Typesetting and Proofreading | Gan Huiqi

During the Spring Festival, the hottest topic of conversation was undoubtedly DeepSeek.

During the Lunar New Year, I posed a question to DeepSeek: Compared to GPT-4, how much computational power can you save?

However, DeepSeek did not provide a specific answer; instead, it directed me to the company's product description.

Since the surge of generative AI in 2022, the relationship between AI, large model development, and electricity consumption has been a hot topic. The AI industry generally views energy as a bottleneck for the sector.

A paper published in the journal Joule predicts that by 2027, the global AI industry's annual electricity consumption will reach 85.4 to 134 terawatt-hours (TWH), roughly equivalent to the total annual electricity consumption of countries like the Netherlands, Sweden, or Argentina, accounting for about 0.5% of global total electricity consumption. The study also found that the popular generative AI ChatGPT processes about 200 million requests per day, consuming over 500,000 kWh of electricity in the process, equivalent to the daily electricity consumption of 17,000 American households.

At the 2024 Bosch Connected World Conference, Elon Musk warned that the rapid growth of AI and electric vehicles could lead to global shortages of electricity and transformer supplies. Meanwhile, Sam Altman, CEO of OpenAI, stated at the 2024 Davos World Economic Forum that the AI industry is facing an energy crisis, with the energy demands of the new generation of generative AI clearly exceeding expectations, and existing energy supplies struggling to cope. For a time, theories about the AI energy crisis flew around the world.

However, with the emergence of DeepSeek, this situation may change. Through more efficient computing, DeepSeek's computational demand is significantly reduced compared to OpenAI, and the "distributed" deployment of computing power, where everyone has a set of DeepSeek, becomes one of the options, making AI potentially no longer an electricity "behemoth".

Nonetheless, currently, there is no research comparing DeepSeek's energy use relative to its competitors, suggesting that energy remains the largest variable affecting AI development.

01

Recently, an article published on the MIT Technology Review website mentioned that in tests of 40 prompts, DeepSeek was found to have similar energy efficiency to Meta models, but DeepSeek tended to generate longer answers, thus using 87% more energy.

An industry insider told me that last year, there was a huge investment in AI computational power, with many companies building data centers in regions with cheaper electricity, such as Gansu, Qinghai, Tibet, Inner Mongolia, and others. Although they didn't anticipate DeepSeek emerging and changing the underlying logic, in fact, DeepSeek's current consumption of computational power is not low. It's not just the cards, it's the electricity, and electricity prices affect its overall cost.

Currently, large AI models are developing rapidly, capable of processing and learning vast datasets. This capability requires enormous computational resources, often provided by high-performance processors such as GPUs, TPUs, and ASIC chips. These processors require significant amounts of electricity to power data center servers, storage devices, and cooling systems. Generative pre-trained large language models, represented by GPT, have computational demands that depend on the number of parameters and tokens, greatly affecting the potential scale and complexity of the models. The exponential increase in the number of parameters, from billions to trillions, has only taken 3 years.

Since 2012, the power demand for AI training applications has doubled every 3 to 4 months. Training large models requires a lot of energy, and AI servers typically require higher power density hardware compared to traditional servers. For example, AI servers may require 4 high-power power supplies of 1800W, while general-purpose servers may only require 2 power supplies of 800W. Some institutions estimate that compared to traditional servers, AI servers have nearly 6-8 times higher power and a synchronous increase in power supply demand of 6-8 times.

According to research by Zhang Yaxin and Mu Qijian from the Resources and Environment Business Department of China International Engineering Consulting Corporation, by 2030, the power consumption for computational power in the field of large AI models alone is expected to reach 1.93% to 5.25% of the country's total electricity consumption.

Under a conservative growth scenario, it is estimated that AI computational power will consume about 217.7 billion kWh of electricity, accounting for 1.93% of the country's total electricity consumption, equivalent to about 26.75 million tons of standard coal, accounting for 0.45% of the country's total energy consumption. If predicted under a rapid growth scenario, the proportion of electricity consumption and energy consumption is expected to rise to 5.25% and 1.21%, respectively.

