Rethinking China’s Data Center Strategy for AI dominance: Cooling Methods, Energy Efficiency, and Heat Recycling
Written by Henry Robinson; Edited by Andrew Ma
Published on April 11th, 2025
Introduction
With the proliferation of communication and network technology, data centers have become increasingly more important to maintain low latency and increase the pace of online services. This growth is driven by expanding needs for data processing, storage, and digital communication, which will naturally lead to higher energy consumption. Without sufficient investment in data centers and promotion of efficient energy use, generative AI is unlikely to realize its full potential. It is estimated that AI could unlock between 2.6 and 4.4 trillion dollars throughout the global economy if data center electricity demand is met. As a result, energy efficiency has become the primary concern for operators, surpassing even security. Improving cooling efficiency, often measured by achieving a lower Power Usage Effectiveness (PUE), is the most effective way to reduce energy consumption.
Data centers are in essence large supercomputers that are prone to overheating, and as such need to cool off or expel the excess heat or risk breaking down in a matter of hours. For every 10 degree celsius increase, the data center failure rate increases by 50%. Data center cooling and energy consumption are inextricably linked, since cooling systems alone account for about 40% of a data center’s energy consumption. IT equipment consumes 45 percent of the energy in data centers, however, maintaining computational capacity remains a priority and so there isn’t incentive to reduce power consumption within servers. The remaining 15 percent is split between lighting and auxiliary power systems which have little capacity to be improved. Currently, data centers account for approximately 1.3% of global electricity consumption, with projections suggesting a 12–14% increase in demand over the next two to five years.
In 2022, China’s National Development and Reform Commission (NDRC) launched the Eastern Data, Western Compute (EDWC) initiative, which plans to redistribute data centers into China’s west for lower energy costs compared to urban servers. The EU has taken a different course, prioritizing data centers in Nordic Countries. Iceland, Sweden, Finland and Norway ranked as top European locations for data centers based on an evaluation of risk factors including energy costs, connectivity and climate stability. Both China’s western relocation and the EU’s northern focus aim to take advantage of climate and geographic conditions.
This paper examines the evolving strategies for data center placement and cooling efficiency, with a focus on China’s Eastern Data, Western Compute (EDWC) initiative and its implications for AI development. While the EDWC aims to leverage the west’s lower energy costs and favorable climate, this paper argues that China’s path to AI dominance may lie in reorienting its strategy toward air-cooled data centers in its northern regions following the success of European data centers.
By analyzing case studies from Xinjiang, Stockholm, Hohhot, and Finland, the paper highlights the advantages of naturally cold climates and waste heat reuse in achieving superior Power Usage Effectiveness (PUE). It also critiques the limitations of water-cooled systems in water-scarce western provinces and emphasizes the untapped potential of northern China’s severe cold climates for district heating integration. Ultimately, the paper contends that prioritizing air-cooled data centers in the north, rather than the west, could enhance energy efficiency, reduce costs, and solidify China’s position as a global AI leader.
Temperature-based Free Cooling
Low temperatures, reduced relative humidity, and dry climates are considered ideal for data centers since HVAC economizers don't need to lower the temperature of outside air for server cooling. Air economizers, responsible for air-cooling, are better suited in colder temperatures since mechanized refrigeration is not required. For fans to create a 2°C decrease in indoor environmental temperature, they consume 2.8%–8.5% more energy. This incentivizes data centers to be placed in naturally cold locations to reduce energy costs.
Xinjiang’s low humidity and 10.0°C average temperature is well-suited for efficient power usage through direct air cooling. When the ambient air temperature exceeds 15°C, indirect cooling with fans is activated. However, 65.9% of operating hours at data centers in Xinjiang utilize direct air cooling, resulting in an average PUE of 1.62. Despite these favorable conditions, Urumqi’s data centers have not yet achieved PUE levels as low as those in Stockholm, Sweden.
Stockholm’s data centers that do not use direct air cooling have an average PUE of 1.54, while those that do achieve an average of 1.40. Stockholm’s naturally cool ambient temperatures, ranging from -4°C to 16°C, allow data centers to remain below the threshold requiring mechanized cooling (24°C) for most of the year. Based on the city’s monthly temperature data, mechanized cooling is only required during the summer months. From May to September, Stockholm’s PUE rises above 1.40, peaking at 1.41 in July and August. During the colder months, however, PUE remains below 1.39.
