A supercamings of artificial intelligence will require up to 4 nuclear power plants

A new study shows that if the current trend maintains, the construction of a computer supercomputers will cost $ 5 billion and 2 GW of electricity in the year, which is approximately equivalent to the output of two nuclear power plants.

Why is electricity consumption and the cost of building a supercomputer artificial intelligence?

Increasing the popularity of the use of artificial intelligence -based services, including chats among users, has made the processing power in this area. A new study by researchers at the University of George Town shows that increasing demand in this area by the end of this decade will increase the cost of developing and maintaining special AI data centers. In the interval between years 1 and 2, the cost has witnessed a 1 -fold mutation, and if the current trend is maintained, we should expect $ 2 billion in datacenters by year 2, each of which will require 2 GW of energy.

The most important factor in raising the cost of developing a computer supercomputer, or data center, is the high price of chips and a lack of production. Researchers estimate that by year 3, every datacenter in the AI ​​will require 5 million chips, which will cost the technology giants to provide. It should not be forgotten, however, that the electricity consumption of the chips is also a serious challenge; So that the power output of several nuclear power plants may be devoted to the activity of an artificial intelligence computer.

Of course, optimizing chips and improving processing power for power consumption is another important factor that can reduce the cost of maintaining future data centers. Current statistics show that we have seen an increase in processing power against any watts consumed in datacenters from year 1 to 5, and we may see improvement in the near future with future improvements. Currently, the Colossus XAI CoSAI is the most powerful of its kind, which has been developed over a $ 2 billion cost of $ 5 billion and receiving 6,000 chips with a power consumption of 2 megawatts. For comparison, it should be noted that this amount of energy is enough to supply 6,000 home subscribers.

The United States has the highest dominance of dedicated data centers for AI with a 5 % share, and China stands in second with a 5 % share. Other advanced industrialized countries, such as Japan and Germany, are far less in the next rank. Of course, the location of the datacenters has nothing to do with the users, and customers can rent the processing power they need from anywhere in the world. In the coming years, the share of countries with numerous nuclear power plants and cheap electricity such as France will increase in the field, and we will see more distribution of artificial intelligence datacenters around the world.