Is China’s energy system ready for the AI boom?

In early June, I travelled to Guiyang in southwestern China’s Guizhou province to visit China Mobile’s Gui’an Data Centre, which has been designated a “national green data centre”. The facility is located in a cluster of modern buildings in the Gui’an New Area. Since 2021, this cluster has become a hub of China’s internet infrastructure, hosting data centres operated by major tech companies.

Engineer Jian Chonghai walked me through the centre’s energy-saving innovations – the modular cooling cabinets, or “maglev” air-conditioning system. Unlike traditional AC, maglev air conditioners eliminate friction in compressors, achieving the same cooling effect with 30-40% less electricity, he said.

Jian’s focus on energy efficiency is driven by an ambitious expansion plan. As product manager Li Haiyan told me, training and running AI models is set to become a part of their business – potentially bringing much higher electricity consumption, carbon emissons and operating costs.

China Mobile isn’t alone. From big tech companies like Alibaba, Tencent and Huawei to startups like DeepSeek, Chinese companies are locked in fierce competition to develop AI services.

This represents a new challenge for China, particualrly as experts believe AI will become one of the country’s most energy-intensive industries. The central challenge for China is how to become the global leader in AI services while not jeopardising its climate action targets.    

While efficiency improvements and AI’s potential to help other sectors decarbonise faster offer some hope, building a new power system that prioritises renewables and meets computing needs has become crucial.

Growing appetite

The electricity use of China’s data centres is expected to increase by 170% between 2024 and 2030, according to the International Energy Agency (IEA).

Last year, data centres around the world consumed about 1.5% of total electricity generated, a recent IEA report estimates. This share is rapidly growing. Mainly driven by AI-specific servers, electricity use in data centres is rising at about 12% per year. That is four times the pace of overall electricity demand.

The report predicts that by 2030, China and the US will account for nearly 80% of global data centre energy growth.

Recommended

Wang Yongzhen, an associate professor at the Beijing Institute of Technology, told me that he considers data centres to be one of China’s major energy-intensive industries – alongside steel, cement and petrochemicals.

He estimates that by 2030, Chinese data centres will demand around 105 gigawatts of electricity, use 26.3 billion litres of water, and emit 310 million metric tonnes of CO2.

This amount of electricity is more than half of China’s residential electricity demand in 2024.

East Data, West Computing

In 2022, China launched an initiative called East Data, West Computing. Under the plan, western provinces and regions like Guizhou, Inner Mongolia, Gansu and Ningxia are tasked with handling computing jobs such as AI training and data storage – workloads that don’t require real-time response.

Meanwhile, clusters in regions like Beijing-Tianjin-Hebei, the Yangtze River Delta, Chengdu-Chongqing and the Greater Bay Area are focused on real-time services such as video streaming and AI chatbot use.

One major goal of this initiative is to reduce energy consumption by leveraging western China’s favourable climate and abundant renewable energy. For example, Guiyang’s average annual temperature of 18-20C naturally lowers cooling needs, while Inner Mongolia offers rich wind and solar resources.

By the end of this year, newly built data centres in national hub nodes are expected to run on more than 80% renewable power, according to the plan. Regions with strong solar, wind or hydropower resources will build low-carbon data centres handling computing-intensive but less time-sensitive work.

Wang Yongzhen notes that green computing power not only aligns with national strategies but also brings tangible economic benefits to companies. Making data centres more energy efficient means smaller electricty bills, he said.

When AI meets the grid

One important piece of the puzzle is the National Integrated Computing Network, which the government is developing to bring together public and private cloud computing resources into a single platform. It was highlighted in a recent report by research organisation RAND, titled “China’s Evolving Industrial Policy for AI”.

Kyle Chan of Princeton University, lead author of the report, likens the network to a “public utility” model for AI compute resources – an approach that echoes China’s typical infrastructure development strategies aimed at reducing regional inequalities.

The network aligns with another major initiative to build a “clean and efficient, flexible and intelligent” power system. From 2024 to 2027, China plans to build this “new power system” that will rely on diverse renewable sources and smart technologies like AI.

Under these two initiatives, electricity and computing are seen as public services – which means they should be easy to access and low-cost. But to hit the country’s climate targets, they also need to be clean and green. The big question is: how can they be made both affordable and sustainable at the same time?

Wang Yongzhen envisions that computing power and electricity will create a synergy. “One of the key goals is to increase the share of green electricity used in data centres. The second is to reduce energy consumption – not just with isolated tech upgrades, but through improved system-wide efficiency. The third is enabling data centres to interact with the power grid.”

Recommended

To put it simply, when areas face tight power supplies, computing tasks can shift to data centres that meet speed requirements but cost less, using electricity spot market prices as guidance. This approach boosts AI reliability for customers while easing local power crunches.

“When confronted with intermittent wind and solar power supplies, data centers can also act as buffers by ‘switching on’ certain computing tasks like AI training during periods of oversupply.”

Wang also noted that data centers can be coordinated to allow the grid to tap into their energy storage systems during emergencies. This has much in common with the vehicle-to-grid model which sees electric vehicles give power back to the grid during peaks in demand.

Still, challenges remain.

First of all, the synergy between computing and power networks that Wang described requires extensive coordination – across different parties, government agencies, and even at the staff level. It also needs to be tested across various scenarios over a long period.

Wang Yongzhen said he noticed a communication gap among staff members at data centres, between IT engineers like Li Haiyan and operation staff like Jian Chonghai. This gap could become a major bottleneck in the future. “They need to stay on top of both computing and energy systems if they want to make the vision of this highly coordinated network a reality.”

Another challenge is the market design.

The China Academy of Information and Communications Technology notes that the Green Electricity Certificates system, which tracks and verifies renewable power generation, still lags behind market demand. In 2023, only 1.5 million megawatt-hours (MWh) of the 3.8 million MWh green power that was traded involved the certificates.

This could hinder data centres who want to purchase more green electricity or track the share of green power in their own energy use.

AI marches on

A 2017 State Council document set a goal for China to become a world leader in AI theory, technology and application by 2030. It shows that AI development in China will not slow down despite pressures on energy, climate goals and from other countries such as the US.

Kyle Chan notes that for China, the goal isn’t just to “win the race” with the US, but to build a “resilient AI industry that will boost productivity across sectors – from manufacturing and healthcare to education and governance”.

Wang Yongzhen says that while advances in AI are driving up power demand from data centres, increased computing power is also helping to support decarbonisation efforts in other sectors. Therefore, when considering AI’s energy consumption in the broader context of societal decarbonisation, it’s overly simplistic to equate rising energy use with a setback to climate goals.

Recommended

One of the biggest hurdles for efforts to build a national carbon-footprint management system is data sensitivity, Xu Ming, a professor at Tsinghua University, told me in a podcast. Full carbon accounting requires tracking everything from suppliers to delivery. Many companies fear exposing trade secrets.

Acting as a neutral data steward, an AI-powered agent could securely store emissions data and grant access to regulators while protecting confidentiality. In this way, AI becomes a critical tool in slashing industrial emissions.

But these ideas are still in the experimental stage, and it will take time to see concrete results on how – or if – AI can actually limit or reduce emissions.

Back at the Gui’an data centre, Li Haiyan likened mainstream AI models to undergraduates. In the future, engineers will train more specialised, domain-focused AI models – akin to postgraduates, PhDs and postdocs – that can deliver more targeted and practical solutions. She believes this shift will also reduce the demand for chips used in both training and operating AI models, and this will likely reduce demand for power to run and cool them.