Moreover, as the AI industry continues to develop, more complex and diverse AI models and application scenarios may emerge in the future, which also means that the impact of energy consumption will show a continuous growth trend to a certain extent.

More importantly, the temporal and spatial distribution of future AI energy consumption impacts may become unbalanced.

From a temporal perspective, AI energy consumption exhibits seasonal fluctuations, being "high in summer and winter, low in spring and autumn." Seasonal temperature changes are a key factor affecting the energy consumption of computational infrastructure. For example, rising temperatures in summer require data centers to rely on a large number of cooling equipment and electricity to maintain ambient temperatures for normal equipment operation, greatly increasing energy consumption. In severe winter cold, data centers need additional energy to maintain a suitable indoor temperature to prevent equipment performance degradation due to low temperatures.

From a spatial perspective, China's "East Data, West Compute" project is reshaping the geographical layout of AI energy consumption. Overall, China is establishing eight major computing power hub nodes in Beijing-Tianjin-Hebei, the Yangtze River Delta, Guangdong-Hong Kong-Macao Greater Bay Area, Chengdu-Chongqing, Inner Mongolia, Guizhou, Gansu, and Ningxia, and planning ten national data center clusters based on the hub nodes, including Zhangjiakou Cluster, Wuhu Cluster, Yangtze River Delta Ecological Green Integrated Development Demonstration Zone, Shaoguan Cluster, Tianfu Cluster, Chongqing Cluster, Gui'an Cluster, Qingyang Cluster, and Hohhot Cluster. With the construction and development of computing power infrastructure hub nodes and data center clusters, the energy consumption pressure of AI will mainly shift to these concentrated areas.

In the short term, the rapid development of AI technology may bring significant seasonal and localized energy consumption pressure.

On the one hand, due to seasonal temperature changes, such as high summer temperatures and severe winter cold, computational infrastructure requires additional energy consumption to ensure its stable operation. In situations where energy supply and demand are already tight in summer and winter, a surge in energy demand for computational infrastructure may further exacerbate the pressure on energy supply and demand.

On the other hand, with the advancement of the "East Data, West Compute" project's integrated computing network, the impact of AI energy consumption is mainly concentrated in eastern regions such as Beijing-Tianjin-Hebei and the Yangtze River Delta, as well as western provinces such as Gansu and Inner Mongolia. Considering the differences in energy supply and demand between eastern and western regions, this spatial distribution may lead to a situation where energy supply is "tight in the east and surplus in the west," which not only affects energy efficiency but may also become a constraint on the coordinated development of the regional economy.

In recent years, China's total energy consumption has continued to grow rapidly, with the rate of decline in energy intensity narrowing, and the decline in energy intensity in some regions falling short of expectations, making it difficult and challenging to achieve energy-saving targets in the latter half of the 14th Five-Year Plan and to carry out energy-saving work during the 15th Five-Year Plan.

The seasonal and localized energy consumption pressure caused by AI in the short term may conflict with the current severe energy-saving situation. Some regions may be limited by the space for completing energy-saving targets or may strengthen the approval management of data center projects during the energy and environmental impact assessment stages due to energy supply and demand considerations, or even take restrictive measures. For example, local governments mostly use the power usage effectiveness (PUE) of data centers as the main regulatory focus in the energy assessment of data center projects and continuously tighten relevant standards. Currently, Beijing, Shanghai, and other places have raised the approval requirements for the PUE of newly built data centers to 1.15, far stricter than the average level in European and American countries.

02

So, where is the solution to the AI energy consumption problem?

The primary solution to the problem is, of course, to optimize large models, which is not only a feasible choice for all parties in the industry but also an implementable strategy within their capabilities. For instance, simplifying model complexity by reducing unnecessary layers and structures, or reducing the total number of parameters of the model through parameter sharing. Simultaneously, more efficient neural network architectures, such as convolutional neural networks (CNN) instead of fully connected neural networks (FCN), can be selected to reduce the computational load during training.

