What's Happening?
AI systems are encountering significant power shortages, leading to stalled computing power projects. A notable instance involves a green power direct connection project in western China, which has been deadlocked for six months over a mere 2-cent difference
in electricity price between the power generator and the computing power side. This seemingly small difference translates to millions of yuan in cost for a 100-megawatt data center, where electricity can account for 60% to 70% of operating costs. The demand for computing power is rising exponentially, with gigawatt-level intelligent computing centers becoming common. Gartner predicted that by 2030, the power consumption of national computing centers could account for about 6% of the total social electricity consumption, up from 1.6% in 2025. The core issue is not a lack of total energy in China but rather the uneven distribution of resources and loads, coupled with insufficient support from local power grids. Some new intelligent computing projects demand up to 75% of a regional power grid's upper supply limit, necessitating massive investments in new grid nodes or even rebuilding local power grids from scratch.
Why It's Important?
The severe power shortages and the resulting deadlock in computing power projects pose a critical challenge to the advancement and scalability of AI in the U.S. and globally. The high energy consumption of AI, particularly with the advent of agent-based AI requiring 5 to 30 times more token consumption than standard generative AI, means that energy infrastructure is becoming a bottleneck for technological progress. This situation highlights that the 'end of AI is computing power, and the end of computing power is energy.' The inability to agree on electricity prices, even for small differences, can halt projects for extended periods, impacting innovation and economic growth. The need to rebuild local power grids or invest heavily in new infrastructure to support AI demand underscores the significant financial and logistical hurdles. Furthermore, the low average GPU utilization rate (less than 30%) in domestic intelligent computing centers exacerbates the problem, as resources are 'occupied but not used,' further straining power supporting facilities and increasing demand.
What's Next?
To address the deadlock, the industry is exploring 'computing-power and electricity coordination,' which involves data centers strategically managing electricity usage from site selection to operation. Three main paths are being pursued: advancing energy planning for new projects, optimizing electricity costs for existing computing centers, and incorporating energy storage into new project planning. For instance, Damao Technology used energy large models and agents at SenseTime Shanghai Lingang AI Computing Center to optimize energy storage, reducing the comprehensive cost per kWh by 9.73%. Policies are also being developed, with 'computing-power and electricity coordination' included in government work reports and the 15th Five-Year Plan outline. The National Development and Reform Commission has mandated an 80% green power consumption ratio for new data centers in national hub nodes. However, despite over 100 approved green power direct connection projects, only a few have started construction, indicating a gap between policy and implementation. The industry anticipates significant changes in the next one to two years as policies, technologies, and cognition align to tackle this energy challenge.
Beyond the Headlines
The energy crisis facing AI systems reveals a deeper, systemic challenge in balancing technological advancement with sustainable resource management. The competition for AI dominance, particularly between countries like China and the U.S., is increasingly becoming a race for energy efficiency and availability. The U.S. is exploring natural gas, nuclear power, and energy storage solutions, while China is focusing on green power direct connections. This global energy competition for AI could accelerate the energy revolution, pushing for more efficient and sustainable power sources. The ethical implications extend to environmental sustainability, as the massive energy demands of AI contribute to carbon emissions if not powered by green sources. The economic impact could also lead to a redistribution of computing power centers, favoring regions with abundant and affordable energy. The concept of data centers evolving from passive power loads to active, flexible, and value-added energy units through smart grid integration represents a paradigm shift in energy management, potentially transforming the entire energy landscape.
