How to Balance Deployment of Power-Hungry Data Centers? (AI Translation)

【周刊提前读】电力是算力之源 东西部之间如何平衡算力布局？

[para. 1] The rapid advancement of artificial intelligence (AI) is placing unprecedented demands on electricity, positioning energy supply—especially clean electricity—as a primary limiting factor for computing power. Tech leaders like Elon Musk and Jensen Huang repeatedly highlight the necessity of robust, green energy to sustain the AI industry's growth. With the proliferation of AI technologies, global data center electricity consumption is projected by the International Energy Agency (IEA) to rise from approximately 415 terawatt-hours (TWh) in 2024, or 1.5% of global demand, to 945 TWh by 2030, surpassing Japan’s current annual usage [para. 2]. By 2030, data center electricity needs will represent around 10% of the world’s new power consumption, a growth rate outpacing most other sectors [para. 3]. The United States will lead this growth, followed by China, whose data centers are expected to consume 140 billion kWh in 2024—a 31% year-on-year increase, compared to the nation’s overall growth rate of 6.8% [para. 4]. By 2030, Chinese data centers are forecast to use 400 billion kWh, raising their share of national power consumption to 6% [para. 5].

[para. 6] China enjoys superior power supply capacity due to significant expansion in the past two decades, but faces issues of regional supply mismatches, particularly with renewable resources concentrated in the west and demand in the east [para. 7]. Although western China has abundant renewable energy for data centers, a lack of technical experts limits development [para. 8]. Success depends on new mechanisms for green, affordable electricity and better synergy between power markets and computing needs [para. 9].

[para. 10] The electricity consumption for AI operations is considerable: generating about 5,000 words with a large AI model or 1,000 images both use roughly 1 kWh, while one minute of 4K video production may require up to 3–5 kWh [para. 10]. The transition toward intelligent computing centers—requiring up to 130 kW per rack (with liquid cooling)—is driving capital investment, especially as major U.S. and Chinese tech firms double or triple data center spending [para. 11][para. 13]. From 2022–2027, China’s intelligent computing capacity is predicted to grow at 33.9% annually [para. 14]. In 2024, over 300 intelligent computing centers were under construction or planned in China [para. 15].

[para. 17] However, telecom operators, who control about 60% of the market, are slowing investment after early expansion [para. 17][para. 18]. Infrastructure is robust, with massive GPU clusters established by leading operators like China Telecom and China Mobile, but utilization rates remain below 30%—a sign of potential overcapacity [para. 20][para. 21]. In contrast, U.S. data center expansion faces grid saturation and energy shortages, with electricity demand projected to rise 2% yearly through 2027 [para. 23][para. 25]. U.S. data centers primarily train large models and mine cryptocurrency, while less than 5% of Chinese centers focus on large-scale AI training [para. 27].

[para. 29] Advances in chip efficiency and cooling (liquid cooling reduces energy from 1.5–1.6 PUE down to 1.2 or below) are moderating some of the increased demand [para. 31][para. 32].

[para. 35] Achieving low-carbon operations is crucial: China’s 2030 data center emissions could exceed 200 million tons if relying on fossil fuels [para. 35]. Policies now push for new centers to obtain more than 80% of power from renewable sources [para. 36][para. 37]. The west is favored for large-scale training due to cheaper, cleaner energy, while eastern centers support low-latency, high-performance applications [para. 39][para. 40]. Example projects in Ningxia reflect this trend [para. 43].

[para. 47] Yet, challenges remain, including workforce shortages, intermittent renewable supply, and grid operation mismatches with rapidly fluctuating computing loads [para. 48][para. 49][para. 52]. The transition is further slowed by fragmented electricity and computing markets, and nascent green electricity trading mechanisms [para. 55].

[para. 57] Data centers are experimenting with new approaches such as multi-year green power purchase agreements (PPAs) and on-site storage to ensure stable, green energy supply [para. 57][para. 59]. On-site solar typically covers less than 5% of data center needs [para. 74], but projects integrating local wind, solar, and storage—such as Tencent’s Huailai microgrid and a major Ulanqab computing center—show promise [para. 65][para. 83].

[para. 73] Localized direct green power and load-storage integration could supply up to 40% of large data centers’ power, minimizing costs and raising renewable shares [para. 73]. National policies encourage, but implementation is difficult outside exceptional cases like the Sanxia Dongyuemiao hydropower data center [para. 80][para. 81].

[para. 88] Local governments vie to attract data centers with low electricity prices—down to 0.26 yuan/kWh in Inner Mongolia, boosting local economies [para. 89][para. 91]. Some cities, such as Qingyang, expect digital industries to account for up to half their economic output by 2025 [para. 95]. Non-hub regions are rapidly increasing investment, often with provincial support measures [para. 97][para. 99].

[para. 101] Comprehensive infrastructure (land, grid, fiber) and affordable green energy are often more critical than electricity prices alone for competitiveness. Locations with these advantages—such as Zhangjiakou and Ulanqab—continue to attract top tech firms and large-scale investments [para. 107].

In summary, as AI and digital economies expand, energy supply—especially clean, stable, and affordable electricity—has become both the primary constraint and key enabler for global computing power infrastructure. The future lies in efficient, scalable integration between power and computing through innovations in technology, policy, and market coordination [para. 1][para. 2][para. 3][para. 109].