Defense Agrees on Need for AI Chips, but Divided on ‘Korean CUDA’
The defense industry needs a platform and development ecosystem to run AI and high-performance computing software. (Photo source: riazhaq.com)
The defense industry, which produces weapon systems, argued that a common software platform like CUDA is necessary to use domestically produced AI chips. However, Baek Joon-ho, CEO of NPU industry frontrunner FuriosaAI, countered that the platform should be led by chip companies within an ecosystem, not designed and imposed by the state as a standard. CUDA is a parallel computing platform and development ecosystem created to run AI and high-performance computing software exclusively for NVIDIA GPUs. It is considered a decisive factor that makes it difficult for NVIDIA GPU users to switch to other GPUs.
While a consensus has formed on the need to foster domestic NPUs as an alternative to NVIDIA GPUs, a gap has emerged regarding the feasibility of a supporting ‘Korean CUDA’.
At the ‘26-3rd Defense AI Innovation Network Seminar’ held on March 18 at MODU Labs’ Gangnam Campus in Seoul, participants including Lee Hyeong-jin, a Deputy Director at the Defense Acquisition Program Administration (DAPA); Baek Joon-ho, CEO of FuriosaAI; Shin Sung-kyu, Vice President of Rebellions; Shin Dong-ju, CEO of Mobilint; Lee Jin-won, CTO of Hyperaccel; Kim Jong-hoon, a leader at Naver Cloud; Seo Young-woo, a Senior Executive at Hanwha Aerospace; and Lee Seung-young, CTO of LIG Nex1, shared their opinions on the adoption of domestic AI semiconductors in the defense sector.
There was no disagreement on the necessity of domestic AI semiconductors. DAPA Deputy Director Lee Hyeong-jin explained that if AI systems are developed on foreign semiconductors and platforms, they may be operable domestically but could face various control and certification issues at the export stage, necessitating redevelopment.
DAPA stated that it began to fully recognize the importance of semiconductors in 2024, secured a budget for ‘Defense Semiconductors’ for the first time in 2025 to start research and development, and is also pursuing the enactment of related laws.
The defense industry also agreed on reducing dependence on foreign semiconductors. Seo Young-woo, a Senior Executive at Hanwha Aerospace, emphasized, “We want to use domestic, not foreign, AI software and hardware to carry out defense missions,” and called for national-level support to foster a domestic NPU and a domestic-based AI model ecosystem. Lee Seung-young, CTO of LIG Nex1, the developer of the Cheongung interceptor missile, also stated that domestic defense companies are willing to enhance intelligence by introducing domestic NPUs into their weapon systems.
The point of disagreement was the next step after the semiconductor hardware. CTO Lee argued that domesticating the chip alone is not the end of the story. Lee Seung-young emphasized the need for a CUDA-level software stack that can run commonly across NPUs from various vendors. He pointed out that if algorithms developed for a specific NPU have to be re-verified every time they are moved to a different chip, the speed of development and application would inevitably drop significantly due to the defense industry’s unique reliability and certification procedures. This is a practical argument that without a common platform, the actual application of domestic chips to weapon systems would be difficult.
In response, Baek Joon-ho, CEO of NPU frontrunner FuriosaAI, offered a different perspective. He clearly opposed the method of the state setting a platform standard and driving it uniformly. CEO Baek asserted, “It is completely unworkable for the state to suddenly create a standard and drive it. The chip companies must be the ones to drive it.” Citing the success of CUDA, he added, “There are no cases where a third party creates and drives a platform standard. That is a difficult thing for the government to do.”
When the moderator, Shim Seung-bae, a principal researcher at the Korea Institute for Defense Analyses (KIDA), asked if this meant ‘winner-takes-all’, CEO Baek went as far as to say, “The essence of platform competition is not much different from that.” CEO Baek expressed confidence that a platform standard would naturally emerge through market competition.
NPU company officials who attended the discussion each presented their own arguments. Rebellions Vice President Shin Sung-kyu diagnosed that the key in defense AI semiconductors is not the technology itself, but the supply chain and legal/regulatory issues. Mobilint CEO Shin Dong-ju introduced that his company will begin mass-producing on-device AI chips for drones and robots, as well as on-premise chips, starting this year, and said that the window of time for Korea to gain a competitive edge in the AI and defense semiconductor fields is not long. Hyperaccel CTO Lee Jin-won held up a newly developed chip and stated that if the defense sector’s requirements are clear, they can certainly create an optimized chip to match.
Meanwhile, amid ongoing controversies surrounding the U.S. Department of Defense’s (the Pentagon) adoption of artificial intelligence (AI), including the ousting of Anthropic and internal backlash at OpenAI, Google is expanding its opportunities to increase supply based on its relatively stable position.
On the 18th (local time), The New York Times (NYT), citing multiple sources, reported that Google Cloud CEO Thomas Kurian met with the Pentagon’s chief of AI adoption, Michael Emily, on February 26. Following this, Google appears to be expanding its position in a relatively comfortable situation compared to its competitors.
The U.S. Department of Defense has designated Anthropic as a ‘supply chain risk factor’ and is in the process of phasing it out of its systems. The two parties are also engaged in a legal battle. OpenAI filled the void left by Anthropic but is facing strong backlash from its own employees over this contract.
Amid this situation, the Department of Defense has agreed to expand its use of Google’s technology and will add Google’s ‘AI Agent’ to its unclassified network.
Jung Suk-yee
댓글
0