By Kang Jae-eun
SEOUL, Oct. 18 (Yonhap) -- As artificial intelligence (AI) increasingly moves into the physical world, neural processing units (NPUs) are becoming more important for running AI workloads efficiently on devices.
NPUs are specialized processors designed to handle AI and machine learning workloads. They are among the key technologies enabling AI applications on smartphones, laptops and other devices, such as robots, self-driving cars and smart factories.
South Korean AI chip startup DeepX specializes in designing low-power, high-efficiency chips for edge devices.
Kim Lok-won, chief executive officer (CEO) of DeepX, said the company is expanding partnerships with global technology companies, including Intel Corp., Samsung Electronics Co. and Hyundai Motor Group.
"We are engaging in technical cooperation with Intel, which includes integrating DeepX's NPUs into the OpenVINO ecosystem and optimizing models and software development kits (SDKs)," Kim said during an interview with Yonhap News Agency on Thursday.
The partnership aims to provide a seamless plug-and-play experience when connecting DeepX's chips to Intel's systems.
"An initial outcome is expected to come out sometime early next year," he added.
Kim Lok-won, chief executive officer (CEO) of DeepX, speaks with reporters during an interview at the company's office in Pangyo, south of Seoul, on Oct. 15, 2026. (Yonhap)
Since its founding in 2018, DeepX has developed four types of AI chip, including its flagship DX-M1, which has entered full-scale mass production using Samsung Electronics' 5-nanometer process.
Kim, a former engineer at Apple Inc., said DeepX is seeing growing demand for its chips both at home and abroad as more companies seek to integrate AI into hardware products such as smart glasses, autonomous vehicles and drones.
DeepX received 77 purchase orders from more than 10 countries over the 12 months from August 2025 to July 2026.
The company has signed contracts with 27 semiconductor distributors worldwide, while about 300 entities, including more than 100 South Korean firms, are evaluating its products.
Kim said DeepX plans to open branches in Japan and Europe.
"We are pursuing a first-mover strategy to lead the global market by securing core source technologies in cutting-edge fields," Kim said, emphasizing the company's ambition to "plant the first flag" in the emerging physical AI market.
Kim Lok-won (1st row, L), the chief executive officer (CEO) of DeepX, poses for a photo with AMD's CEO Lisa Su (1st row, C) after a round-table discussion with local artificial intelligence semiconductor companies in central Seoul, in this file photo taken Oct. 7, 2026. (Yonhap)
DeepX is also in the early stages of pursuing business cooperation with U.S. chipmaker AMD following Kim's recent meeting with its CEO, Lisa Su, who visited South Korea earlier this month.
DeepX was among the South Korean AI chip companies invited to a roundtable with Su.
During her visit, Su expressed AMD's willingness to build an open, heterogeneous computing infrastructure with the South Korean government and companies, combining different types of processors, including graphics processing units (GPUs) and NPUs.
"I could see AMD's willingness to collaborate and build an open ecosystem with other companies in the field of physical AI," Kim said. "We also received an email from AMD's technical representative, saying they would conduct follow-up discussions."
Kim said DeepX is discussing the possibility of a partnership with AMD and preparing for further working-level talks.
Kim Lok-won (R), the chief executive officer (CEO) of DeepX, speaks during an interview with Yonhap News Agency at the company's office in Pangyo, south of Seoul, on Oct. 15, 2026. (Yonhap)
Regarding its partnership with Samsung Electronics, Kim said DeepX plans to complete a tape-out of a chip designed using the tech giant's 2-nanometer process this month.
Tape-out is a key stage in semiconductor development when a finalized chip design is sent to a foundry for manufacturing.
"The first set of chips is expected to come out around next March or April," he said.
Kim also said DeepX plans to unveil DX-Newton, software designed to manage robot operations using physical AI, later this year. The software was developed to align its inputs and outputs with Nvidia's Isaac robotics development platform, he said.
"When the number of robots reaches thousands or tens of thousands, it becomes difficult to manage them by having humans install an AI model one by one," Kim said. "It's important to have a system where we can distribute new AI models through a network."
Asked about the timeline for a potential initial public offering (IPO), Kim said a stock market debut was not the company's ultimate goal but that he hoped to take DeepX public within the next three years.
fairydust@yna.co.kr
(END)