[Drifting Korean-style AI] ① One Year Since HyperCLOVA X Debut... Widening Gap with Big Tech

Naver's R&D Expenses Reach 898.8 Billion KRW in H1
R&D Investment Ratio to Sales Declines
"Computing Power and Capital Make Competition Difficult"

Image source=Yonhap News

Naver's large language model (LLM) 'HyperCLOVA X,' recognized as a leading domestic artificial intelligence (AI) model, will mark its first anniversary on the 24th. The AI ecosystem survey report by the Center for Research on Foundation Models (CRFM) at Stanford University in the United States highlighted Naver HyperCLOVA and its improved version HyperCLOVA X as representative AI models of South Korea. However, as it gradually falls behind in competition with overseas big tech companies backed by substantial capital, it is being described as a "faded year."

According to the Financial Supervisory Service's electronic disclosure system on the 19th, Naver's research and development (R&D) expenses for the first half of this year totaled 898.8 billion KRW. Although this is a decrease of 66.2 billion KRW compared to 965 billion KRW in the same period last year, it remains relatively high compared to other domestic companies included in the AI ecosystem survey. SK Telecom, which owns the AI model 'A.Dot X,' recorded 190.1 billion KRW, and KT, which developed 'Mideum,' recorded 104.3 billion KRW. Naver's R&D expenses have increased annually from 1.3321 trillion KRW in 2020 to 1.655 trillion KRW and 1.809 trillion KRW in 2021 and 2022, respectively. Last year, it was confirmed at 1.9926 trillion KRW.

Of course, not all of Naver's R&D expenses are invested in AI research and development. However, since 2019, Naver has been conducting research and development on NSML (Naver Smart Machine Learning), an AI technology R&D platform. Additionally, since 2021, it has added research on overcoming and advancing large-scale AI limitations and the development of HyperCLOVA, a large-scale language model for NLP (natural language processing) services. Beyond this, it is also conducting R&D on creative multimedia generation AI technologies and the advancement of the existing Photo AI inference system in operation.

HyperCLOVA X attracted significant attention when it was first introduced last year. The Center for Research on Foundation Models at Stanford University described HyperCLOVA X as "a competitive LLM product line developed specifically for the Korean language and culture, based on strong capabilities in English, mathematics, and coding."

Choi Soo-yeon, CEO of Naver, stated upon unveiling HyperCLOVA X, "Thanks to the advanced technology and high-quality data steadily built through investing 22% of our revenue in R&D," adding, "Our decades of experience in understanding users, service operation know-how, and technical capabilities all serve as a strong foundation that powerfully supports the competitiveness of HyperCLOVA X, the backbone model of current generative AI."

However, compared to overseas big tech companies, Naver is clearly at a disadvantage in terms of capital. Considering that AI success depends heavily on massive data, capital is a critical factor in enhancing AI competitiveness.

The proportion of Naver's R&D investment relative to total revenue is also decreasing. It was 22% in 2022, but dropped to 20.6% last year and further to 17.5% in the first half of this year. OpenAI reportedly invested over 100 billion KRW solely in training the GPT-4 model. Additionally, although Meta has not disclosed the development costs of its LLM LLaMA 3.1, which was unveiled last month, it is estimated that at least 650 billion KRW was spent on chip purchases such as NVIDIA's H100 graphics processing units (GPUs).

There are concerns that falling behind in capital competition also limits performance. In particular, technologies such as multimodal (simultaneous processing of image, video, audio, and other data) remain insufficient. Google owns the multimodal LLM Gemini series, and OpenAI has developed multimodal AI like GPT-4o. In contrast, Naver's conversational AI agent service Clova X, which incorporates HyperCLOVA X, still primarily relies on text or document-based inputs.

Professor Lee Sung-yeop of Korea University's Graduate School of Technology Management commented, "The development is moving toward multimodal forms or general-purpose AI resembling humans, but Naver seems not to have reached that level yet," adding, "Lack of computing power and insufficient capacity to invest massive capital are factors making competition difficult."

A Naver representative said, "Uploading images to Clova X and having conversations based on them is our top priority, and an update will be made soon," adding, "Various tasks such as image or voice generation through AI are scheduled for updates at a later stage."