Kim Jae-won, CEO of Ellis Group, introduces the AI full-stack strategy during a press briefing at the company's headquarters in Gangnam-gu, Seoul, on the 15th./Courtesy of Ellis Group
"Because a modular data center can be built in three months, it can respond in a tailored way to the rapidly increasing demand for artificial intelligence (AI) infrastructure."
Kim Jae-won, Ellis Group CEO

Ellis Group is expanding its portable modular data center (PMDC) business, which can shorten data center construction times that once took more than two years to three to four months. This year, the plan is to accelerate its push into the cloud infrastructure market with a PMDC that supports Nvidia's next-generation graphics processing unit (GPU) "Vera Rubin."

Ellis Group, which started as an Edtech startup and transitioned into an AI infrastructure company, announced its business strategy for this year at a press briefing held at its headquarters in Samseong-dong, Seoul, on the 15th. Kim Jae-won, Ellis Group CEO, said, "This year, the PMDC-based cloud business is expected to account for more than half of annual revenue," adding, "Based on our self-developed PMCD technology, we will step up efforts to localize the AI infrastructure ecosystem and also push for an initial public offering (IPO) within the year."

Founded in 2015, Ellis Group was originally an Edtech corporations operating an AI education service, but five years ago it jumped directly into building AI data center infrastructure while developing the cloud software needed for its hands-on training platform "Ellis LXP."

Ellis Group's self-developed PMDC is a modular data center that packs core facilities such as servers, power, and cooling into a container. The PMDC's strength is speed. It can reduce data center build times that used to take more than two years to three months, allowing flexible response to demand from corporations that need an AI data center right away.

Kim said, "According to BCG data, corporations that adopted AI proactively had double the sales growth rate compared with those that did not, and expense was found to be reduced by more than 40%," adding, "From a corporate standpoint, safely scalable AI infrastructure is urgently needed."

The company has built more than 10 PMDC units that support Nvidia's flagship GPU lineup (A100, H100, B200). Most recently, it completed development of a PMDC that supports Nvidia's next-generation "Vera Rubin NVL72," which requires high power of 230 kW per rack. It is also developing a cooling technology that uses hot water, not chilled water, to efficiently dissipate heat generated when running high-power GPU equipment.

Based on Blackwell Ultra (B300), it has also secured design capabilities to cluster up to 13,068 GPUs. Clustering is a technology that connects multiple GPUs over a network to improve large-scale computation speed and efficiency. The industry has mainly used Nvidia's expensive InfiniBand, a data-center-only network, but Ellis Group plans to cut expense by implementing Ethernet-based clustering technology, a general-purpose network standard.

Kim said, "AI infrastructure competitiveness does not depend on the number of GPUs, but on how well they are utilized," adding, "No matter how good the GPU is, if you do not have software such as a tailored storage system, speed and performance will degrade."

Ellis Group is also accelerating the creation of a "sovereign AI" infrastructure ecosystem that protects security and Data Sovereignty in cooperation with domestic corporations such as AhnLab, LG Uplus, and MakinaRocks. It also took part in the government's call for proposals for a "GPU acquisition, build-out, and operation support" project worth 2 trillion won. It launched a "GPU spot plan" that uses idle GPU resources to offer services at up to 50% of on-demand rates so that domestic startups and research institutes can use Ellis Group's services at low expense.

Kim said, "Through next-generation AI infrastructure completed with homegrown technology, we will raise the self-sufficiency of the domestic AI ecosystem and lay the groundwork so that corporations can respond to surging AI demand without expense barriers."

It is also preparing to list on KOSDAQ within the year. Funds raised through the IPO will be used to expand the GPU-based PMDC business. Ellis Group posted revenue of 39.5 billion won and operating profit of 5.2 billion won last year, and unlike many startups that run losses, it touts as a strength its stable annual earnings based on steady performance of its existing education platform.

※ This article has been translated by AI. Share your feedback here.