Summary of Key Points
This news article reveals a truth that is often overlooked: although AI appears to be a virtual digital entity, it consumes a significant amount of real-world water resources. From the cooling of chips to the production of electricity, every process involved in AI’s thinking and training requires water. The article not only quantifies the amount of water consumed by AI but also analyzes how China balances the demand for computing power with water distribution through the “East Data West Computing” initiative. It proposes technical approaches to reduce AI’s water footprint, emphasizing that the digital world cannot exist independently without real-world resources.
1. Why Does AI Consume So Much Water? – The Need for Cooling
The “brain” of AI consists of GPU chips in data centers, which generate high amounts of heat (one chip is as power-consuming as a microwave oven). When thousands of these chips are running simultaneously, the temperature inside the cabinets can exceed 85°C, leading to system lagging or even damage. To reduce this heat, various cooling methods are used:
- Air cooling: Similar to using a fan on a computer, it is inexpensive but inefficient and suitable only for traditional data centers.
- Immersion liquid cooling: Chips are immersed in a special coolant, which is effective but costly (similar to advanced water cooling systems for computers, not feasible on a large scale).
- Plate-type liquid cooling: This is currently the most efficient method as it separates the chips from the coolant, balancing performance and cost.
More importantly, the cooling systems themselves consume electricity (pumps, fans, etc., accounting for 15%-25% of the total power usage in data centers), and electricity production itself relies on water (thermal power plants all require water). This means that AI indirectly consumes water as well.
2. How Much Does AI Cost in Water? – The Heavy Burden of Training and Operation
The most water-intensive phase of AI is training. For example, training GPT-4 required thousands of GPUs running continuously for weeks, with the cooling systems operating at full capacity, consuming a total of 18,900 tons of water—equivalent to the monthly water usage of 5,000 Chinese households.
The daily operation of AI systems also consumes significant amounts of water. A large data center can use millions of tons of water per year, equivalent to the total water consumption of a medium-sized city. Moreover, most of the water used for cooling is lost through evaporation and cannot be recycled.
3. Where Does Data Center Power Come From? – The Competition for Water Resources
China’s demand for computing power is concentrated in the eastern regions, which have dense populations and many enterprises. However, these areas face challenges such as limited land, high electricity costs, and scarce water resources. As a result, new data centers with high water demands are prohibited, and recycled water is used for cooling (for example, half of the water used in data centers in Nanning and Guangzhou is recycled).
The western regions, on the other hand, have abundant water and cooler climates, making them ideal for data centers. The government has promoted the “East Data West Computing” initiative, with the east handling online services and the west handling offline training tasks. However, the western regions also face challenges, such as hard water quality that requires costly treatment and high evaporation rates in some areas.
4. Can We Reduce AI’s Water Consumption? – Technological Innovation Is the Key
While it is unrealistic to eliminate the use of AI, technological advancements can help reduce its water footprint:
- Reducing heat generation: Using low-power chips (e.g., neuromorphic chips) or more efficient models that require less computation.
- Improving cooling efficiency: Enclosed liquid cooling systems that prevent water evaporation, and underwater data centers (using seawater for cooling, tried by Microsoft and China).
- Reusing waste heat: Turning the waste heat generated by data centers into heating energy or for driving refrigeration systems. In Europe, some data centers even connect their waste heat to the urban heating network.
5. The Digital World Is Not an Island – The Importance of Real-World Resources
Although AI’s current water consumption is still lower than that of industrial and agricultural activities, it highlights the reality that the digital economy is not self-sufficient. Every interaction with AI is connected to real-world resources such as water and electricity. In the future, we need to develop more powerful AI systems while ensuring that every resource is used efficiently. After all, virtual intelligence cannot exist without real-world infrastructure.