The 33rd China International Exhibition on Electric Power Equipment and Technology
Shanghai International Energy Storage Technology Application Expo / Hydrogen Energy Expo
AI power capacity refers to the specialised electrical power infrastructure needed to support artificial intelligence computing workloads, which are among the most energy-intensive applications in the digital economy. Training large AI models requires clusters of thousands of high-performance GPUs or AI accelerators, each consuming 300–700 W, housed in data centres with total IT loads of 10–100 MW or more. AI inference — running trained models at scale — creates sustained, predictable power demand that differs from traditional data centre workloads in its density (up to 100 kW per rack for liquid-cooled GPU clusters), its continuous operation profile, and its sensitivity to power interruptions. The rapid growth of AI computing is driving unprecedented demand for new data centre capacity, grid connections, and power infrastructure. In China, major AI hyperscalers (Alibaba, Baidu, Tencent, ByteDance, Huawei) and cloud providers are investing heavily in AI-optimised data centres, creating substantial demand for high-density power distribution, precision cooling, backup power, and grid connection infrastructure.
5 Key Questions About AI Power Capacity
AI computing is extraordinarily energy-intensive. Training a single large language model (LLM) such as GPT-4 is estimated to consume 50–100 GWh of electricity — equivalent to the annual consumption of thousands of households. AI inference at scale consumes continuous power: a major AI service provider may operate tens of thousands of GPUs drawing 300–700 W each, for a total IT load of 10–50 MW per data centre. The International Energy Agency (IEA) projects that global data centre electricity consumption will double by 2026, driven primarily by AI workloads. In China, the government's AI development strategy is accelerating investment in AI computing infrastructure, with significant implications for power grid planning.
AI data centres require specialised power infrastructure compared to traditional data centres: higher power density (up to 100 kW per rack for liquid-cooled GPU clusters vs 5–15 kW for standard IT racks); more robust UPS and backup generation to protect against power interruptions that would abort long AI training runs; precision power distribution units (PDUs) with high-density outlet configurations for GPU servers; liquid cooling infrastructure (direct liquid cooling or immersion cooling) to manage the heat generated by high-density AI compute; and direct high-voltage grid connections (typically 110 kV or 220 kV) for large AI campuses to avoid the losses and reliability limitations of lower-voltage connections.
The rapid growth of AI data centres is creating new challenges for power grid planning in China. Large AI campuses with 100–500 MW of connected load require dedicated transmission connections and substation capacity that must be planned years in advance. The geographic concentration of AI data centres in regions with low electricity prices (Inner Mongolia, Guizhou, Gansu) and renewable energy resources is creating localised grid congestion. China's 'East Data West Computing' (东数西算) policy directs AI data centre development to western regions with surplus renewable energy, requiring long-distance data transmission infrastructure and creating new demand for UHV transmission capacity.
Power Usage Effectiveness (PUE) is the primary energy efficiency metric for data centres, defined as total facility power divided by IT equipment power. World-class data centres achieve PUE of 1.1–1.2; the Chinese national standard GB/T 32910 sets a PUE target of 1.3 for new data centres. For AI data centres with high-density liquid cooling, PUE values approaching 1.05 are achievable. The AI-specific metric of Performance per Watt (operations per second per watt) measures the energy efficiency of AI computing hardware. Liquid cooling — direct liquid cooling, rear-door heat exchangers, and immersion cooling — is essential for achieving low PUE in high-density AI deployments.
Major AI hyperscalers are committing to 100% renewable energy for their data centre operations through a combination of on-site renewable generation (rooftop solar, wind), power purchase agreements (PPAs) with renewable generators, and renewable energy certificates (RECs). In China, the Green Power Trading market enables data centres to directly purchase renewable electricity from generators, providing a credible renewable energy claim. Some AI data centres are being co-located with large renewable energy projects — solar farms and wind parks — to maximise direct renewable power supply and minimise transmission losses. Battery energy storage is increasingly deployed alongside renewable generation to provide firm power for AI workloads that require continuous operation.
Key Takeaways
AI power capacity is one of the fastest-growing segments of power infrastructure demand, driven by the explosive growth of AI computing workloads that require high-density, highly reliable, and increasingly renewable-powered electrical infrastructure. China's national AI development strategy is accelerating investment in AI data centres and the grid infrastructure to power them. EP Shanghai showcases the power distribution, cooling, backup power, and grid connection technologies that enable the AI computing revolution.
Firefox
Chrome
Safari
QQ
Edge