The 30 Southern European data centers in Italy, Malta, Spain, and Greece have an average PUE of 2.00 (Avgerinou 12). However there exists reason to believe that data centers in central Europe can match or exceed those in Nordic countries with investment from the proper avenues. The Meo Covilha data center in Portugal, constructed by Facebook has a PUE of 1.25 and The FRA 01 data center in Germany has a PUE of 1.3.
Similarly, cloud computing projects by Shanghai Data Solutions (SDS) could help reconcile the PUE disparity between Shanghai and Urumqi. Responsible for the Unionpay and IDX Blue Light data centers, SDS also built the IDX Jinqiao which relies on air free cooling. With a PUE of 1.5, it boasts a 13,000 sqm floor space with “cabinet-style” server cabinets. Shanghai has a high annual relative humidity of 80.3% and an average temperature of 16.6°C, yet has been able to remain competitive with Nordic data centers.
On the macro-level, the 122 air-cooled data centers in central Europe have an average PUE of 1.71 in 14-28 degree celsius temperatures and a relative humidity of up to 75 percent. By contrast, the 42 air-cooled data centers in Hohhot, China exist in average ambient temperatures of 17.3 Celsius and 75.7 percent humidity and have PUE as low as 1.29. This shows that free air cooled data centers in China can compete or exceed those in northern Europe. One prominent example in China is that of AliCloud, which opened the Qiandao Lake Data Center in 2015. This facility pumps 17°C water 35 meters below the surface of Qiandao Lake. This temperature allows for free cooling 90% of the year and an 80% reduction in energy costs. The Qiandao Lake data center, part of the Yangtze Delta Cluster, maintains a PUE of 1.27 and a Water Usage Effectiveness (WUE) of 0.2.
While the Qiandao Lake Data Center is an example of utilizing the naturally low temperature for free cooling, other data centers have optimized their cooling even further. In 2011, Google purchased a facility previously owned by Finnish paper company Stora Enso. Located in Hamina, Finland, Google spent $273 million to transform the old Summa paper mill into an advanced data center that relies on direct water intake from the Gulf of Finland for cooling, where 80 percent of the cooling comes from recycled seawater. The annual average seawater temperature off the coast of Hamina is 9°C, reaching 12°C in summer and falling to 1°C in winter. The Hamina data center achieves a PUE of 1.09, outperforming the Qiandao Lake Data Center, which has a PUE of 1.27.
However, temperatures far below that of Google’s data center in Hamina may not necessarily be the most efficient. One study concluded that the optimal water temperature for data center cooling is around 16°C, with temperatures below 12°C being detrimental to energy efficiency . Seawater off the coast of Hamina reaches the optimal 16°C temperature only in July, August, and September. For the remaining nine months, the seawater temperature is typically below 12°C. Despite these colder temperatures, the PUE remains at 1.09 year-round, according to Google reports. While water temperatures at Qiandao Lake fluctuate between 12-16°C for most of the year, it performs less efficiently than the Hamina facility with a PUE of 1.27. This case study suggests that although the Chinese have optimal temperatures, they still underperform compared to counterparts in Finland.
It is unlikely the success of the Qiandao Lake Data center will be replicated elsewhere in China since data centers face water scarcity in their push westward. Only 4 out of China’s 29 provinces have reached ideal water resource efficiency and there were relative downtrends in access to water in 14 provinces between 2015 to 2020. Compared to the east and central regions, provinces in the west require even more support to leverage their water resources. For this reason, China would be remiss to rely on water-cooled data centers in place of data centers cooled by air-economizers.
Heat Recycling
In Europe, higher longitudes have a higher need for heating than the climates closer to the Equator. Waste heat from data centers is exceedingly common in Nordic climates as it accounts for 3.3% of heating in Finland and 8% in Sweden. Conversely, data center waste heat usage drops in China’s southernmost cities. In a southward trend, Beijing, Shanghai, Guiyan, and Guangzhou use 35.6%. 17.5%, 19.2%, and 0.0% waste heat respectively. In China, 85% of district heating is powered by coal and natural gas, while less than 4% comes from data centers. This highlights the limited role data center waste heat currently plays in China’s district heating systems nationally.
