Elice Group pushes modular AI data centers with cloud software

Elice Group said it is promoting an AI infrastructure business that pairs modular data centers with cloud software, supplying its Elice PMDC modular AI data center and ECI cloud software, according to ZDNet Korea.

The Elice PMDC can be built in about three months in an environment where the site and power supply are secured, the company said. It modularizes server racks together with power, cooling and networking equipment for on-site installation, and allows phased expansion based on demand.

Elice Group combines the PMDC with ECI, its cloud software that applies GPU virtualization and orchestration technology to allocate and operate GPU resources by workload.

Separately, an edge AI data center based on GPU and NPU is being built at the Myeongji Noksan National Industrial Complex in Busan for shared use by companies in the complex. Elice Group, Busan Technopark, Easy AI, Gun Solution and Fomite are taking part in the project to support AI transformation of manufacturing companies.

Elice Group CTO Park Jung-kook said that as companies expand private AI adoption, control over efficiently operating AI infrastructure itself will become important. He added that modular data centers that can be quickly built where needed and expanded based on demand will be an effective option for companies and institutions handling sensitive data or operating large-scale AI compute.

Background

Interest in on-premises and self-operated AI environments has been building. A paper published by MIT and Georgia Tech researchers found open models performing at about 90% of closed state-of-the-art AI models while costing about one-sixth as much. In Databricks research, fine-tuning Meta's open model Llama 3.1 8B to a company's own code produced a 1.4 times higher code adoption rate and twice the response speed compared with GPT-4o. US insurer Travelers developed its own insurance-focused model, TravelersLLM, and runs its AI environment using external models for general or complex requests.