Additionally, optimizing power management solutions and improving circuit design are excellent energy-saving ideas. New power management technologies, such as power management units (PMU) and intelligent power management chips, can dynamically adjust power supply strategies based on actual needs. This dynamic adjustment can ensure optimal energy efficiency of the device under different load conditions. In circuit design, reducing power consumption points, lowering switching frequencies, and optimizing layout can also significantly reduce circuit energy consumption during device operation.

Secondly, at the policy level, strengthen the element guarantee for the development of the AI strategic industry.

On the one hand, thoroughly implement the national policy of separate energy consumption for major projects and actively promote the inclusion of AI-related projects in the scope of separate energy consumption for major national projects, reserving sufficient energy space for industrial development. On the other hand, strengthen the energy supply guarantee for areas with concentrated deployment of AI computing power, such as computing power hub nodes and data center clusters, improve power matching and grid support capabilities, and promote the establishment of a "computing and power coordination" institutional mechanism to ensure the steady development of the AI industry.

To this end, industry insiders suggest that relevant institutions should quickly investigate the status of the AI industry, comprehensively grasp the current situation of the industry, and it is necessary for relevant national departments to take the lead in organizing research institutions and industry experts in related fields to conduct research on the AI industry in key domestic regions.

By conducting on-site visits to relevant enterprises and projects, systematically investigate the operating status of AI enterprises, project progress, technological innovation directions, etc., and select representative enterprises and projects to conduct regular monitoring and analysis of energy consumption in AI enterprises, providing a solid data basis for accurately assessing the energy consumption situation of the industry and scientifically guiding the development of the AI industry.

Thirdly, while providing adequate element guarantees, seize the strategic window period for the development of the AI industry, actively guide its industrial transformation and upgrading towards green and low-carbon directions, and accelerate the exploration and establishment of a green and low-carbon management mechanism suitable for the development of China's AI industry.

In fact, due to the fact that cooling often consumes a lot of electricity, data centers will tend to be deployed in relatively cool places in the future to reduce air conditioning load consumption. With the gradual increase in domestic demand for green electricity consumption, water (storage), wind, solar integration, or nuclear power will become preferred options.

Furthermore, AI-driven collaborative optimization will be one of the technical supports and innovation directions for the coordination of computing and power, that is, simultaneously optimizing computing power allocation tasks and power scheduling through AI models, such as task migration, which migrates high-energy consumption computing tasks to periods with sufficient green electricity; dynamic frequency reduction, which reduces server frequency during power shortages to reduce energy consumption, etc.

In this regard, industry insiders suggest that it is necessary to conduct extensive research on the advanced energy efficiency levels and green and low-carbon management systems and mechanisms of the AI industry at home and abroad, form successful case experiences for reference, and adopt management measures such as industry standards, incentive funds, and graded evaluation, from aspects such as AI model optimization, chip and algorithm efficiency improvement, data center green architecture, and green energy use, explore the establishment of a green and low-carbon management mechanism covering the entire life cycle of AI. Finally, select representative regions as pilots to accumulate experience in green and low-carbon management of the AI industry through regional pilot projects, and promote the green and low-carbon development of the AI industry with high standards.

Optimizing the spatial layout of the AI industry is imperative.

Given the current scarcity of resources like land and energy, the eastern region faces challenges in expanding data centers on a large scale. Consequently, it is crucial to further refine the spatial layout, harmonize the distribution of computing power and electricity, and direct AI projects with low latency requirements to prioritize the deployment of computing infrastructure in the western region. This will harness the abundant computing resources in the west to cater to the computing demands of AI projects in the east.

An industry insider concluded, "I believe 2025 will witness comprehensive integration, with energy and technology inextricably linked."

Indeed, AI is increasingly becoming an energy-intensive sector. Coordinating electricity and computing power is not merely a technical matter but a systematic endeavor encompassing economics, policies, and ecology. The industry must proactively take measures and plan ahead.