The GPT-3 consumes 1287 megawatt hours of power during a single training session, which can support 3000 Tesla electric vehicles to run together, with each vehicle running 200000 miles. If all the electrical energy required for GPT-4 training is converted into heat energy, it can heat approximately 1000 Olympic standard swimming pools to boiling. A small AI, but an energy consuming giant. With the popularization and application of technology, AI is accelerating both technological progress and global energy consumption. The problem of huge energy consumption in AI is severe but not unsolvable. In order to solve this problem, countries and the technology industry are also actively exploring corresponding countermeasures. From a global perspective, the energy consumption of AI is constantly increasing with the development of technology and demand. In the United States, Boston Consulting Group predicts that by 2030, the electricity consumption of data centers in the United States will triple compared to 2022, and this increase will mainly come from artificial intelligence. In Ireland, the power consumption of data centers has exceeded the total household electricity consumption of all cities in the country, accounting for approximately 21%. In China, according to relevant institutions' predictions, the electricity consumption of data centers in China will exceed 950 billion kilowatt hours by 2030, which is more than 3.5 times that of 2022. Nowadays, the power consumption of global data centers has increased from 10 billion watts (GW) ten years ago to the level of 100 billion watts. It is obvious that the energy consumption of AI is increasing day by day, and the reason lies in the operation of AI system devices. In terms of hardware configuration, high-performance hardware is one of the reasons why AI consumes a lot of energy. AI systems, especially deep learning models, require a large number of computationally intensive tasks such as matrix and floating-point operations, which cannot be separated from the support of high-performance computers and graphics processing units (GPUs). Especially in the process of training large models, multiple GPUs need to run continuously, and the energy consumption of one GPU is much higher than that of a central processing unit (CPU). For example, a Nvidia A100 GPU has a power consumption of 400 watts, GPT-3 training uses 1024 A100 chips, and GPT-4 has climbed to 25000 chips. The increase in model size and quantity has led to a significant increase in energy consumption. In terms of computing operation, the large-scale data processing and extremely complex algorithm models further exacerbate AI's energy consumption. AI systems typically need to process large amounts of data to extract useful features and information. Training a large language model may require analyzing and processing billions or even trillions of textual data. Both training data and real-time input data require a large amount of energy support when the data scale is large. With the increasing complexity of AI models, the required computing resources and internal resources are constantly increasing, and the energy consumption is also increasing. In terms of equipment maintenance, effective cooling systems and equipment upgrades also consume a significant amount of energy. High performance computing hardware generates a large amount of heat during the operation of applications. In order to ensure the normal operation and prolong the service life of the equipment, a cooling system is needed to maintain the stable operation of the hardware, and the operation of the cooling system also consumes energy. Microsoft's data centers in the United States directly consumed 700000 liters of clean fresh water to cool the system, solely for training GPT-3 models. And the update and replacement speed of AI devices is fast. While the elimination and disposal of old devices bring certain environmental and energy burdens, the manufacturing and deployment of new devices also consume a large amount of energy and resources. Moreover, the development of AI technology is accompanied by significant carbon emissions and freshwater consumption. A research team from Harvard University analyzed the operations of 2132 data centers in the United States and found that carbon emissions from data centers have tripled since 2018. As of August 2024, these facilities emitted a total of 105 million tons of carbon dioxide, accounting for 2.18% of the total carbon emissions in the United States. The energy consumption problem is expected to be solved. In response to the huge energy consumption of AI, some large technology companies are considering purchasing nuclear power plants to satisfy AI's appetite. According to the Wall Street Journal, Google is considering signing a power purchase agreement with developers of small modular reactors (SMRs). Developing clean energy such as nuclear energy, solar energy, and wind energy can alleviate the tight demand for AI energy to some extent, but the core lies in strengthening the innovation and optimization of AI technology, government policy guidance and regulation, and addressing the problem of huge AI energy consumption by addressing both the symptoms and root causes. In terms of technological innovation and optimization, the industry continuously improves AI models and algorithms, updates the use of high-performance hardware, and optimizes data center design. By optimizing AI models and algorithms, developing more efficient and intelligent technologies, reducing unnecessary computation, improving model performance, and reducing energy consumption. According to foreign media reports, the novel integer addition algorithm proposed by BitEnergy AI can significantly reduce energy consumption while ensuring the calculation results. By utilizing advanced semiconductor manufacturing technology to customize AI chips, it has higher energy efficiency compared to traditional CPUs and GPUs. For example, Nvidia's Tensor Core and Google's TPU have designed chips specifically for AI computing, which provide efficient matrix multiplication and convolution operations, thereby reducing energy consumption; Explore the use of new chips, such as quantum chips, to disrupt traditional computing models and achieve higher energy efficiency; Adopting efficient cooling technologies such as liquid cooling to reduce the energy consumption of data centers for heat dissipation. In terms of government policy guidance and regulation, both domestic and international efforts are actively promoting data sharing and governance, as well as the development of energy-saving standards and regulations. The EU is promoting the green transformation of the AI field through policies such as the European Green Deal, and has established a special fund to support the research and application of AI energy-saving technologies. The Japanese government promotes the improvement of energy efficiency of AI products through the "Top Runner" system. The US government also encourages companies to use renewable energy and energy-saving technologies through tax breaks, subsidies, and other means, and requires data centers to have an energy efficiency ratio no higher than a certain value from 2025, and to increase the proportion of renewable energy use year by year. China has also launched a series of measures in the green and low-carbon development of the AI industry. Computing power is an important pillar of artificial intelligence. In 2021, China launched the "East West Computing" project to promote the construction of green data centers, reasonably allocate computing power resources, and deploy computing power resources to the western regions with lower energy costs and cooler climates to reduce energy consumption. In July 2024, the National Development and Reform Commission and the Ministry of Industry and Information Technology released the "Special Action Plan for Green and Low Carbon Development of Data Centers", aiming to promote the green and low-carbon development of data centers, accelerate energy-saving and carbon reduction renovations, and update energy consuming equipment. The plan proposes specific goals by the end of 2025, including a more rational layout of data centers nationwide, an overall shelving rate of no less than 60%, an average electricity utilization efficiency of less than 1.5, and an average annual growth rate of 10% in renewable energy utilization. China's basic telecommunications enterprises are also making efforts in AI and strengthening the construction of green and low-carbon energy. China Telecom has launched a green cloud computer using zero carbon cloud as its computing power carrier; Build a high standard AIDC demonstration base for "Two Bombs and One Optimization" to effectively reduce the PUE value of data centers. China Mobile is comprehensively promoting the "AI+" action plan, accelerating the evolution of network infrastructure towards the integration of space, sky, and earth, and creating a diverse, ubiquitous, green, and low-carbon intelligent computing cluster. China Unicom is actively building its "dual carbon" cloud platform and promoting new technologies such as evaporative cooling, fresh air, and graphene in Hebei, Xinjiang, Ningxia, and other regions to improve the level of data center greening. From the steam age to the electrical age, and then to the information age, technology has changed the way society produces and lives. With the promotion and application of technology, AI is gradually becoming an indispensable part of us. Blockage is better than diversion, and diversion is better than diversion. Perhaps the energy crisis brought by AI has not been solved yet, but it is not an unsolvable problem. I believe that in the future, with the iterative optimization of intelligent technology, policy guidance, and the development and utilization of new energy, the development and utilization of AI will be more fully utilized, and the problems caused by AI's consumption of energy such as electricity will be solved. (New Society)