Summary of Key Points
This news article highlights the new trends in data centers in the AI era. Ulanqab, with its low electricity prices, cold climate, and proximity to Beijing, has become a popular location for companies such as Alibaba, Huawei, and ByteDance to set up their data centers. Behind this trend is the explosive demand for AI computing power, which has led to a shift in the logic of data center选址. The focus is no longer on being close to users, but on balancing costs and latency. In terms of operation, data centers need to be cost-effective while also ensuring reliability through redundant designs. The physical form of data centers has evolved from traditional large buildings to modular systems that can be easily assembled. In the future, there will be a two-way distribution of tasks: training algorithms will take place in the west, while inference processes will be handled in the east.
Detailed Analysis
1. Why Ulanqab Became the “Power Hub for AI?”
Ulanqab's appeal to large companies stems from three key advantages that address the challenges of the AI era:
- Excessively Low Electricity Prices: The cost of electricity ranges from 0.32 to 0.35 yuan per kilowatt-hour, which is more than half lower than in eastern cities. Alibaba has calculated that by moving its data centers here, it could save over 5 billion yuan in annual electricity costs. AI training requires a large amount of power (with cabinet capacities increasing from 10千瓦 to 100千瓦), making electricity expenses the largest cost factor. Low prices are a significant advantage.
- Natural Cooling: With an average annual temperature of 4.3°C, Ulanqab has naturally cold weather that eliminates the need for air conditioning, further reducing energy costs.
- Proper Distance from Beijing: Located 350 kilometers from Beijing, the network latency is only 4 milliseconds. Since AI training does not require real-time responses (unlike applications in finance or gaming), this distance allows companies to take advantage of the lower costs in the west while still being able to serve the eastern markets quickly.
2. AI Has Changed the Logic of Data Center Location
The approach to data center selection has changed dramatically between the cloud computing and AI eras:
- Cloud Computing: Locations were chosen close to users (e.g., Alibaba Cloud’s former base in Zhangjiakou, near the Beijing-Tianjin-Hebei region) for applications that required immediate responses.
- AI Era: The need for massive computing power has increased electricity costs significantly. Since chip prices remain high, reducing electricity expenses is crucial. AI training can tolerate some latency (e.g., several days of training with a few milliseconds of delay), allowing companies to relocate data centers to lower-cost areas in the west.
3. Data Center Operations: Balancing Cost Efficiency and Reliability
Operating a data center involves carefully managing both costs and safety:
- Cost-Efficient Measures: Electricity must be converted multiple times before it can be used (e.g., from alternating current to direct current), resulting in energy losses. Alibaba has developed its own “Panama System” to minimize these conversions and uses more energy-efficient DC air conditioning.
- Reliability Enhancements: Data centers are designed with redundancy to prevent power outages, including large batteries that can provide power for 15 minutes and 14 diesel generators (with one extra as a backup). These measures are necessary because an interruption in AI training can result in significant losses.
4. Modular Data Centers: Speeding Up Delivery
The demand for computing power in the AI era has made delivery speed a competitive factor. Alibaba Cloud’s CUBE 5.0 architecture uses modular components that can be manufactured and assembled on-site, similar to building with Lego bricks:
- Time Savings: Construction time has been reduced from six to twelve months to just over three months.
- Flexibility: These modules are compatible with at least three generations of chips, eliminating the need for costly rebuilds when chip technology evolves.
5. Future Trends: Distributed Data Center Layout
The layout of data centers will become more targeted:
- Training in the West: Large-scale training clusters will be moved to areas west of the Huaihuan Line (e.g., Xinjiang and Gansu) where energy is more abundant and costs are lower.
- Inference in the East: Inference tasks (e.g., answering user queries) will remain near users in the eastern regions (Beijing-Tianjin-Hebei, Yangtze River Delta). This ensures that different types of computing tasks are allocated to the most suitable locations, maximizing efficiency.
Conclusion
The landscape of AI computing power is not static; it is a dynamic process of finding the best options. Ulanqab represents this trend, but in the future, every city will need to determine its role based on local advantages (electricity prices, location, and industry needs). Companies will place different types of computing resources in the most cost-effective locations.