It’s reported that there are only about 2,000 NVIDIA ‘H100’ GPUs in stock in Korea. SK announced that it will deal with ‘infrastructure’ to make up for this.
SK Telecom (CEO Young-sang Yoo) announced that it would lead changes based on AI infrastructure by unveiling its plan to construct an ‘Artificial Intelligence (AI) Infrastructure Super Highway’ on the ‘SK AI Summit 2024’ held for 2 days on the 4th and fifth.
About 35,000 people gathered at the location. The primary agenda emphasized by Big Tech CEO and SK at this event was ‘infrastructure’.
SKT CEO Yoo Young-sang said that there are statistics showing that while global big tech firms akin to Meta are securing a mean of 150,000 GPUs, Korea is securing only 2,000 H100 GPUs in total.
In relation to this, a survey of AI firms conducted by the Software Policy Research Institute under the Ministry of Science and ICT showed that a complete of 1,961 H100s were owned by 1,441 major domestic AI firms as of the top of last yr.
In keeping with what CEO Yoo said, H100 holdings have barely increased since then.
Nonetheless, Big Tech’s GPU holdings have been rapidly increasing recently. xAI announced that it had deployed 100,000 GPUs right into a single cluster within the second half of this yr, and Meta had built an infrastructure larger than that.
This is a component that makes you realize the difference in level between domestic and overseas when it comes to foundation model constructing capabilities. If there are fewer GPUs, the time it takes to develop the model will inevitably be longer.
To quickly resolve this GPU supply shortage, SKT plans to convert the Gasan data center situated within the metropolitan area into an AI DC and launch GPU as a service (GPUaaS), which provides GPUs within the cloud. Because it is difficult to buy GPUs directly, which means that they borrow the infrastructure of other cloud firms.
To this end, the corporate plans to launch GPUaaS based on NVIDIA’s ‘H100’ in December through cooperation with US cloud startup Lambda, and plans to introduce GPU ‘H200’ for the primary time in Korea in March of next yr. Although specific figures weren’t disclosed, we plan to steadily expand the amount in accordance with customer demand.
In this fashion, SK announced that it plans to construct a nationwide AI infrastructure centered on three axes: ▲AI data center ▲GPU cloud service (GPUaaS) ▲Edge AI, and based on this, enter the worldwide market with domestic and foreign partners.
We plan to open an AI data center (DC) test bed in Pangyo in December that mixes SK Group capabilities and partner solutions. The one test in Korea that implements cutting-edge AI semiconductors akin to NVIDIA’s latest chips and Hynix HBM, three sorts of next-generation liquid cooling solutions (chip contact cooling, tank-type immersion cooling, and chassis-type immersion cooling), GPU virtualization solutions, and AI energy optimization technology. He explained that Peter would take his place.
They said that they expect to draw greater than 50 trillion won in latest investment, create greater than 550,000 jobs, create greater than 550,000 jobs, achieve economic effects of greater than 175 trillion won, and foster regional AI high-tech industries within the mid to long run through gigawatt (GW)-level AI DC.
Plans to implement ‘Korean Sovereign AI’ were also announced. SKT plans to take a position a complete of 100 billion won starting in 2025 to mix Insurrection’s NPU (neural network processing unit), SK Hynix’s HBM, and various AI DC solutions owned by SK Telecom and its partners.
Lastly, we plan to introduce edge AI that may solve the gap between AI DC and ‘on-device AI’ by utilizing communication infrastructure connected across the country. Currently, SKT is constructing AI DC and developing customized servers using communication infrastructure with global partners. As well as, we’re working with various domestic and foreign firms to pursue edge AI service development technology proof of concept (PoC) in six areas, including healthcare, AI robots, and AI CCTV.
SK Chairman Chey Tae-won also appeared on the event and emphasized the importance of infrastructure and energy.
Specifically, it was explained that a knowledge center of not less than 10 gigawatts (GW) is required to construct one large language model (LLM). Considering that greater than $40 billion (KRW 54.956 trillion) is required to construct a 1GW level data center, the associated fee can be about 10 times higher, he added.
Ultimately, he said, issues akin to the quantity of power, transmission methods, and carbon reduction should be resolved. Accordingly, SK emphasized that it’s considering bringing in an independent type of energy fairly than a grid form. A small reactor type is currently being developed and is anticipated to be a great solution for future AI data centers.
Next, the corporate announced that it’s developing the bottom technology for its own AI data center in an effort to have AI data center heat control (temperature control, cooling) technology.
As well as, Microsoft CEO Satya Nadella, NVIDIA CEO Jensen Huang, and TSMC CEO CC Wei sent congratulatory videos.
Specifically, Chairman Choi said, “The last time I met with CEO Jensen Huang, I used to be asked to increase the availability of HBM4 by six months,” and added, “I agreed to accomplish that.”
He then emphasized, “SK is one among the few firms that may encompass all elements of AI, including infrastructure, energy, and services.”
Reporter Jang Se-min semim99@aitimes.com
