Inside the relentless race for AI capacity

These screws threaten successive power outages, affecting homes and companies that feed the network itself. Last summer, Virginia service providers had to wrestle with a sudden increase in power after a group of facilities turned into reserve generators as a protective measure of safety, which displays the supply of network infrastructure.

With abundant strength, operators also ended in areas with large water restrictions. Consuming lobes and group sites in the United States consumes 55 billion liters of water in 2023, according to researchers in LBNL. The indirect consumption associated with the use of energy is significantly higher at 800 billion liters per year, which is the annual equivalent of water of about $ 2 million. American homes.

In 2023, Microsoft said 42 percent of its water came from “areas with water stress”, while Meta said nearly 16 percent of its water use was derived from similar areas during the same time period. Google said last year that about 30 percent of its water came from the water gatherings with an average or large risk of “draining water or scarcity.” Amazon does not reveal Its number.

Data centers in drought -vulnerable states such as Arizona and Texas led to anxiety among the local population, while the inhabitants of Georgia complained that the development of dead in the state had affected water wells, pushed the cost of municipal water and led to a shortage of that. The water was legalized.

Some believe that this endless race for medium computing power is misplaced.

Sasha Lakkoni, Amnesty International, Climate Guns in the face of open source embrace, said that alternative techniques for training artificial intelligence models, such as distillation or the use of smaller models, were gaining popularity and could allow developers to build strong models in a small part of Cost.

“It is almost similar to the mass hallucinations, as everyone is at the same wavelength that we need with more data centers without actually wondering about the reason.” She said.

Increasing the density of chips has another unwanted effect: heat. About five power used by the artificial intelligence data center stems from cooling and equipment chips, according to Consultants McKinsey.

Early data centers operate cloud work burdens that publish industrial air conditioning units similar to those used in offices to cool servers. But when the chips began to draw more energy, it became difficult to keep them within the safe operating range between 30 and 40 degrees Celsius, with data centers that require more advanced cooling methods to avoid malfunctions. The challenge has led to significant investments in advanced innovations.

The operators have turned to install the pipes filled with cold water in the server room to transfer heat away from the equipment. Then this water is directed to the large cooling towers that use evaporation to reduce the temperature of the facility to a safe range. But the approach leads to water loss, with a single tower that passes through about 19,000 liters In the minute.

Microsoft and others adopted a closed loop system based on coolers-in reality, refrigerator-to cool water. This process is less wasteful and more efficient than Evacuation options.