[Science Scope] The Hidden Side of AI: Consuming Both Electricity and Water

The Secret of Data Centers
To the Arctic, Under the Sea, in Abandoned Mines, and Beside Nuclear Plants
The AI Race Is About Power, Cooling, and Location—Not Algorithms
Even 'Space Data Centers' Are Emerging

The generative artificial intelligence (AI) seen on screens is often perceived as a digital technology that answers questions and creates images. For the general public, AI still feels like an invisible, intangible entity. However, the story changes when we step into the world of infrastructure. The data centers that power AI are highly physical industrial facilities. They consume vast amounts of electricity and generate considerable heat. This is why the technology race propelled by generative AI is expanding beyond semiconductors and algorithms, now encompassing competition for electricity, cooling solutions, and site locations.

A data center that withstood two years under the sea. Images show the 'Project Natick' data center, which Microsoft installed on the seabed of the Orkney Islands in Scotland, being raised after two years (left) and undergoing internal inspection (right). It drew attention as the world's first underwater data center experiment utilizing cold seawater as a natural coolant. Provided by Microsoft (Jonathan Banks).

Why Are Data Centers Heading North?

The International Energy Agency (IEA) has projected that global data center electricity consumption will reach around 945 terawatt-hours (TWh) by 2030. This figure is nearly equivalent to Japan’s total annual electricity consumption. The IEA analyzed that the spread of generative AI will be the core driver behind the surge in data center power demand. The recent trend of data centers moving toward the Arctic, submerging under the sea, and delving deep into abandoned mines vividly illustrates this shift.

The first places AI data centers seek out are ‘cold regions.’ The reason is the reduced cost and energy required for cooling servers. Meta's large-scale data center is located in Luleå, northern Sweden. Close to the Arctic Circle, this area uses cold outside air year-round to cool servers. Meta's choice to build its first major data center outside the United States here was influenced by the low temperatures and the abundance of renewable energy available.

A panoramic view of Meta's data center located in Luleå, northern Sweden. This region, close to the Arctic Circle, is considered a prime data center location for global big tech companies as it leverages low temperatures year-round to reduce power consumption for server cooling. Provided by Meta

The Google data center in Hamina, Finland, is also distinctive. Google repurposed the seawater tunnel of an old paper mill to cool its servers with cold Baltic Sea water. This is a case where an industrial-era facility has merged with the natural environment to transform into a core piece of infrastructure for the AI age.

In the past, data centers prioritized easy access to telecommunications networks and proximity to users when selecting locations. Minimizing network latency was the top criterion. However, the advent of the generative AI era is changing these location formulas.

Noh Sangmin, Data Center Director at Naver Cloud, explained, “For training infrastructure that repeatedly trains large-scale models, what matters most is a stable large-scale power supply, cooling efficiency, and site scalability, rather than real-time response speed. These training clusters are thus likely to follow favorable conditions for power and cooling.”

On the other hand, inference infrastructure—which responds to user requests in real time—requires low latency for optimal service quality. Chatbot or search services, for example, are still likely to favor access to communication networks and sites close to metropolitan areas. In short, not all AI data centers are looking for the same kinds of locations.

Beneath the Sea and in Abandoned Mines: Site Experiments Beyond Imagination

Even bolder experiments are underway. Microsoft pioneered ‘Project Natick,’ placing an entire data center beneath the sea. In 2018, Microsoft deployed a container-type underwater data center at a depth of 30 meters in the North Sea off the coast of Scotland and analyzed its performance after raising it two years later.

Underwater data centers can cut cooling costs by using cold seawater as a natural cooling source and are easily integrated with offshore wind and tidal power generation.

Conceptual diagram of the Shanghai underwater data center (left) and the actual platform built at sea (right). Electricity is generated on the sea surface in conjunction with offshore wind power, and the data center module installed about 35 meters below the water surface uses cold seawater as a natural cooling source to cool the servers, drawing attention as a new form. Provided by Shanghai Underwater Data Center Project and Hiwin Technology

Although challenges remain—such as corrosion, maintenance, and building undersea cables—experts say these projects have shattered the assumption that data centers must be located on land.

China has gone a step further. Recently, a commercial underwater data center went online near Shanghai, powered by a nearby offshore wind complex. About 2,000 servers installed 35 meters below sea level use seawater for cooling. This new type of infrastructure generates electricity atop the sea and cools heat using seawater below.

Dark abandoned mines are also being transformed into data centers. The ‘Lefdal Mine Datacenter’ in western Norway is a prime example of remodeling a former mine. It is resilient to external shocks and climate change, and it uses cold seawater from the nearby fjord as cooling water. Ready access to abundant hydropower is another advantage.

Spaces that once extracted coal and minerals have now been reborn as digital infrastructure for data storage and AI computation.

