돈經濟

In 2026, the AI infrastructure market will no longer be decided solely by chips and servers. The real key has become how to reliably draw power and how to dissipate heat. The situation is characterized by a split into super-large facilities for training and distributed networks for inference, traditional infrastructure companies fixing profits with long-term contracts, and cooling components and skilled labor emerging as bottlenecks.

Training and inference are physically separated

The most noticeable change is that infrastructure is completely divided according to the nature of the AI task. Training requires so much power that it moves to isolated, massive facilities near natural gas power plants or SMRs to avoid existing power grid queues. Conversely, inference, where user response speed is critical, is fragmented into small edge networks that utilize communication networks and surplus power in urban or small and medium-sized cities.

In Korea, this trend is already becoming a reality. GS is promoting gigawatt-class data centers near its Donghae and Dangjin power plants. This is a strategy to avoid transmission network bottlenecks by directly connecting to self-generated power. SK Group is preparing regional facilities, centered in Ulsan, utilizing LNG and LPG power generation from SK Gas and SK Multi Utility. Doosan Enerbility is investing hundreds of billions in a dedicated SMR plant to enhance its long-term power supply capabilities. In construction, SK ecoplant, Hyundai E&C, Samsung C&T, and Daewoo E&C are actually building these facilities. For inference, LG Uplus's Paju Center and KT Cloud's commercialization of liquid cooling are representative examples.

From an investor's perspective, the power and construction value chain directly connected to power plants is more likely to be a stable growth axis than data center operators. Training facilities are large-scale and have long contract periods, so once secured, their performance lasts for a long time.

Traditional infrastructure companies are transforming like software

The second change is the business model itself. In the past, it was about delivering equipment or completing a one-time construction project. Now, they are entering into ultra-long-term exclusive contracts with big tech companies for 10 to 15 years. As revenue transforms into predictable, recurring revenue, the method of valuing companies also changes.

HD Hyundai Electric, Hyosung Heavy Industries, and LS ELECTRIC are the companies that most clearly demonstrate this trend in Korea. They have successively signed long-term supply contracts worth trillions of won with North American big tech companies, accumulating tens of trillions of won in order backlogs. In Korea, if the government's 18.4GW AI data center mega-project is added, mid-to-long-term performance visibility will further increase. LS ELECTRIC is also securing next-generation technology by jointly demonstrating 800V DC power infrastructure with LG Uplus.

Investors should now focus more on the proportion of long-term contracts and the speed of order backlog growth of these companies rather than their short-term performance. Key points to check are whether margins are stably maintained and whether production capacity can actually keep up. The re-evaluation of valuations, similar to software companies, has already begun.

Contradiction between eco-friendly goals and reality

The third is the discrepancy between eco-friendly goals and reality. Big tech companies advocate for net-zero, but they are immediately increasing off-grid natural gas power generation due to a lack of electricity. At the same time, to avoid concerns about water depletion, they are adopting closed-loop cooling and two-phase direct liquid cooling, which use almost no water, as essential specifications.

In Korea, GS, SK Gas, and SK Multi Utility are strengthening their self-generation-based data center models. In terms of cooling, GS Caltex and SK Enmove supply immersion cooling fluids, and LG Electronics manufactures D2C cold plates and cooling distribution units. GST and KNSOL are entering the market by developing immersion cooling systems. KT Cloud and LG Uplus have already commercialized a hybrid method that combines air and liquid.

As deregulation and technology demonstration are progressing rapidly in this area, it is necessary to carefully observe the commercialization speed of related companies and their entry into the global supply chain. If waterless cooling technology becomes the standard, the bargaining power of these companies will significantly increase.

The real bottleneck is cooling components and skilled labor.

Fourth, and most realistically, the bottleneck is not the chip. It's components like special chemical fluids to physically cool ultra-high-density AI chips, flawless magnetic pumps, and aerospace-grade quick connector valves. A few companies globally monopolize these, and there's an even greater shortage of skilled personnel who can perfectly install ultra-high-pressure electrical facilities and precision piping.

In Korea, GST is standing out by developing 2-phase immersion cooling technology. Hyosung Goodsprings supplies sensorless pumps optimized for data centers, and LG Electronics is expanding its CDU solutions. The capabilities of construction companies' specialized construction teams are also becoming important, as supply delays directly push back data center operation schedules.

From an investor's perspective, companies that can resolve bottlenecks are more likely to command a premium than equipment suppliers. Those with cooling fluids, special components, and skilled construction capabilities will have a long-term advantage.

Opportunity Cost of Mining Farms Shifting to AI

Finally, the rapid conversion of cryptocurrency mining farms in the US is also worth noting. Gigawatt-scale sites connected to the power grid are starting to generate much higher profits from AI hosting than from Bitcoin mining, causing mining rigs to shut down en masse and pivot to AI clouds. While Korea is not a major mining hub, the same logic applies in terms of the reversal of opportunity cost. The moves by GS, SK, and SGC Energy to convert existing power-intensive sites or power generation facilities into AI data centers are prime examples.

Points investors should pay attention to

In summary, the core of AI infrastructure investment in 2026 will not be chips, but rather the ability to stably supply power, efficiently dissipate heat, and secure profits through long-term contracts. Domestically, the three power equipment companies, HD Hyundai Electric, Hyosung Heavy Industries, and LS ELECTRIC, appear to be the most visible beneficiaries in the short term. GST, GS Caltex, SK Enmove, and LG Electronics, which are part of the cooling value chain, and SK ecoplant, GS, and Doosan Enerbility, which are EPC companies linked to power generation, are highly likely to emerge as mid-to-long-term growth drivers. Of course, execution risks such as delays in power grid permits, a shortage of skilled personnel, and global component supply bottlenecks still remain. Therefore, it is reasonable to review portfolios focusing on the scale of long-term contracts, the commercialization progress of cooling and power solutions, and linkage with government mega-projects, rather than the short-term performance of individual companies. As this trend accelerates, the re-evaluation of traditional infrastructure companies' valuations is likely to become more pronounced. Now is a time when the ability to discern structural changes is more important than mere numbers.