Billions of liters of clean water are evaporating into the sky every year from giant data centers. It is not a machine immersed in water, but a harsh physical law that is silently operating, turning technology into a big trade-off problem.

Every time a large-scale data center project is announced, public opinion often focuses on sensational figures on water consumption. Many people imagine millions of gallons of water being pumped through state-of-the-art servers like engine coolant, then completely disappearing inside electronic chips.

In fact, that picture is completely misleading. The server was never exposed to water. All liquid flows at these facilities flow in closed pipe systems, and much of the water loss actually goes straight into the sky as vapor from rooftop cooling towers.

On chilly mornings, people can easily see columns of white smoke billowing above computer buildings, a phenomenon often mistaken for smoke from a fire.

The essence of the story is not where the water has gone, but why technology businesses actively release billions of liters of clean water into the air. The answer comes from a basic law of physics: latent heat of vaporization.

Currently, the data center system in the US alone consumes more than 4% of the total national electricity output. Almost all of this huge amount of electricity, after operating billions of transistors on silicon wafers, is turned into heat. If you don't quickly push that heat source out, the computer system will immediately be damaged.

To dissipate heat most effectively and cost-effectively, engineers are forced to take advantage of the evaporation process of water. According to thermodynamic calculations, heating one kilogram of water by 10 degrees Celsius absorbs only about 42 kilojoules of energy. However, if one kilogram of that water is converted into a gas at a temperature of 30 degrees Celsius, it will remove up to 2,430 kilojoules of heat - about 58 times more.

With each kilowatt-hour of electricity equivalent to 3,600 kilojoules of heat emitted, the cooling tower only needs to evaporate about 1.5 liters of water to release all this heat. If you choose the method of heating water and then draining it, the amount of water needed will increase nearly sixty times.

In actual operation, cooling towers are mechanical equipment blocks placed on the roof of the building, spraying hot water from the server room through honeycomb plastic membranes for fans to suck air through. About 80% of the water taken in will evaporate to cool the remaining water, the remaining 20% ​​must be periodically drained to remove accumulated mineral salts, avoiding scale buildup and pipe blockage.

In 2023, data centers in the US will evaporate about 66 billion liters of water on-site. However, the actual level of waste in each region is very different. Some large cloud operators average 1 liter per kilowatt-hour by taking advantage of solar wind, but in extreme climates like Arizona summers, this number can jump to 9 liters per kilowatt-hour.

Many people ask why not use outside air directly for cooling. The biggest barrier lies in the physical limit of wet bulb temperature - the lowest temperature achievable by evaporation under a given humidity condition.

When air humidity increases or the ambient temperature exceeds 29 degrees Celsius, the dry heatsinks completely lose their effectiveness, forcing the evaporative cooling system to intervene.

Another key aspect that is often overlooked in environmental debates is the indirect water footprint. Evaporation at server facilities is actually just the tip of the iceberg.

Most of the electricity supplied to digital infrastructure comes from thermal power plants - where water is continuously boiled into steam to drive turbines and then evaporated through giant condensing towers. In 2023, the amount of water indirectly evaporated from power plants supplying data centers in the US will reach nearly 800 billion liters, which is 12 times higher than the amount of water consumed directly on site.

Even the story about artificial intelligence models like ChatGPT consuming resources clearly reflects this mechanism. Studies estimate the GPT-3 model requires about 500 milliliters of water for a typical 10 to 50 responses.

Notably, only about 2 milliliters of water evaporated at the algorithm processing server, while up to 15 milliliters of water evaporated at the power plant. The thirst for artificial intelligence, after all, is the shadow behind which power consumption lurks.

Under pressure from the community, technology corporations are finding ways to adjust but cannot eliminate the trade-offs. Microsoft commits that its new generation of data centers operating from the end of 2027 will completely eliminate the use of cooling water by using a closed liquid cycle, but this means that power consumption must increase.

Other directions such as switching to using treated wastewater, industrial water, sea water, or relocating server infrastructure to cold countries like Finland are just ways to shift costs to power sources or transmission infrastructure problems. The white column of steam on the roof of data buildings is, after all, the only visible part of a very high price that the digital economy is paying.