When considering implementation, consult with providers about the setup period, the number of units that can be installed, required power, cooling methods, and efficiency. These factors vary significantly depending on the provider and installation conditions, so they cannot be judged solely on common figures. If your use case requires land or power supply, ensure that these can be prepared.
Summary: What should you remember?
A containerized data center consists of distribution boards, cooling equipment, racks, servers, and airlocks all housed within shipping containers. Their key features are the ability to be installed without constructing a permanent building and the ease of relocation since they are standard shipping containers.
GPUs generate significant heat, and air cooling alone may not be enough to unlock their full performance. Liquid cooling uses water to transport that heat, which is then captured by cold plates mounted directly onto the GPUs. Your deployment options include renting GPUs, dropping off your own GPU servers, installing them at your own location, or placing dedicated equipment on an operator's premises.
When evaluating these options, first organize the types of workloads you want to run and how you prefer to manage the hardware. Then, confirm the details regarding installation conditions, deployment timelines, quantity, power consumption, and cooling capabilities with the provider.
References
- Build times for gigawatt-scale data centers (Epoch AI, November 2025)
- Meta packs GPU clusters into tents to speed up construction of AI data centers (GIGAZINE, July 25, 2025. Originally reported by Fast Company and The Information)
- Azure Modular Datacenter: mission resiliency for the field (Microsoft, December 7, 2020)
- AWS rolls out modular datacenters for JWCC (The Register, February 15, 2023)
- Google Unveils Its Container Data Center (Data Center Knowledge, April 1, 2009)
- xAI's Colossus turbines (The Register, May 8, 2025)
- Crusoe Announces New Manufacturing Facility to Produce Modular AI Factories (Crusoe, March 12, 2026)
Top comments (0)