News Center

Text Zhao Xiaodong Deng Liangchen Wang Juan Fei Jiaying China Macroeconomic Research Institute, Energy Research Institute

Since November 2022, marked by the release of the landmark product ChatGPT by the American company OpenAI, AI large models have become a global technology hotspot, rapidly setting off a wave of technological innovation and application, and are reshaping all industries. The explosive growth in AI computing power demand in the United States has driven a rapid surge in electricity consumption for computing infrastructure, becoming a new "power-hungry beast" impacting the power system. The contradiction between computing power and electricity is becoming increasingly prominent. The power grids in areas with concentrated computing resources are overburdened, the carrying capacity of the power system has significantly decreased, and the risk of safe and reliable power supply has sharply increased, forcing IT giants such as Microsoft and Google to compete in cross-border, on-site power generation to meet the electricity needs of newly built computing infrastructure. China's computing power is also in a stage of accelerated development, and it is advisable to be prepared for any eventuality and deploy in advance to achieve organic synergy between computing power and electricity.

► Soaring US Computing Power Drives Surge in Electricity Consumption

1) US Generative AI Ignites Computing Power Demand

Generative AI generates new content by learning from large-scale datasets. Compared with traditional AI, it has a larger model scale, more complex scene algorithms, and rapidly climbing computing power demands. In terms of model scale, GPT-3, which can only process text, has 175 billion training parameters, while GPT-4, which has text and image processing capabilities, has approximately 1.8 trillion training parameters. The large-scale and complex nature of the model leads to a sharp increase in computing power demand. In terms of scene algorithms, with the iteration of generative AI, application scenarios have "upgraded" from simple text generation to video creation, leading to another explosion in computing power demand. According to relevant research, using the same text as input, the computing power required to generate a video is 1500 times that of text.

2) Electricity is a Key Element Supporting Computing Power Development

The massive demand for computing power from generative AI is driving the accelerated upgrading of computer hardware such as central processing units and AI computing cards. While the performance of computer hardware continues to improve, the power consumption of devices is also rapidly increasing, and the dependence on electricity is constantly deepening. For example, the new generation B200 computing card released by Nvidia in 2024 has 5 times the performance of the previous generation, but the power consumption per card has also increased from 700 watts to 1200 watts. Affected by this, servers equipped with new computing cards are expected to have a single-rack power consumption of 40-60 kilowatts, approximately 10 times the power consumption of conventional server racks.

3) Data Centers Become New "Power-Hungry Beasts"

Currently, generative AI innovation and iteration are active, and the penetration rate in various industries in the United States is constantly increasing. According to a survey conducted by McKinsey in the fall of 2023, 28% of respondents said they frequently use generative AI at work, with the technology sector accounting for as high as 33%, and the energy sector accounting for at least 14%. Although the efficiency of computing cards is constantly improving, the rapid expansion of generative AI applications will lead to a sharp increase in computing power and electricity demand. Research shows that OpenAI alone needs nearly 30,000 AI computing cards to support ChatGPT, with a daily electricity demand exceeding 500,000 kilowatt-hours. According to the latest data from the International Energy Agency, the United States has more than 2,600 operating data centers, and it is estimated that by 2026, the electricity consumption of US data centers will increase from approximately 200 billion kilowatt-hours in 2022 to nearly 260 billion kilowatt-hours, accounting for 6% of total electricity demand.

►   Three Manifestations of the Contradiction Between Computing Power and Electricity in the United States

1) Insufficient Existing Power Grid Capacity, Unable to Support Rapid "New" Data Centers

To meet the surging demand for computing power, the US AI industry's demand for new data centers is becoming increasingly urgent. It is estimated that by 2030, the total power consumption of US data centers will increase from 17 million kilowatts to 35 million kilowatts. Affected by this, data center bases close to power sources and close to user terminals have experienced severe insufficient power grid capacity, leading to obstacles in new data center projects. According to surveys, with the recent unexpected commissioning of a large number of data centers, the commissioning time of some new data centers has been extended by 2 to 6 years. Northern Virginia, the world's largest data center base, has advantages in energy and network infrastructure, but the scale of new data centers is less than 0.2% of expectations; Silicon Valley, Dallas-Fort Worth, and other data center bases close to end users have less than 0.5% and 1.9% of the expected capacity for new data centers, respectively.

