Original title: Will AI lead to power shortage

"AI consumes far more power than the data center" "AI development will collapse the existing power grid" "Technology companies will build nuclear power plants"... At several international conferences held recently, some technology giants have expressed their concern about AI energy consumption. Faced with the potential impact of AI development on the power system, the National Energy Administration will respond publicly, closely follow the trend of AI and other high-tech development, analyze and judge the growth trend of power demand, systematically plan measures to ensure the stable supply of energy and meet the demand growth.

Will AI really lead to power shortage? At present, all walks of life are trying to get a ride on AI. However, the development and operation of AI are inseparable from energy security. Taking the popular AI image generation function on the Internet as an example, each image generated consumes the same amount of power as a smartphone is fully charged. Nvidia H100 GPU chip is one of the most commonly used chips for artificial intelligence algorithm at present. One such chip consumes more power in a year than a medium-sized family. The energy consumption of large-scale deployment of AI is even more surprising. It is estimated that from 2027, only the new servers of AI will consume the same amount of power as the total power consumption of countries like the Netherlands or Sweden.

The latest warning comes from Elon Musk, who openly pointed out that the computing power of AI consumption seems to increase 10 times every six months. More than a year ago, chip was in short supply, and then power will be in short supply. "Soon people will see that there is not enough power to run all AI chips". In addition to the technology giants, the rise of AI has also caused more discussion and concern. Everyone is worried that AI will cause the unemployment rate to soar. To this end, some netizens jokingly said that "as long as the electricity cost is more expensive than steamed bread, AI can never completely replace people".

Why does AI consume so much power? In people's impression, traditional industries such as steel and cement are big power consumers, and it is difficult to link technology companies with high energy consumption. In fact, AI industry is also an energy intensive industry. Today's electronic computer industry is based on integrated circuits, which are often called chips. There are many transistors in each chip, which are used to undertake the "switch" function. The computer toggles these switches by rapidly changing the voltage to achieve binary logic operations. With the progress of technology, the number of transistors on a unit chip is increasing, and even hundreds of millions of transistors can be integrated per square millimeter area. No matter how small the power consumption is, the power consumption is also very considerable.

The training of large models requires repeated iterations on large-scale data sets. Each iteration requires the calculation and adjustment of billions, tens of billions or even hundreds of billions of parameter values. These calculations will ultimately show the on-off of transistors, and each work requires power to drive. For neural networks, the larger the scale, the better the calculation results, the more problems can be solved, but the greater the power consumption. More importantly, this kind of calculation is not once and for all. It needs to constantly re train and re optimize the model, which also means that there seems to be no "ceiling" for AI iteration's demand for computing power and power.

The premise of AI development is that it cannot significantly increase carbon emissions and endanger energy security. Although the overall power consumption scale of AI is relatively low at present, which is not enough to threaten the power system, in response to the rapidly growing demand for power, it is necessary to reconsider the investment in power infrastructure and industrial layout to achieve sustainable development.

At the power supply side, it is necessary to coordinate the power and power coordination layout. China has built the world's largest power supply system and clean power generation system, and the scale of hydropower, wind power and photovoltaic power generation has ranked first in the world for many years. In the vast western and northern regions, renewable energy and land resources are abundant, and with the continuous improvement of new energy installation, there are challenges in the future power consumption; The eastern region is affected by land, water and electricity, operation and maintenance and other factors, so the operation cost of the data center is high. Therefore, it can guide the data centers with low delay requirements to gather in the western region and realize the differential processing of data, which is not only conducive to green power consumption and economic development in the western region, but also can promote the low-carbon transformation of data centers, ensure energy security, and truly transform China's green power advantages into computing advantages.

On the power side, we should accelerate the development of low-carbon and energy-saving technologies. The current operation mode of AI is very energy consuming, so we should actively seek more energy-efficient algorithms and electronic components to improve the availability of data centers. At the same time, by improving the design of the data center, optimizing the cooling system, and setting up waste heat recovery equipment, the energy efficiency level of the data center is effectively improved and energy waste is reduced.

In the future, artificial intelligence will be deeply integrated into our production and life. Although energy consumption is worth paying attention to, there is no need to stop eating because of choking. It is believed that with technological progress, market development and optimization of productivity layout, the industry will find appropriate solutions to balance computing power and energy consumption demand and achieve high-quality development. (Wang Yichen)