Cooling technology emerges as key to overcoming AI infrastructure constraints

As Big Tech companies invested approximately $320 billion into US AI infrastructure in 2025, cooling has become a vital method to boost computing power amid shortages of land, chips, and electricity. With next-generation AI racks expected to reach 1 megawatt of heat, the global cooling market is projected to grow from $18.8 billion in 2025 to $54.4 billion by 2034.

The essence of the AI competition is ultimately a battle over "who can calculate more." To this end, Big Tech companies are investing massive amounts of capital to build data centers, secure GPUs, and draw in electricity. In 2025 alone, Big Tech companies in the United States poured approximately $320 billion (about 426 trillion won) into AI infrastructure. However, increasing land, chips, and electricity immediately is nearly impossible. Physical limitations are clear, as projects are being derailed due to data center location issues or because building power grids takes several years.

According to this analysis reported by T Times TV, the almost unique technical breakthrough to increase computing power under these constraints is "cooling." Improving cooling efficiency offers benefits in three areas. First, power previously used for cooling can be redirected to computation, improving the Power Usage Effectiveness (PUE). Second, better heat dissipation allows GPUs to be placed more densely within the same area, increasing land efficiency. Third, it can maximize chip performance and lifespan by preventing "throttling," a phenomenon where chips reduce performance due to overheating.

AI racks must withstand the heat of 600 hair dryers

The heat generation problem in AI data centers is beyond imagination. NVIDIA's latest GPU, Blackwell (B200), consumes 1,400 watts per chip, and the upcoming Rubin chip is expected to use more than 1,800 watts. This means a single palm-sized chip emits as much heat as a household hair dryer running at maximum capacity. The problem is the heat at the "rack" level, where hundreds of such chips are installed. The heat emitted by a single rack of a next-generation AI platform is expected to reach 1 megawatt (MW), which is equivalent to running 600 hair dryers simultaneously.

The most widely used method currently is "air cooling." However, air has a very low capacity to carry heat. To cool more heat, fans must be run harder, which consumes massive amounts of additional power. This creates an inefficiency similar to waiting to get cool while running hundreds of electric fans inside a sauna in the middle of summer. Consequently, analysis suggests that air alone has reached a threshold where it cannot handle the explosively increasing heat of AI chips.

Immersion cooling and NVIDIA's struggle for dominance

To overcome the limits of air, cooling technologies using liquids are gaining attention. Water can absorb about 3,500 times more heat than air. Technologies are largely divided into three types: Rear Door Heat Exchangers (RDHx), which attach a coil plate with cold water flowing behind the server; Direct-to-Chip (D2C) cooling, which attaches metal cooling plates onto GPUs to flow coolant; and immersion cooling, which submerges the entire server in a special non-conductive liquid.

In particular, immersion cooling is a powerful technology that can lower PUE to the 1.05 level. However, while NVIDIA currently officially supports the D2C method for its GPU servers, it has not yet certified immersion cooling. The prevailing view is that this is a matter of "dominance" rather than technical capability. In D2C, cooling companies essentially install air conditioners on servers designed by NVIDIA, allowing NVIDIA to maintain design authority. In contrast, immersion cooling requires the entire server to be redesigned to match the cooling method, creating a risk that cooling companies could seize dominance in the server manufacturing market. However, as power density per rack approaches 1MW, it is expected that NVIDIA will find it difficult to ignore immersion cooling.

From oil refiners to power equipment makers, the battle for supremacy in the cooling market

The cooling market is evolving beyond a mere accessory for data centers into an independent growth industry. The global data center cooling market is expected to grow nearly threefold, from approximately $18.8 billion in 2025 to $54.4 billion by 2034. Key players in this market include Vertiv, which leads reference designs through close cooperation with NVIDIA; Schneider Electric, which locks in customers with software platforms; and Johnson Controls, which possesses large-scale refrigeration technology.

Interestingly, the shift in cooling technology is breaking down industrial boundaries. The entities most advantaged in producing the "special non-conductive cooling oil" essential for immersion cooling are oil refiners. Accordingly, global companies such as Shell and ExxonMobil, as well as domestic refiners like SK Enmove and GS Caltex, are emerging as key players in the AI infrastructure market. With the addition of power equipment company Eaton and water treatment company Ecolab, the hegemony of AI infrastructure is shifting beyond securing GPUs to "who can manage heat more efficiently."

Source: original video (YouTube)