AI Boom Puts Pressure on China’s Power Grid as Data Centers Eat Up Electricity
By Du Zhihang, Bao Hongyun, Liu Peilin and Han Wei
China’s rapidly expanding artificial intelligence (AI) industry is eating up electricity at an alarming rate — a factor analysts say could disrupt the country’s power grid if infrastructure doesn’t keep pace.
By 2035, China’s data centers will consume 400 billion kilowatt-hours (kWh) of electricity a year — around four times their 2024 usage, Bloomberg analyst Lv Jinghong estimated at the BloombergNEF Summit in Beijing Thursday. That will account for 3.2% of national power consumption, comparable to the entire electricity usage of Sichuan province.
According to S&P Global figures, China’s data centers consumed 140 billion kWh in 2024, representing about 1.4% of total electricity usage. While national consumption rose 6.8% year-on-year, data center usage soared 31%. That growth is expected to continue at an average yearly rate of 20% through to 2030.
Behind the spike in demand for electricity is the energy-intensive nature of AI model training. BNEF data shows GPT-4 consumed 20 megawatts in a single training run. Some models require up to 154 megawatts per session — equal to the annual electricity use of 4,000 American households. This level of energy use demands a new approach to infrastructure, Lv said.
It’s not just computing equipment driving power consumption. Cooling systems account for 27% of a data centers’ energy use.
While China has improved energy efficiency — cutting the average Power Usage Effectiveness (PUE), the metric used to measure energy efficiency, from 1.7 in 2016 to 1.48 in 2023 — rising AI workloads and the growth of massive GPU clusters are reshaping the industry, placing increasing demands on land, cooling and power distribution, Lv said.
The Chinese government is pushing for greener data centers. On Tuesday, five agencies released new guidelines calling for further increases in green power usage at data center hubs, building on an existing 80% benchmark. China’s most advanced facilities now operate with PUE as low as 1.04, according to the National Data Bureau.
Nevertheless, grid stability is becoming a growing concern. In the AI era, GPU training and inference can cause power spikes exceeding five times normal loads — a serious challenge for grid management, Lv noted. While electric vehicles may drive broader electricity growth, data centers’ concentrated and volatile loads are harder for grids to absorb.
One solution is better site selection. China’s East-data-West-computing initiative, launched in 2022, is redirecting data center development to regions such as Inner Mongolia, Ningxia and Gansu — areas with cooler climates and abundant wind and solar energy.
Tech companies are already moving in. In November 2024, Tencent Holdings Ltd. launched China’s first integrated wind-solar-storage microgrid data center in Hebei province, with an installed capacity of 10.99 megawatts. Green power now accounts for 71% of electricity used at Tencent’s self-built centers. A month later, state-owned telecom operator China Unicom began construction on a similar green microgrid project in Qinghai province.
Despite such efforts, Lv cautioned that most data centers still rely on purchased green power due to land constraints limiting on-site generation.
Looking ahead, breakthroughs in chip design and AI algorithms could reduce the power burden. Lv pointed to DeepSeek, which achieved GPT-4-level performance using only 2,000 chips, as a sign that smarter software may ease hardware demands. Nvidia Corp. and others are also rolling out more energy-efficient AI chips — though geopolitical tensions could hinder global access to advanced technologies.
Contact reporter Denise Jia (huijuanjia@caixin.com)
caixinglobal.com is the English-language online news portal of Chinese financial and business news media group Caixin. Global Neighbours is authorized to reprint this article.
Image: Creator88 – stock.adobe.com