[Edaily Reporter Han Kwangbeom ] While countries around the world are accelerating efforts to build “sovereign AI” to ensure data sovereignty and technological independence, it has been revealed that they remain absolutely dependent on U.S.-based NVIDIA in the field of AI semiconductors, a core infrastructure.

According to the “Sovereign AI Large Language Model (LLM) Research Report” released on the 10th by market research firm Counterpoint Research, 92.4% of the approximately 170 sovereign AI LLMs developed in 55 countries worldwide—excluding the U.S. and China—are trained and perform inference using NVIDIA’s AI semiconductors.

Counterpoint Research analyzed the status of sovereign AI infrastructure across more than 80 countries, including the Middle East, Western Europe, Eastern Europe, Asia-Pacific, Central Asia, Africa, Latin America, and Canada. While there is a growing trend toward building data storage and server hosting capabilities within domestic data centers, dependence on the hardware and the proprietary software platform “CUDA” ecosystem that powers them appears to be intensifying.

Amid NVIDIA’s proprietary ecosystem, Counterpoint Research highlighted key examples of Korean companies collaborating with latecomers to lead the trend of moving away from NVIDIA. The analysis found that AMD, a latecomer, is expanding its presence in the Asia-Pacific region by leveraging its collaboration with Korean companies.

“Moreh,” a South Korean AI startup and member of the Foundation Model Consortium, developed “Llama-3-Motif-102B,” an AMD-based large language model (LLM), through its subsidiary Motif. Similarly, “Upstage,” a leading South Korean AI company, is expanding its AMD-chip-based infrastructure by adopting the “MI355.”

The report also highlights the moves by Korean companies targeting the AI inference market, moving beyond the capital-intensive field of model training. Rebellion, a Korean AI semiconductor fabless company, has partnered with SKTelecom(017670)to enter the domestic sovereign AI inference market and is accelerating the development of an alternative ecosystem.

Looking at the global market as a whole, AMD secured a 4.1% market share, driven by the European consortium’s “LUMI” supercomputer, its partnership with Australia’s Maincode, and cases in South Korea. Cerebras recorded a 1.7% market share based on a $900 million investment partnership with UAE tech firm G42, the construction of the “Condor Galaxy” AI supercomputer, and its collaboration with India’s C-DAC. In addition, Japan’s “Fugaku LLM,” based on Fujitsu’s A64FX, and Israel’s AI21’s “Zamba 2,” based on Google Cloud TPUs, made their mark in the niche market.

Mark Einstein, Research Director at Counterpoint Research, stated, “Our analysis of sovereign AI infrastructure shows that while the localization of server hosting is strengthening, independence from AI chip suppliers has not been achieved,” adding "While NVIDIA is expanding its partnerships—worth trillions of won—and broadening its reach through inference and the establishment of sovereign AI factories, latecomers are actively targeting the inference market by leveraging tokenomics and ecosystem support," he added.