2) Significant Mismatch Between Lagging Power Infrastructure Construction and Rapid AI Development Trends

The inter-regional power mutual assistance capacity in the United States has been weak for a long time. The unexpected increase in electricity consumption caused by generative AI far exceeds the US power infrastructure renewal and new construction planning targets, further exacerbating the power supply and demand contradiction. In terms of power infrastructure, the iteration and upgrading of US power infrastructure is slow, resulting in 70% of transmission lines and transformers nationwide having an operating life of more than 25 years, and 60% of circuit breakers having an operating life of more than 30 years. In 2023, the surge in computing power demand intensified the power supply and demand contradiction, and the US Department of Energy allocated nearly $3.5 billion for the first time to support the expansion of transmission capacity and enhance grid resilience. In terms of power planning, Dominion Energy, the leading power company in Northern Virginia, has included the project progress of new data centers in its operating area into its future power planning priorities in response to the rapid increase in data centers and the serious lag in power planning. In terms of regional mutual assistance, to cope with the surge in electricity demand from data centers, the PJM market operating organization, departing from its previous conventional approach of mainly relying on new power sources within the region, proposed planning to add multiple 500-kilovolt cross-regional ultra-high-voltage lines to improve regional power mutual assistance capacity.

3) New Characteristics of Data Center Loads Increase the Difficulty of Safe and Reliable Power Supply

Data centers have strong rigidity in electricity consumption and high reliability requirements. With the popularization of generative AI applications, the current electricity consumption of US data centers is characterized by a surge in electricity consumption and large load fluctuations. According to surveys, the peak-to-valley difference in electricity consumption for model inference in US data centers has reached as high as 4 times, and the peak-to-valley period   is highly matched with US economic activity, making load regulation more difficult. Because most US data centers require a high proportion of renewable energy supply, the new characteristics of data center electricity consumption and the volatility and randomness of renewable power generation will constitute a "double fluctuation" of source and load, making it increasingly difficult to ensure the reliable power supply of data centers.

►   Insights and Suggestions

The electricity challenges faced by the development of generative AI in the United States stem from the mismatch between the surge in computing power demand and the conventional development of the power system. Insufficient power grid capacity in the United States, lagging power infrastructure construction, and the "double fluctuation" of source and load between the new characteristics of data center electricity consumption and renewable electricity are the three main factors further exacerbating the contradiction between computing power and electricity.

Currently, China's AI large models and other artificial intelligence industries are developing rapidly, with applications becoming increasingly widespread across multiple sectors, accelerating the demand for computing power. To this end, China is actively building a nationwide integrated computing power network. By the end of 2023, the total computing power will reach 230 billion billion operations per second, ranking second globally, with expectations of continued growth at around 20% in the future. Simultaneously, China possesses the world's largest power system, with continuously strengthening cross-provincial and cross-regional resource allocation and mutual adjustment capabilities. Based on this, China adheres to a nationwide strategy, coordinating energy resources and computing power spatial layout. Through the "East Digital West Compute" project, it promotes the orderly transfer of medium-to-high latency computing power needs from eastern regions to western energy-rich areas, effectively addressing the surge in electricity demand caused by generative AI. In the next step, we should fully learn from the experiences and lessons of the United States in terms of how electricity constrains computing power development, promoting the synergistic development of computing power and electricity in the following aspects.

1) Proactive Planning of Computing Power and Electricity Spatial Layout

The distribution of data centers is positively correlated with population density, economic level, and commercial demand. In the United States, data centers are mainly concentrated in the capital (Washington), financial centers (New York, Chicago), and technology centers (San Francisco, Seattle). To address the contradictions of surging electricity consumption by data centers, relevant authorities and enterprises have been forced to adopt temporary countermeasures. China should coordinate the spatiotemporal layout of computing power and electricity, while scientifically planning "East Digital East Compute," "West Digital West Compute," and "East Digital West Compute" scenarios; it should conduct forward-looking assessments of potential electricity demand growth and grid capacity. Following the principle of "moderately ahead of schedule and with sufficient margin," it should promote the construction of power infrastructure and lay a solid foundation for the rapid development of computing power.

2) Comprehensive Improvement of Power Grid Optimization and Operation Capabilities

In recent years, affected by frequent extreme weather and the widespread access of new types of loads, global power load fluctuations have intensified, and the difficulty of electricity demand forecasting has increased daily. The surge in generative AI computing power demand in the United States has further exacerbated the peak-valley difference in regional power grid loads. Faced with the accelerated evolution of load characteristics under the new situation, China should combine the construction process of new power systems, strengthen the prediction of new loads such as data centers, electric vehicles, new energy storage, and hydrogen energy, analyze the trend of electricity demand changes, adjust electricity supply in a timely manner, and optimize   power grid resource allocation and dispatch operation, improving the safe and stable operation level of the power grid.

3) Strengthen Data Center Power Load Management  

Currently, the power supply systems of computing power infrastructure generally adopt redundant operation strategies, with considerable potential for energy saving and emission reduction. It is suggested to further improve real-time electricity monitoring of various types of data centers, strengthen the coordinated control of data center power supply systems and data services, implement refined management of energy and electricity use, and achieve overall optimized operation in terms of energy efficiency, cost, and computing power.

-------------------------------------------------------------------------------

Information Source: [Think Tank Voice] The Current Situation and Implications of the Impact of Artificial Intelligence Development in the United States on Power Systems｜China Investment