Korea's H1 Venture Investment Jumps 54% as AI, Bio Draw Capital

* Editor's note: AI PRISM (Personalized Report & Insight Summarizing Media) is an AI-based personalized news recommendation and summarization service developed with support from the Korea Press Foundation. It selects and provides six customized news items by reader type.

1. First-Half Venture Investment Pours Into AI and Chips at Record 8.9 Trillion Won

- Key summary: New venture investment in the first half of this year came to 8.8676 trillion won, up 54.3% from the first half of last year and a record high. According to the "2026 First-Half New Venture Investment and Venture Fund Formation Trends" released by the Ministry of SMEs and Startups on the 19th, 1.86 trillion won, or 21% of total investment, went into the AI-led ICT services sector, followed by 1.5359 trillion won in electronics, machinery and equipment and 1.5041 trillion won in bio and medical. New venture fund formation rose 33% to 8.4366 trillion won, the second-highest first-half figure on record, driven largely by financial institution contributions that climbed 54.9% to 2.6061 trillion won. Kim Bong-deok, director general of venture policy at the ministry, stressed the role of the private sector and said performance would trend upward in the second half as well. Yoon Kun-soo, CEO of DSC Investment (241520.KQ), cited the formation of large funds but said whether recovery markets such as the KOSDAQ become more active would be a variable.

2. Unitree Soars Sevenfold Right After Listing as Robotics Firms Gain Fundraising Momentum

- Key summary: Shares of Chinese humanoid robot maker Unitree surged as much as 629.44% above their offering price of 150.8 yuan to 1,100 yuan immediately after listing on the Shanghai Stock Exchange on the 19th, before closing at 845 yuan, up 460.34%. Through the listing, Unitree plans to raise 6.099 billion yuan (about 1.28 trillion won) for robot production, and its market capitalization stood at 341.7 billion yuan (about 71.29 trillion won) as of the close. The stock held by founder Wang Xingxing, who owns a 33.36% stake, was valued at more than 100 billion yuan based on the opening price, while Sunwei Capital and DeepSeek founder Liang Wenfeng are estimated to have gained unrealized profits of 15.2 billion yuan and 1.1 billion yuan, respectively. Still, with a price-to-earnings ratio (PER) of 219.23 times at the offering price — about six times the robotics industry average of 38.56 times — some point to the high valuation as a burden.

3. "After Technology Comes Profit": Robotics Firms Court Sales and Finance Experts

- Key summary: Korean robotics firms that have focused on technology development are recruiting outside sales and finance figures one after another. Wearable robot specialist Angel Robotics (455900.KQ) reorganized its management structure in the first half of this year by successively bringing in executives overseeing finance, sales and quality, and it is expanding exports to Southeast Asia, including Thailand, Malaysia and Vietnam, while preparing to enter the U.S. and European markets. Industrial robot software firm CMES Robotics (475400.KQ) created a chief financial officer (CFO) position in February, and autonomous driving robot software firm Klobot recruited Presidents Ryu Jung-hoon and Son Jun-bae, among others, and hired sales staff for humanoid solutions. Meanwhile, an industry official said that even among listed firms, quite a few post revenue of around 5 billion won or run losses, adding that companies are at a stage of trying to secure profitability by increasing recurring revenue and overseas business and strengthening financial management.

4. Robots With "Vision AI": Judging Danger on Their Own From 750 Million Images

- Key summary: "Argos," a quadruped patrol robot developed by Lion Robotics, relays on-site danger situations to a control center in real time by linking with the AI robot control platform of vision AI specialist Intellivix. Existing systems required a control center to manually analyze video after an incident, taking at least several minutes, but the AI control system enables real-time response. Choi Eun-soo, CEO of Intellivix, said the company is building safety-focused AI that understands the context of danger using about 750 million images accumulated in the field over 26 years. The two companies also plan to deploy Argos at SK Incheon Petrochem's plant by taking part in a quadruped patrol robot demonstration under the Ministry of Science and ICT's "Hyper-AI Network Infrastructure Project." Intellivix also won a "Public AI CCTV Conversion Project," in which the Korea Expressway Corporation participates as the demanding agency, last month.

5. Mobilint NPU Installed on Navy Ships for AI-Based Video Analysis

- Key summary: AI chip startup Mobilint said on the 19th that it is supplying its own AI chip-based inference platform to the Navy's "NPU-Based Intelligent Video Surveillance System Pilot Project." The supply marks a case of Mobilint's AI chips, which have expanded their application mainly across public and industrial sites, extending into the defense sector. Through the project, an AI-based video analysis environment will be built on about 50 Navy ships, with AI analyzing intrusion, fire, flooding, falls into the sea, abnormal behavior and abandoned objects in real time to support responses through control screens and alarms. Mobilint said its business is expanding beyond technology verification into actual product supply and commercial application, and that it plans to broaden its customer base into various edge AI environments going forward.

6. CCTV Footage as a Robot Textbook: Physical AI Widens Learning Models

- Key summary: Video capturing human behavior from external sources such as CCTV is emerging as key data for training physical AI, because gathering training data by physically moving robots takes considerable time and cost. A research team led by Professor Zhou Daquan of Peking University built "HumanNet," a training video dataset totaling 1 million hours, and published a paper in May this year. A model pre-trained on 1,000 hours of HumanNet video outperformed a model trained on 100 hours of actual robot data. A research team led by Professor Joo Jae-gul of the Kim Jaechul Graduate School of AI at the Korea Advanced Institute of Science and Technology (KAIST) developed an AI model called "EgoX" that generates a person's actual field of view from third-person video alone, while Nvidia and Google are also using large-scale video data to train robots and to advance models for understanding the physical world.