The technology race among the three telecom operators in AI data centres is moving to a stage of testing real operational capabilities. As centres grow, heat generation and power consumption surge. Cooling efficiency, stable power supply and network connectivity are emerging as key factors that will determine competitiveness.

On Aug. 10, the industry said the three telecom operators, including SK Telecom, KT and LG Uplus, are speeding up efforts to apply their own capabilities to AI data centres. They are not only expanding capacity. They are also pooling capabilities across servers, semiconductors, cooling, power and networks to improve AI computing efficiency.

SKT brings full-stack technology into hyperscale AI data centres.

SKT has moved most aggressively among the three telecom operators to expand its AI data centre business. Last month, it launched SK Hyper, a dedicated AI data centre unit, and decided to invest 750 billion won by 2030. Centered on SK Hyper, it plans to open AI data centres with a total capacity of 5 gigawatts in stages by 2029 and expand to 15 gigawatts by 2035.

The first test is the Ulsan AI data centre. SKT is working with Amazon Web Services to build a large AI data centre in Ulsan. It aims for synergy with group energy affiliates, including SK Hynix semiconductors and SK Broadband's network capabilities.

SKT is also moving to secure technology to raise AI computing efficiency. It is partnering with global chip designer Arm and domestic AI chip company Rebellions to develop a server solution. The solution will include Arm's AGI CPU and Rebellions' AI inference accelerator, Rebellcard, on the same platform. The effort aims to complement GPU-centred AI infrastructure by combining a CPU and an AI inference neural processing unit, while cutting power use and costs.

It is also developing technology to improve computing resource utilisation. SKT is developing a next-generation AI data centre architecture with Panmnesia using Compute Express Link, or CXL, to flexibly connect resources such as CPUs, GPUs and memory. The approach uses other resources to make up for memory shortages on a specific server to increase GPU and memory utilisation.

For centre construction and cooling and power, it is highlighting global cooperation. It is working with Supermicro, Schneider Electric and others to strengthen efficiency across server and power and cooling infrastructure.

Jung Seok-geun (정석근), head of SK Hyper, said at the company's launch last month that it would secure key infrastructure and customers through systematic and fast execution. How quickly it can fill hyperscale centres with actual demand and operate them efficiently alongside building them is expected to determine the business outcome.

KT connects infrastructure through DCI.

KT is focusing on its nationwide telecom network and long-accumulated data centre operating technology. It plans to invest about 5 trillion won over the next 5 years to build AI data centres with 1 gigawatt of capacity. In the Seoul metropolitan area, it plans to supply centres mainly in areas where power and permits are secured. Outside the metropolitan area, it plans to build centres after confirming tenant demand.

KT's differentiator is connectivity. It says data centre interconnection, or DCI, technology that links multiple data centres through dedicated networks makes it easier to scale in line with computing demand. Using DCI allows it to expand infrastructure to another data centre when one centre lacks space or when setting up disaster recovery systems.

That can also be a strength in an AI data centre environment. Even if AI training and inference resources are distributed across multiple locations, they can be used like a single infrastructure through high-speed network connections.

KT Cloud is also advancing liquid cooling technology. It commercialised direct-to-chip liquid cooling in Korea for the first time, using coolant to cool heat generated by GPUs directly. Yoon-young Park (박윤영), KT's chief executive, recently visited the Mokdong 2 data centre to inspect key AI data centre operating technologies, including liquid cooling and DCI.

Park said at a press briefing last month that "even in an era of AI transformation where the target of connection expands from people to AI, the essence of KT, which is responsible for connectivity in the Republic of Korea, does not change." It is interpreted as meaning KT will seek differentiation in AI data centres by using its existing telecom network and operational capabilities.

LG Uplus strengthens cooling and power technologies with its "One LG" strategy.

LG Uplus is seeking differentiation through its "One LG" strategy, combining technologies from group affiliates. The key hub is the Paju AI data centre, which is under construction and targeted for completion in 2027.

The Paju AI data centre, being built with 200 megawatts of capacity, the largest in the Seoul metropolitan area, will apply a wide range of new group technologies. For cooling, it is working with LG Electronics. It will apply direct-to-chip liquid cooling that attaches a cold plate to GPU chips and circulates coolant directly. In power, it will use LG Energy Solution's high-performance uninterruptible power supply battery to respond to outages and voltage fluctuations and reduce risks of fire and thermal runaway.

LG Uplus will add an AI-based data centre infrastructure management system to its experience operating 15 data centres nationwide. The aim is to analyse power use, temperature and humidity, cooling status and equipment abnormalities in real time to check in advance for potential power shortages or overheating.

Hyung-gyun Ahn (안형균), head of LG Uplus' AI Business Group, said at the Paju AI data centre last month that it would "advance into an AI factory that operates and scales GPUs most stably." The strategy is to move beyond a centre that merely houses AI servers and instead integrate and optimise cooling, power and operations.