Inside the Lefdal Mine Datacenter in Norway. This facility converted abandoned mining tunnels into a data center, utilizing the mountain rock and the cold seawater from the nearby fjord as cooling sources. It is considered a prime example of a former mineral extraction site reborn as data infrastructure for the AI era. Provided by Lefdal Mine Datacenter

Park Jongbae, professor at the Department of Electrical and Electronic Engineering at Konkuk University, commented, “AI data centers are becoming a kind of ‘industrial nomad’ migrating in search of electricity. The reason that Arctic regions, coastlines, and abandoned mines are attracting attention is that they simultaneously address the two critical objectives of securing power and reducing cooling costs.”

Kim Sungjin, professor at KAIST’s Department of Mechanical Engineering, also noted, “Cooling systems account for around 30 percent of total power consumption in data centers. If we can reduce the energy required for cooling, it will significantly broaden the range of possible sites for data centers.” In other words, AI data centers now follow sources of electricity rather than communication networks.

"AI Consumes More than Power": The Water-Drinking Algorithms

One lesser-known fact is that AI uses a substantial amount of water. Evaporative coolers and cooling towers run to dissipate the heat produced by servers, consuming significant quantities of water in the process.

According to the Brookings Institution in the U.S., a typical data center uses about 300,000 gallons (roughly 1.13 million liters) of water per day, while ultra-large-scale facilities can consume up to 5 million gallons (around 18.93 million liters) daily. That’s enough to fill about 7.5 Olympic-size swimming pools.

The problem is that the heat generated by GPUs dedicated to AI is rapidly increasing, making it harder for traditional air-cooling methods to keep up.

Professor Kim explained, “New AI semiconductors generate hundreds of watts of heat per chip, and next-generation products are expected to exceed 1,000 watts. Air-blowing cooling methods are reaching their limits, so switching to liquid cooling is unavoidable.”

A representative from the SK Telecom AI Data Center Business Unit stated, “For AI data centers, not only is ultra-high-performance computing necessary, but also the extreme management of heat and a massive, balanced electricity supply. Racks with GPU servers consume ultra-high-density electricity, often tens of kilowatts or more, so next-generation cooling technologies are essential.”

The technology the industry is now focusing on is “Direct Liquid Cooling” (DLC), which directly delivers coolant close to the server’s internal chipsets to remove heat. DLC has emerged as a core technology in recent AI data center design.

Director Noh explained, “With the increased power density of the latest GPU servers, DLC is no longer optional but a mandatory infrastructure technology. This is not merely swapping out equipment—it's a transformation that changes the entire data center design, from power supply structure and coolant piping to server room layout, leak detection, and water quality management systems.”

Professor Kim added, “Recently, ‘dry cooling’ technology, which recirculates coolant or releases heat into the outside air, has also drawn attention. Improving cooling efficiency is critical because it can reduce both power consumption and water usage.”

Big Tech Moves Next to Nuclear Power Plants

Given these circumstances, global big tech firms are naturally turning their attention to nuclear power plants.

Three Mile Island Nuclear Power Plant in Pennsylvania, USA, being reactivated to supply power for Microsoft's AI data center. Photo by Getty Images

AI data centers must run 24/7. While solar and wind power are environmentally friendly, their output varies with the weather. In contrast, nuclear plants offer a stable, large-scale power supply unaffected by climate.

This is why companies such as Microsoft, Amazon, and Google are signing power purchase agreements with existing nuclear plants or investing in small modular reactor (SMR) companies. The International Atomic Energy Agency (IAEA) has also noted that the growing power demand of the AI industry and data centers is spurring renewed discussion of next-generation nuclear power.

Professor Park predicted, “Looking ahead, data centers may move away from simply buying power from utilities and instead move towards co-locating generation facilities with data centers. We may even see an era where the data center itself becomes a mini utility.”

Challenges for Korea’s AI Strategy

This global infrastructure race is posing new challenges for Korea as well.

In Korea, demand for new AI data centers is rising rapidly, but the capacity of the power grid and the availability of land in the Seoul metropolitan area are steadily worsening. When data centers cluster in Seoul and nearby areas, the load on the grid and transmission network inevitably escalates. As a result, the nation’s AI strategy is also expanding beyond just securing GPUs to include site selection, electricity supply, and securing sufficient cooling water for data centers.

Conceptual image of a space data center. Data centers are already heading to the Arctic Circle, underwater, and abandoned mines, while some companies are even considering establishing data storage and computing facilities in low Earth orbit. Photo by Conceptual Image of Space Data Center

Recently, Europe and the U.S. have begun discussions around the ‘in-orbit data center’ concept—building data storage and computation facilities in low Earth orbit. Although still at an early stage, it is a symbolic example of how the data center location race is expanding from the Arctic, sea beds, and abandoned mines all the way to outer space.

Even if the AI we encounter on our smartphone or PC screens appears to be merely a few lines of text, in reality, there is a vast infrastructure battle over electricity, water, land, and power stations unfolding behind the scenes. In the AI era, competitiveness is no longer decided solely by algorithms. Who can secure power, water, cooling, and sites most reliably will determine the winners in the next generation of AI competition.

This content was produced with the assistance of AI translation services.