As the "Dok-Pa-Mo" (independent AI foundation model) project enters its second round, participating companies are betting that domestic AI chip utilization could be the next decisive factor in winning government support. The Ministry of Science and ICT has been emphasizing technology sovereignty across the AI ecosystem and is now putting significant weight behind nurturing Korean-made NPUs (neural processing units).
According to industry sources on Sunday, LG CNS (064400.KS) is partnering with AI chip startup FuriosaAI to develop NPU-based AI services. FuriosaAI's second-generation NPU "RNGD (Renegade)" meets the high-performance requirements for large-scale AI services while significantly reducing power consumption and operating costs compared to GPUs.
The collaboration is particularly notable as it represents a coordinated effort between consortium members that passed the first round of the Dok-Pa-Mo project. LG CNS plans to apply FuriosaAI's RNGD to LG AI Research's K-EXAONE model to optimize performance and maximize synergy. The goal is to strengthen a sovereign AI ecosystem comprising homegrown AI models, services, infrastructure, and AI semiconductors—the core objective of the Dok-Pa-Mo project—and achieve public sector AI transformation using a fully Korean AI stack.
"Through our partnership with FuriosaAI, we will secure NPU-based AI infrastructure technology and expertise," said Kim Tae-hoon, executive vice president of LG CNS. "We will support the advancement of the national flagship AI model in collaboration with LG AI Research and contribute to the development of Korea's AI industry."
SK Telecom's (017670.KS) consortium, which also passed the first evaluation alongside LG AI Research, is similarly focusing on NPU utilization. The consortium used Rebellion's NPU—a member company—in developing its AI model "A.X K1." The NPU was partially adopted to improve efficiency, as Rebellion's chips offer better computational performance per watt in AI inference compared to Nvidia GPUs.
Trillionlabs consortium, which has expressed intent to rejoin the project, has also reached internal consensus to utilize domestic NPUs if officially accepted. Trillionlabs previously developed an AI model using Rebellion's NPU and plans to leverage that experience in its Dok-Pa-Mo challenge.
NPU validation among consortium members is expected to better serve the project's purpose. The government is pursuing the Dok-Pa-Mo project with the goal of creating a Korean AI model that does not depend on foreign technology. Accordingly, the degree of foreign technology dependence is a key factor in determining project participants' success or failure. The government also aims to promote mutually beneficial relationships between large corporations and startups. Companies like FuriosaAI and Rebellion have developed NPUs but faced limited opportunities for real-world validation, constraining their business expansion.
The second round of Dok-Pa-Mo evaluation is scheduled for around June. Three teams selected from the first evaluation and one team chosen through additional recruitment will compete. "Even though the government provides GPUs to Dok-Pa-Mo consortiums, the reason they want to use NPUs is to demonstrate that everything from top to bottom is made with our own technology," a consortium official said, speaking on condition of anonymity.