South Korean Government Continues Efforts to Secure Cutting-Edge GPUs
On February 18, 2025, the ruling People Power Party and the South Korean government held a policy consultation meeting, agreeing to increase the country’s inventory of state-of-the-art GPUs tenfold by the first half of next year, reaching a total of 20,000 units. To achieve this, the government plans to allocate a supplementary budget of 2 trillion KRW.
Additionally, the government aims to secure 10,000 high-performance GPUs within this year to accelerate the launch of a national AI computing center. Simultaneously, it seeks to reform the scholarship system for university students engaged in research on AI, semiconductors, secondary batteries, and biotechnology, with a long-term goal of providing tuition-free education. Discussions are also underway to offer tuition assistance and overseas study funding to students conducting research in strategically significant technologies.
The ruling party and the government plan to expand the budget to propel South Korea into the ranks of the world's top three AI powerhouses. They are currently reviewing a proposal to augment the 1.2 trillion KRW AI infrastructure budget, which was agreed upon last year in the National Assembly’s Science, ICT, Broadcasting, and Communications Committee, by an additional 800 billion KRW—bringing the total allocation to 2 trillion KRW.
Meanwhile, the Democratic Party welcomed the government's AI competitiveness initiative but raised concerns that the plan to secure 20,000 GPUs might be an exaggeration of the previously proposed government budget target of 18,000 units. The opposition party also urged the administration and ruling party to swiftly present a detailed AI-focused supplementary budget plan that prioritizes public welfare.
Source:
🔗 Chosun Ilbo: "AI Supplementary Budget of 2 Trillion KRW... 10-Fold GPU Expansion by Mid-Next Year"
The Data Center Construction Process for Grok 3
On February 18, 2025 (KST), Elon Musk’s AI company, xAI, hosted a live demonstration of the Grok 3 model. The discussion surrounding the development of Grok 3’s data center underscored that simply amassing high-performance GPUs is not enough. The successful implementation of such AI infrastructure requires securing adequate space, ensuring stable power supply and transmission, managing power fluctuations, and maintaining an efficient cooling system.
🔗 Watch the Grok 3 Demo on YouTube
During an online event held via social media platform X (formerly Twitter), Musk emphasized that Grok 3 delivers significantly enhanced performance compared to its predecessor, Grok 2.
According to xAI, Grok 3 boasts over ten times the computing power of its previous version, enabling it to solve complex problems with unprecedented efficiency. The company attributed this breakthrough to Grok 3’s ability to learn from vast datasets and leverage advanced reasoning capabilities. At the event, xAI showcased real-time game creation and orbital simulation processing, demonstrating the model’s enhanced capabilities.
A key factor behind Grok 3’s development was xAI’s massive investment in building its own data center. The company currently operates approximately 200,000 GPUs, with over 100,000 of them being Nvidia H100 units used in the early development phase. Remarkably, within just 92 days, xAI doubled its GPU capacity, rapidly scaling up the AI model.
Building Grok 3’s data center was a formidable challenge. Initially, xAI had no plans to construct its own facility. However, after consulting external data center providers and being told it would take 18 to 24 months to deploy 100,000 GPUs in one location, xAI decided to undertake the project itself. In an extraordinary feat, the company successfully designed and operationalized a data center capable of running 100,000 GPUs within just 122 days.
Key Challenges in Data Center Construction
Facility Acquisition
Since new construction was not feasible within the required timeline, xAI sought an existing, well-maintained structure. The company identified and repurposed an abandoned Electrolux factory in Memphis, a city with historical significance as both the birthplace of Elvis Presley and the namesake of the ancient Egyptian capital.Power Supply
Operating the facility required at least 120MW of power, but the building’s existing infrastructure could only support 15MW. The goal was to scale up to 250MW—enough to power 200,000 GPUs. To achieve this, xAI leased large-scale generators, positioning them on one side of the building while deploying massive mobile cooling units on the opposite end. The cooling demand was so high that the company reportedly rented a quarter of all available mobile cooling capacity in the U.S.GPU Deployment and Cooling
Given the high GPU density, xAI opted for liquid cooling, an uncharted territory for large-scale AI data centers. Implementing an efficient liquid-cooling system required extensive piping installation and optimization.Power Fluctuations
The biggest technical hurdle was power instability. As 100,000 to 200,000 GPUs rapidly switched between active and idle states—sometimes within 100 milliseconds—power fluctuations disrupted generator operations. To mitigate this, xAI integrated Tesla’s Megapacks as power buffers. However, the default software of the Megapacks was not designed for such rapid energy fluctuations, prompting xAI and Tesla to co-develop new firmware for stable power management.Network Stability
Ensuring seamless communication across all GPUs posed another significant challenge. Engineers worked through the night, debugging network cables until they discovered a BIOS configuration issue. By comparing LSPCI outputs between functioning and non-functioning machines, they finally resolved the network instability at 4:20 AM.
Through these efforts, xAI successfully deployed and activated 200,000 GPUs within an unprecedented 122-day timeframe.
Lessons for South Korea from xAI’s Experience
As South Korea strengthens its AI infrastructure, it must recognize that merely stockpiling GPUs is insufficient. The Grok 3 case demonstrates that a holistic approach—including space acquisition, power supply and fluctuation management, cooling solutions, and network stability—is crucial for building an effective AI ecosystem.
For South Korea to emerge as one of the top three AI powerhouses, investment must extend beyond GPU procurement. It must also focus on data center design and operations, large-scale energy management systems, high-performance networking, and advanced cooling technologies. Additionally, to sustain long-term AI industry growth, the country must invest in AI semiconductor development, software optimization, and cloud computing infrastructure while cultivating top-tier talent capable of managing these advancements.
Expanding scholarships for AI research and providing financial support for students studying strategic national technologies will enhance the country's AI competitiveness. Furthermore, fostering an environment that attracts global AI talent and improving domestic research conditions will be key to establishing a world-class AI research ecosystem.
South Korea’s Latest AI Ambitions: A 3GW AI Data Center?
A group of investors recently announced plans to build one of the world's largest AI data centers in South Korea. This move highlights South Korea’s potential to become a global AI data center hub amid soaring global demand for AI infrastructure.
The project, projected to cost up to $35 billion (approximately 46 trillion KRW), aims to construct a facility capable of utilizing up to 3 gigawatts (GW) of power—three times the scale of the "Stargate" AI data center (1GW) currently under development by OpenAI and SoftBank in the U.S.
🔗 WSJ: "AI Data Center with Up to 3GW Planned for South Korea"
However, skepticism remains. The lead investor, Stock Farm Road, is a little-known entity, and Brian Koo—its co-founder and grandson of LG Group's founder—has faced significant past investment failures. Crucially, this project is still in its early investment proposal phase, with no confirmed progress, necessitating careful scrutiny.
댓글
댓글 쓰기