AI Startups Enter GPU 'Subscription' Era as GPUaaS Becomes Essential Infrastructure

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

1. Time and Cost Burdens Ease… AI Startups Flock to 'GPUaaS Subscriptions'

- Key Summary: AI startups are adopting GPUaaS (GPU as a Service), a cloud-based subscription service, as essential infrastructure instead of purchasing graphics processing units (GPUs) directly. Medical AI startup SoombitAI used Elice Group's GPUaaS to train on 14 million chest X-ray data sets. LLM (large language model) developer Upstage used Samsung SDS's Nvidia H100-based service to simultaneously resolve upfront investment costs and GPU supply constraints. AI security startup Theori Korea chose NHN Cloud to avoid time-zone and language barrier issues with foreign cloud providers, illustrating a growing trend of adopting GPUaaS from domestic cloud companies for their reliability. As technologies such as AI agents and world models advance, GPU demand is expected to grow further, and the GPUaaS market's growth trajectory is likely to continue for the foreseeable future.

2. New Titles Pour Out… But K-Game Companies See Diverging Results

- Key Summary: Major new titles from Korean game companies launched in succession, but their commercial performances diverged sharply. Pearl Abyss's 'Crimson Desert' surpassed 4 million cumulative global sales within 12 days of launch after eight years of development and 200 billion won in investment, fully recouping its development costs within just four days. Despite harsh initial reviews, the company reversed sentiment through eight rapid patches and responsiveness to user feedback, while short-form video content maximized promotional impact. In contrast, Devsisters' 'CookieRun: Oven Smash' debuted at No. 1 on the app store but saw its stock price slide to a 52-week low due to optimization defects including device overheating and network instability, making a comprehensive revision of its business strategy unavoidable.

3. Driverless AI Trucks Roam Mines Until 300 Tons Are Loaded

- Key Summary: Applied Intuition, a U.S. autonomous driving company valued at $15 billion (approximately 22.68 trillion won), has chosen mines as the front line for implementing Physical AI, a technology combining AI with the physical world. The company partnered with equipment manufacturer Komatsu to deploy 300-ton autonomous trucks at an Arizona mine, using a surround-view system with six wide-angle cameras to enable real-time hazard detection without safety personnel. The structural problem of essential mineral demand quadrupling by 2040 while the workforce for dangerous mining operations shrinks rapidly serves as the company's growth driver. Applied Intuition has been collaborating with LG Innotek (011070.KS), Hyundai Motor (005380.KS) and Fortytwo Dot, and is rapidly expanding global partnerships by signing an autonomous driving partnership with Germany's Traton and unveiling a second-generation autonomous truck with Japan's Isuzu.

4. "Venture Capital from Financial Sector Needed to Overcome Crises Through Innovation"

- Key Summary: At the '2026 Korea Best Banker Awards' ceremony hosted by the Seoul Economic Daily and sponsored by the Financial Services Commission (FSC) and Financial Supervisory Service (FSS), financial authorities publicly emphasized that funds must flow into strategic industries such as AI and semiconductors. Shin Jin-chang, secretary general of the FSC, said, "Venture capital and patient capital from the financial sector are needed to achieve innovation." Kim Sung-wook, deputy governor of the FSS, urged that funds be channeled into new growth industries such as AI and semiconductors without concentration in real estate. Jeong Jin-wan, president of Woori Bank, received the top 'Best Banker' award, while KB Financial Group (105560.KS) (Best Finance), Shinhan Financial Group (Best Consumer Protection) and Hana Financial Group (086790.KS) (Best Inclusive Finance) were named as category winners. A consensus that finance must function as an innovation engine amid external crises including Middle East conflicts and prolonged low growth is spreading across financial authorities and the broader industry.

5. First-Ever Semiconductor Model Not Found in Nature Proposed… Opens New Horizon for Low-Power 'AI Chips'

- Key Summary: Professor Shim Woo-young of the Department of Materials Science and Engineering at Yonsei University, selected as the April recipient of the 'Korea Science and Technology Award' hosted by the Ministry of Science and ICT (MSIT), developed the world's first new structure of 'III-V compound semiconductor' materials not found in nature. Professor Shim applied Pauling's rules, established in 1929, to semiconductors to create an ultra-fine gap called a 'van der Waals spacing' between layers, designing a structure in which electrons and ions can move simultaneously. This enables both computation and memory functions within a single material, and is recognized as a core foundational technology for next-generation low-power AI semiconductors and neuromorphic computing, a computing method that mimics the way neurons in the brain operate, by reducing power waste in data transfer processes. Professor Shim plans to apply this new material to the AI semiconductor field while expanding research into a new computational method that utilizes the actual movement of ions.

6. Nvidia Invests 3 Trillion Won in Marvell… Connecting Big Tech's Custom Chips Too

- Key Summary: Nvidia invested $2 billion (approximately 3 trillion won) in Marvell Technology, a fabless semiconductor design company, and signed a strategic partnership. The two companies will combine Nvidia's chip interconnect networking technology, NVLink Fusion, with Marvell's custom AI semiconductors (ASICs) to enable big tech companies to integrate their self-designed chips into Nvidia's infrastructure ecosystem. They will also jointly develop silicon photonics, which transmits data using light instead of electrical signals, aiming to simultaneously improve transmission speed and reduce power consumption. Marvell saw fiscal year 2026 revenue increase 42.0% year-on-year, driven by surging demand for custom AI chips from hyperscalers such as Amazon. The investment is interpreted as a strategy by Nvidia to solidify its central position in AI computing even as big tech companies increasingly design their own chips.