92% of Global Sovereign AI Dependent on Nvidia – A 'Paradox'; South Korea Attempts Diversification

"92% of Global Sovereign AI Dependent on Nvidia – A 'Paradox'; South Korea Attempts Diversification"

Counterpoint Report... "Hosting localized, chip supply concentrated"

SK Telecom and Rebellions' partnership in inference chip market also mentioned

While countries worldwide are pursuing their own artificial intelligence (AI) models, 'Sovereign AI' (domestic AI), for data sovereignty and other reasons, an analysis shows that the hardware is overwhelmingly dependent on the American company Nvidia.

Market research firm Counterpoint Research announced on the 5th (local time) in its 'Sovereign AI Large Language Model (LLM) Research Report' that 92.4% of independent AI worldwide use Nvidia AI semiconductors for training or inference.

Following Nvidia, AMD's market share, which holds the second position, was 4.1%, and the third-place provider Cerebras only reached 1.7%.

This is the result of analyzing over 170 models from 55 countries, including Asia-Pacific, the Middle East, Europe, and Africa, which are competing to build their own AI models and dedicated infrastructure, excluding the United States and China.

Countries are making efforts to secure data sovereignty by installing data centers within their regions, but it was found that they rely solely on Nvidia for AI chips.

Mark Einstein, a senior analyst at Counterpoint Research, pointed out, "We are witnessing the fact that hosting (server operation) is increasingly localized, whereas the source of AI chips is not. The majority of domestic LLM developers are drawn to the CUDA ecosystem (Nvidia's platform) and are utilizing Nvidia AI chips to train their LLMs."

However, within this monopolistic structure, South Korea appears to be moving to break free from Nvidia's influence.

South Korean AI startup 'Moreh' developed 'Llama-3-Motif-102B' based on semiconductors from Nvidia's rival, AMD, and Moreh's subsidiary, Motif, is also building LLMs through the AMD platform.

Another key example mentioned was the adoption of AMD Graphics Processing Unit (GPU) MI355 by another AI startup, Upstage.

In the inference chip market, South Korean semiconductor company Rebellions has partnered with SK Telecom to expand its market share, Counterpoint Research noted.

Although AMD's market share is still small, it is targeting niche markets in the European Union and Australia, while Cerebras is gaining prominence in the United Arab Emirates (UAE) market.

Qualcomm has partnered with Humaine in Saudi Arabia, and Google is providing its own Tensor Processing Units (TPUs) to AI startups in Israel. Japan utilizes its self-developed processor, 'Fujitsu A64FX' ARM Central Processing Unit (CPU), for LLM training.

However, it seems difficult to stop Nvidia's dominance, which has already captured the market.

Analyst Einstein explained, "As the domestic AI infrastructure market expands, Nvidia continues to forge extensive multi-billion dollar partnerships around the world."