Caixin Weekly | The Endgame of Computing Power Is Electricity (AI Translation)

财新周刊｜算力的尽头是电力

Summary

The exponential growth of artificial intelligence (AI) is making energy availability, especially of clean electricity, a critical constraint in the development of data centers—the foundational infrastructure of the digital economy. Key tech leaders such as Elon Musk and Jensen Huang have highlighted power supply as the ultimate limiting factor for future computing growth. The International Energy Agency (IEA) estimates that global data centers consumed 415 terawatt-hours (TWh) in 2024, or 1.5% of worldwide electricity consumption—a figure projected to more than double to 945 TWh by 2030, surpassing Japan's total annual power use. The U.S. is expected to lead data center electricity demand growth globally, followed by China, where demand is accelerating at a much faster rate than overall electricity usage, with consumption by China's data centers projected to increase at a 17% annual rate through 2030, eventually making up 6% of the nation’s power use [para. 1][para. 2][para. 3][para. 4].

Although China’s power supply is abundant due to rapid market development over the past two decades, mismatches persist between regions rich in renewable energy (western China) and regions with high computing demand (eastern China), a problem the "Eastern Data, Western Computing" policy since 2022 aims to address. However, operational bottlenecks such as a shortage of technological talent and insufficient local support in western China hinder further optimization [para. 4][para. 5][para. 6].

The rise of AI—especially large language models—means data generation and processing have become vastly more energy intensive. For example, generating 5,000 words from a language model consumes about 1 kilowatt-hour (kWh), while a minute of 4K video can require up to 5 kWh. Modern AI data centers are much more power-hungry than traditional cloud centers, with energy consumption per rack reaching as high as 130 kilowatts in top-tier liquid-cooled setups [para. 7][para. 8][para. 9]. As tech giants ramp up investments, capital expenditure for data centers is rising sharply, with leading Chinese and U.S. firms doubling down on new construction, even as domestic telecom operators in China slow investment after a period of rapid expansion [para. 9][para. 10][para. 11][para. 12].

Energy efficiency advances may temper future power needs. Improved chips and liquid cooling have reduced the Power Usage Effectiveness (PUE) from 1.5–1.6 in air-cooled systems down to close to 1.2 in liquid-cooled ones, potentially reducing cooling’s share of total power consumption to below 10%. However, average GPU utilization in Chinese AI data centers remains below 30%, and less than 5% of China’s centers are heavily used for AI model training [para. 19][para. 20][para. 21][para. 22][para. 23][para. 24].

Achieving green electricity use is critical to reducing carbon emissions, with China aiming to raise the share of green power in new data centers above the current 80%. Demand for renewable energy in China’s data centers is forecast to triple by the end of 2025, requiring 40–63 billion kWh of green power and posing significant operational and grid coordination challenges. The physical and infrastructural gaps between power-rich and demand-rich regions mean that, despite incentives, talent shortages and unreliable infrastructure in the west, plus renewable intermittency, impede full synergy [para. 25][para. 26][para. 27][para. 28][para. 29][para. 30][para. 31][para. 32].

Solutions explored include long-term green power purchase agreements (PPAs), large-scale integration with renewables and storage (“source-grid-load-storage”), and microgrid deployment. Tencent, for instance, has built a pioneering microgrid data center in Hebei, utilizing on-site solar and wind, providing up to 15% of its needs at a significantly reduced cost. Still, across China, distributed generation typically covers less than 5% of on-site usage due to limited available roof space [para. 33][para. 34][para. 35][para. 36][para. 37][para. 38][para. 39][para. 40][para. 41][para. 42].

Local governments offer competitive electricity prices—sometimes as low as 0.26 yuan per kWh, compared to 0.6 yuan elsewhere—to attract data centers as growth engines for their digital economies. However, sustained low rates pose challenges for grid companies, and infrastructure remains a decisive factor: only regions with strong connectivity, land, reliable power, and proximity to major markets (like Zhangjiakou and Ulanqab) can consistently attract large-scale data center investments [para. 43][para. 44][para. 45][para. 46][para. 47][para. 48][para. 49][para. 50][para. 51][para. 52][para. 53][para. 54][para. 55][para. 56][para. 57][para. 58].