Modular AI data centers bring up to 2,000 GPUs in any packaging

A traditional data center protects valuable assets inside with a “shell” made of steel and concrete. Building a data center shell is not cheap compared to the cost of the equipment and infrastructure inside it, but it is not trivial. It takes time for engineers to consider potential sites, apply for permits, and coordinate with construction contractors.

It’s a problem for those who want to install AI tools quickly, which has led companies like Duos Edge AI and LG CNS to respond in a more traditional way. They use prefabricated, self-contained boxes that can be shipped in months instead of years. The boxes can work alone or in combination with each other, giving the option to add more if needed.

“Just got back from Nvidia’s GTC, and lots of it [companies] they’re sitting in their jobs because their data centers haven’t been fixed, or they can’t find the space,” said Doug Recker, CEO of Duos Edge AI. “We see a need there, and we can deliver quickly.”

GPUs shipped directly to you

Duos Edge AI’s modular compute pods are 55 meters long and 12.5 meters wide. Although they look similar to a trailer, they are actually a bit larger and are designed to be transported by trucks. Each computer pod has racks of GPUs like those used in other data centers. Duos recently entered into an agreement with AI architecture company Hydra Host to install four pods with 576 GPUs per pod. That’s a total of 2,304 GPUs, with the option to later double it to 4,608 GPUs.

Conventional data centers are not new for Duos; the company previously used nearby data centers for rural customers, such as the school district in Amarillo, Texas. However, Hydra Host deployment pods will be optimized to handle heavy AI loads. They will have more racks, draw more power, and use liquid cooling to keep the GPUs running efficiently.

Across the Pacific, Korean tech giant LG is taking a similar approach. Support company CNS, which provides services and IT services, has announced the AI ​​Modular Data Center which, like the Duos unit, contains racks of GPUs and supporting equipment in a pre-installed building.

Also like the Duos deployment, LG’s AI Modular Data Center has 576 Nvidia GPUs with the option to expand in the future. “We are currently developing an expanded version that can support more than 4,600 GPUs in a single unit, with service launch planned for later this year,” said Heon Hyeock Cho, vice president and head of datacenter business division LG CNS. LG’s first Modular Data center will open in the South Korean port city of Busan, where it can house up to 50 units.

LG and Duos are not the only ones. Hewlett Packard Enterprise, Vertiv, and Schneider Electric now have modular data centers available or in development. A report from market research firm Grand View Research estimates that the market for modular data centers could more than double by 2030.

Online, but under the radar

A modular data center is very different from a traditional data center because it is not necessary to build a large steel and concrete shell. Instead, the site can be prepared by pouring a concrete slab. Prefabricated modules are trucked in, placed on the pad where desired, and then connected to the site.

Duos deployments, for example, include power modules placed next to desktop computers, and the pods are connected to passive connections that allow the pods to work individually. Recker compared it to lining up school buses in a parking lot. “Everything is built off-site in a factory, and we can put it together like a jigsaw puzzle,” he said.

That simple fact is the point. Both Duos and LG CNS expect a modular data center to be deployed in about six months, compared to the two or three years that a standard data center would take. Recker said, for Duos, the turnaround is so fast that building a pre-made unit isn’t always a problem. While it’s possible to build a prefabricated unit in 60 or 90 days, site preparation adds time “because you can’t get permits as quickly.”

Modular data centers can also provide good value. Recker said that a five-megawatt facility could be built for $25 million, and that the Duos cost per megawatt about half of the larger rates. For Duos, the savings are in part because its modular data centers can target smaller services where licensing is less difficult. Smaller, modular deployments are also meeting less resistance from local governments, which are increasingly skeptical of data center construction.

While the Duos is aimed at small applications, LG hopes it will be big. Its planned Busan campus of 50 AI Modular Data Centers suggests an ambition to achieve deployments that rival the capabilities of conventional services. An area with 50 units would bring the total number of GPUs to over 28,000. Here, the benefits of a modular approach can largely arise from scalability, as a modular data center can start small and grow as needed.

“Using a common method, the AI ​​Modular Data Center can be expanded incrementally by connecting multiple AI Boxes,” Cho said. “It helps build even state-of-the-art AI data centers.”