In 2015, online services giant Yandex opened its first non-domestic data center in Mäntsälä, Finland. The center has achieved an impressive 1.15 Power Usage Effectiveness (PUE) and an energy reuse rate of 31%, without the need for heat pumps, as the data center’s 40°C temperature is sufficient for local heating.The server-generated heat is then sent to the Mäntsälä district heating system, providing heating for over 2,500 homes.
A similar data center in Hohhot, China, also directs waste heat to local homes but instead uses water cooling. Classified as a "severely cold" region, Hohhot experiences 180 days per year when the outdoor temperature falls below 7.5°C, resulting in high demand for heating. This data center boasts a high energy efficiency rating, with a PUE of 1.17 and a waste heat recovery of 65%. The advantage in heat recovery over the Mäntsälä data center can be attributed to using liquid cooling which retains more heat than air.
Unlike electricity, heat cannot travel long distances without significant loss of temperature, so there must be sufficient demand for heat nearby. This relationship is evident in Latvia, where 23 out of the country's 26 data centers are located within 500 meters of a district heating network. However, In China, the northeastern provinces of Inner Mongolia, Liaoning, Jilin, and Heilongjiang, which are classified as having a severe cold climate, collectively host urban district heating systems covering 3,576 million square meters. This represents 34% of China’s total district heating area, even though these provinces account for only 1% of the country’s data centers. Despite high demand for heating, there are few data centers with the capacity of waste heat reuse which represents a missed opportunity for increasing energy efficiency in Northern China.
The primary motivation for the EDWC is improved scalability of data centers by leveraging more abundant resources in west China. The cost reduction of data centers begins by building them where energy and land are cheap. In Beijing, industrial land costs 981 yuan/m2 while in Guizhou, it is 252 Yuan/m2. Guizhou also has lower energy costs at 0.5787 Yuan/kWh while in Beijing it costs 0.7673 Yuan/kWh. As such, the number of big data enterprises in Guizhou grew from fewer than 1,000 in 2013 to 9,551 in 2018. However, the northern province of Ningxia is even more cost effective for data center replication than Guizhou. In 2021, Ningxia industrial land cost 79 yuan/m2 and energy cost 0.4883 Yuan/kWh outcompeting both Beijing and Guizhou in land and electricity cost. Access to cheap energy coupled with increased access to water in China’s north might be precedent to reorient data center expansion to China’s north as opposed to its west.
Conclusion and Summary of Findings
The rapid growth of data centers is paramount for the global AI revolution, however, the energy demands of these facilities pose significant challenges. China’s Eastern Data, Western Compute (EDWC) initiative aims to address these challenges by relocating data centers to western regions. While this strategy leverages lower energy costs and abundant resources, the success of the European model suggests that China’s path to AI dominance may lie in reorienting its strategy toward air-cooled data centers in its northern regions.
Less than 4% of district heating in China uses data center waste heat, compared to 8% in Sweden and 3.3% in Finland. Construction of air-cooled data centers in the north could replicate the success of facilities like Yandex’s Mäntsälä data center, which achieves a 31% energy reuse rate without heat pumps. The Hohhot data center, with a 65% waste heat recovery rate, shows the potential of liquid cooling systems, but air-cooled systems in colder climates can achieve similar results with increased scalability. Northern provinces like Ningxia offer lower energy and land costs than Guizhou, making them ideal for large-scale data center expansion. For China to solidify its position as a global AI leader, The National Development and Reform Commission (NDRC) should revise its EDWC strategy to focus on northern regions while prioritizing waste heat reuse and sustainable cooling technologies.
While the EDWC initiative represents progress, its focus on western regions overlooks the potential of northern China’s cold climate and district heating infrastructure. By implementing cloud computing infrastructure in the north, China can achieve superior energy efficiency and integrate waste heat into urban heating systems. This approach would not only solidify China’s position as a global AI leader but also set a new standard for sustainable data center development worldwide.
