[AI Talent Lost to Big Tech③] Meta Has 350,000, Korea Only 2,000... Securing GPUs Is Key to Bringing AI Talent Back

Seeking Better Research Environments and New Challenges
U.S. Hired 400,000 Professionals with Visas Last Year
Korea Moves to Secure 10,000 GPUs Through Supplementary Budget

The outflow of artificial intelligence (AI) talent from leading domestic universities to overseas is not solely due to differences in financial compensation. The shortage of AI semiconductors, such as graphics processing units (GPUs), which are core infrastructure for AI, is also cited as a major reason for choosing overseas big tech companies.

Infrastructure like AI semiconductors is absolutely essential for proper AI model development, training, and service operation. However, in Korea’s industrial and research sectors, such infrastructure remains insufficient, leading many to choose the United States in pursuit of a better research environment.

According to Air Street Capital, a U.S. venture capital firm, as of July 18, Meta owns 350,000 units of Nvidia’s latest AI accelerator, the H100. Elon Musk’s xAI and Tesla possess around 100,000 and 35,000 H100s, respectively. For the previous generation accelerator, the A100, Meta alone owns about 210,000 units, and China’s DeepSeek has secured 10,000 units. In contrast, it is reported that only about 2,000 H100s have been deployed in all of Korea.

While the Korean government has set a goal of becoming one of the world’s top three AI powers and has begun efforts to secure GPUs and support AI development at the government level, those in the field?including schools and companies?say they have yet to feel any tangible changes.

An official at a domestic company developing its own large language model (LLM) commented, “The level of AI service development is essentially determined by the level of GPU availability. Some companies facing difficulties in securing additional GPUs are developing AI services using alternatives such as neural processing units (NPUs) or Google’s custom-developed tensor processing units (TPUs).”

The level of AI infrastructure development is closely linked to talent acquisition. Amazon is not only one of Nvidia’s largest customers but also the company that has received the most H-1B professional visas in the United States. In addition to GPUs, Amazon is developing ultra-large AI supercomputers using its own AI-specific chip, “Trainium.” The company is also directly designing and operating eco-friendly power infrastructure needed in the AI era, focusing on building a sustainable AI ecosystem.

As the AI development ecosystem takes shape, talent is rapidly concentrating in the United States. From January to September last year, Amazon topped the list of companies receiving H-1B visas for professional workers from the U.S. government, with 9,265 recipients. Google followed with 5,364, Meta with 4,844, Microsoft (MS) with 4,725, and Apple with 3,873.

Last year, about 400,000 people either renewed or newly obtained H-1B visas, a 123% increase compared to the year 2000. The proportion of H-1B workers holding a master’s degree rose from 31% in 2000 to 57% in 2021, demonstrating that the U.S. is absorbing highly skilled professionals through this visa program. Since 2012, more than 60% of H-1B workers approved each year have held computer-related jobs, including AI. In 2023, this proportion reached 65%, with the average annual salary reported at $123,600 (about 172 million KRW).

The government is also beginning efforts to secure GPUs, the foundation of the AI ecosystem. The Ministry of Science and ICT recently completed the application process for operators to import about 10,000 GPUs using the first supplementary budget, totaling 1.46 trillion KRW. Four companies?Naver Cloud, Coupang, Kakao Enterprise, and NHN Cloud?have reportedly submitted applications. The final operator is expected to be selected this month after a review of which company can bring in abundant resources at the most reasonable price and operate them stably.