SK Telecom's artificial intelligence (AI) strategy is shifting from model development to "infrastructure that runs AI." The company judges that the center of gravity in the Generative AI race is moving from training giant models to inference infrastructure that operates real services stably and cheaply.
Lee Jae-shin (50), SK Telecom's head of AI business development (executive vice president), met with ChosunBiz on the 15th and said, "The center of gravity in the AI industry is shifting from training to inference," adding, "Competitiveness will depend on how efficiently AI data centers and inference infrastructure are operated going forward."
Lee, who holds a master's degree in accounting from Syracuse University in the United States, joined SK Telecom in 2001 and has held key posts across business development and strategy. Since 2021, he has led AI business development under SK Telecom's AI CIC, overseeing investment, business development, strategy, and global big tech partnerships. He led investments in Anthropic and Perplexity and has tied them to the AI personal assistant service adot and the in-house foundation model A.X K1 development. Recently, he has been accelerating a full-stack strategy that links domestic AI Semiconductor and AI data centers.
His criteria for AI investment are clear. It is not simple financial investment; it must align with SK Telecom's business. Lee said, "When we invest, strategic synergy must be proven," adding, "No matter how good a company is, we do not invest if there is no possibility of business cooperation with us."
The Anthropic investment illustrates this principle. From the time the Generative AI market began in earnest, SK Telecom believed it should pursue both in-house model development and collaboration with global leading model corporations. Lee said, "Anthropic first asked why it wanted us to invest and what we could gain from each other," adding, "Its stance of weighing long-term cooperation potential over simple fundraising was impressive."
◇ Bolstering inference infrastructure by adding Rebellions NPUs to Nvidia GPUs
SK Telecom recently began cooperating with Nvidia on a DSX platform-based "full-stack AI cloud." The plan is to secure Nvidia's latest graphics processing unit (GPU) Blackwell and next-generation platform Vera Rubin ahead of time to meet demand for AI training and high-performance inference.
At the same time, SK Telecom is pursuing a strategy that does not leave all AI computation to GPUs. Depending on the workload, it uses a "heterogeneous computing hybrid" structure that combines GPUs and Neural Processing Unit (NPU). GPUs handle training and high-performance computation, while NPUs with higher power efficiency take on part of the repeatedly occurring service inference.
Lee described the role of Rebellions NPUs as "a complement rather than a replacement for GPUs." Lee said, "We are taking a hybrid strategy that divides areas where GPUs excel and areas where NPUs are more efficient," adding, "GPUs remain important for training and general high-performance computation, but in inference, where already-built models are repeatedly run in real services, power efficiency and operating cost are far more important."
SK Telecom's recent move with Arm and Rebellions to develop a combined CPU-NPU AI inference server solution is in the same vein. The three companies plan to develop a server that combines Arm's data center processor "Arm AGI CPU" with Rebellions' AI inference accelerator "Rebel Card," and verify it at SK Telecom's AI data center.
AI inference is the stage where a trained model responds to user requests in real services. Individual computations may be lighter than training, but because user requests repeat constantly, as cumulative usage grows, power efficiency, latency, stability, and operating cost become key competitive factors.
Lee said, "An AI data center is not a provider that simply supplies servers," adding, "It is core infrastructure that must run AI models and services stably around the clock without interruption."
◇ Full-stack AI that ties models, services, and data centers together
SK Telecom views its AI business as a structure where models, services, and infrastructure are combined. It ties together the in-house foundation model A.X K1, the AI personal assistant service adot, the GPU-based AI cluster "Haein," and AI data centers into one.
Lee said, "Some companies can do only infrastructure, and some can do only services," adding, "Our differentiator is that we have both and can connect them." Lee said, "Among telecom operators worldwide, there are very few that have the full stack across services, infrastructure, and models."
The company also sees the operating experience it has built up as a telecom operator as a competitive edge in AI infrastructure. AI services must run continuously without outages and handle massive traffic stably. The company says its capabilities in outage response, quality control, and traffic management accumulated in network operations can be applied to running AI data centers.
The Perplexity investment shows the services side of the full-stack strategy. SK Telecom concluded that Perplexity's AI search capabilities could be integrated into adot. Lee said, "Perplexity had strengths in search, and we were persuaded from the start to collaborate by putting this function into the adot service."
◇ Meeting global demand with the Ulsan AI data center
The hyperscale AI data center that SK Telecom is building in Ulsan with Amazon Web Services (AWS) is also a key pillar of its inference infrastructure strategy. It is cited as a symbolic project in the domestic AI infrastructure market, as a global big tech company and a domestic telecom operator are establishing a large AI data center in Korea.
Lee said, "It is meaningful in that a global big tech company is building a data center in Korea," adding, "It is also meaningful in that it will address overseas demand as well as domestic demand." Lee said, "It does not end with building a data center with AWS; we can accumulate operating know-how while meeting the conditions required by big tech."
Power, land, and network location were behind the choice of Ulsan. AI data centers require more power than general data centers. In the greater Seoul area, land and power grid burdens are high, but Ulsan can secure large tracts of land and is close to Busan's international submarine cable landing stations. The company also says it has an advantage in power procurement because an LNG power plant run by an SK Group affiliate is nearby.
Lee said, "In Ulsan, land for building a data center is important, and it is also a suitable location in terms of submarine cables and network connectivity," adding, "More important than building several small data centers was whether we could expand later and whether land and power were secured."
From a shareholder's perspective, investing in AI data centers may appear to be a large expense. But SK Telecom views it as long-term growth infrastructure beyond its core telecom business. As AI services increase, structural demand for computing for training and inference will grow, and the value of operators that run this stably and efficiently will also rise, the company judges.
Lee said, "The key question for shareholders is whether this investment is a simple expense or a foundation for future growth," adding, "SK Telecom views AI data center investment as long-term growth infrastructure beyond its core telecom business."
◇ Aiming at sovereign AI demand… next investments also in infrastructure
SK Telecom expects demand to grow for sovereign AI infrastructure that can be controlled domestically in the public, finance, telecom, and research sectors. The company judges that where data are stored, who operates the systems, and whether security and control are possible will be important criteria when choosing AI infrastructure.
Lee said, "Trust in Data Sovereignty, data center location, security, and whether operations are domestic is important," adding, "In the public, finance, telecom, and research sectors, there will be growing demand for AI infrastructure that can be controlled domestically."
Future investment directions also tilt toward infrastructure. Lee said, "For now, we need to focus a bit more on infrastructure," adding, "We are looking at companies with technologies and software that can operate data centers more efficiently."
However, rather than immediately choosing a specific country and pushing AI exports, the company is prioritizing improving the completeness of models, services, and infrastructure domestically. Lee said, "More important than export results like which country we broke into first is making AI models more sophisticated and adding diverse services," adding, "We believe it is better to advance overseas after refining these parts further."