CoverStory:China’sAIBoomIsRewiringItsPowerGrid

When Chinese markets reopened after the May Day holiday, state-owned utility Datang International Power Generation Co. Ltd. embarked on a dizzying stock rally.

After hitting the daily trading limit six times and steadily climbing, its shares reached a record high of 9.92 yuan ($1.5) on June 3. In a single month, Datang International’s stock surged more than 130%.

The sudden frenzy prompted the company to issue eight separate risk advisories, urging investors not to be swayed by “overheated market sentiment.”

What caused the buying spree was a buzzy new policy concept: “computing and power synergy.” On May 2, Datang launched a 500-megawatt solar plant in a cloud computing hub in Zhongwei, Ningxia. Markets interpreted this as China’s first large-scale green-energy direct-supply project for computing — a milestone connecting desert solar resources directly to digital infrastructure.

In fact, Datang International was only an investor on the power-generation side and has no role in operating the computing center. In its risk advisories, the company said it had no computing-power-synergy project in actual operation.

Yet the market’s overreaction reflects a profound realization that the relentless evolution of artificial intelligence has led to an explosive demand for computing power, elevating the strategic importance of energy infrastructure. Across China, markets are hunting for equities tied to this new reality.

In March 2026, the concept of coordinating computing power and power generation was officially enshrined in the government work report presented to the National People’s Congress. China’s cabinet called for the building of ultra-large intelligent computing clusters, the coordination of power infrastructure, the enhancement of national monitoring of computing resources and the forging of a new intelligent economy.

On May 8, four top government bodies — including the National Development and Reform Commission and the National Energy Administration — issued a joint action plan to foster a two-way synergy between AI and energy. Its mandate set a clear timeline: establish a secure, green and economical energy guarantee system for AI by 2027 and achieve significant two-way empowerment between AI and the energy sector by 2030.

For the public, the main concern is whether the country will run out of electricity to power these fast-growing computing systems. Industry insiders face a more complex reality, however. The bottleneck is not necessarily the aggregate volume of electricity generated nationwide, but whether the grid can deliver stable, instantaneous power to specific regions at specific times.

“The industry’s discussion of computing growth often focuses on energy volume, but once it enters grid operation, the core issue is instantaneous power,” said Wang Zesen, deputy director of the Power System Institute at the Electric Power Science Research Institute under State Grid Jibei Electric Power Co. Ltd.

Energy volume refers to the amount of electricity consumed over time, while instantaneous power is the load required at a given second. Keeping that balance is a hard constraint for grid operators — one that the AI boom will test constantly.

The new AI variable

From coal and oil in the Industrial Age to electricity in the AI era, the critical resources driving productivity are changing dramatically. For the power system, if the biggest variable of the past decade has been the rise of renewable energy, the next decade belongs to AI.

Data usage is expanding at a staggering pace. By March 2026, China’s daily AI Token usage surpassed 140 trillion, a more than 1,000-fold increase from early 2024, according to Liu Liehong, head of the National Data Administration. As user interactions move from basic chatbots to complex AI agents requiring multi-agent coordination and long-context processing, token consumption is exhibiting explosive growth.

This translates directly into demand for power. Between 2019 and 2025, the amount of electricity consumed by computing in China rose from 82.4 billion kilowatt-hours to 196 billion kilowatt-hours, increasing its share of total power consumption from 1.3% to 1.9%. The China Academy of Information and Communications Technology anticipates that this figure will hit 500 billion kilowatt-hours by 2030. If AI growth is explosive, it could reach 700 billion kilowatt-hours, up to 5.3% of the country’s total demand.

Meanwhile, China’s power supply is moving from coal to high proportions of renewable energy. Wind and solar power accounted for nearly 48% of its total installed capacity by March of this year. Because renewables are intermittent and volatile, combining them with the massive, fluctuating loads AI data centers require makes systemic coordination both essential and challenging.

To bridge this gap, China is leaning into its “Eastern Data, Western Computing” strategy. Launched in 2022, the initiative is intended to guide power-hungry data centers away from the densely populated, land-scarce East toward the resource-rich West, establishing eight national computing hubs. These new data centers must source more than 80% of their power from green energy.

A Huawei Cloud data center in Ulanqab, Inner Mongolia, in Oct. 2020. The city aims to transform its growing computing cluster into China’s “Token Capital” by 2029. Photo: VCG

Wang Yongzhen, an associate professor at the Beijing Institute of Technology, said that this macro layout gives China a unique foundation. While China still needs to close the gap in advanced AI chip manufacturing and software ecosystems, it can compensate for its computing shortcomings by powering its data centers with a vast, cheap supply of green electricity.

Fears that AI will cause widespread power shortages in China are overblown, according to Gao Xing, chief utilities analyst at China Securities Co. Ltd. China’s total power consumption exceeded 10 trillion kilowatt-hours in 2025, growing about 5% each year. Even if AI adds 50 billion to 100 billion kilowatt-hours a year, that is still only around 1% of the total load.

Gao contrasts this with the United States, where electricity demand had been flat for two decades, leading to underinvestment in power infrastructure. When AI suddenly makes massive new demands, a stagnant U.S. grid faces severe localized supply stress. China’s advantage lies in its continuous, growth-oriented grid investments.

The real threat to China is not a macro deficit, but the repercussions of a fundamental shift in how its grid operates.

Power dictates computing

On the northern bank of the Yellow River, bordering the Tengger Desert, data centers in a large cloud computing park in Zhongwei are the thirstiest power users in the sands.

Blessed with abundant sunlight, strong winds and a refreshing climate ideal for server cooling, Zhongwei has attracted six of China’s top ten computing enterprises. By 2025, more than 65% of the park’s electricity came from green sources.

In addition to a 500-megawatt solar farm, Datang Group plans to launch a 1.5-gigawatt wind farm there by late 2026. The integrated project will cut electricity costs to 0.36 yuan per kilowatt-hour, about 20% lower than the local average.

Zhongwei exemplifies an emerging model — power dictates computing. Instead of building data centers and then searching for power, planners are using power availability as the primary determinant for where computing facilities should be located. The capacity of local grids to absorb renewables and provide storage dictates the scale of new AI hubs.

Another flagship project is Ulanqab in Inner Mongolia. Driven by the regional Mengxi power grid’s flexible pricing and a 67% local green power ratio, Centrin Data launched a low-carbon computing base there in July 2025. It employs a “source-grid-load-storage” model — an integration of wind, solar and battery storage alongside the data center — generating 848 million kilowatt-hours of self-used green power each year.

Crucially, Ulanqab sits only 300 kilometers from Beijing. Direct, high-capacity fiber-optic cables link the desert data center to the capital with a latency of just 4.2 milliseconds. This real-time speed allows the western hub to process eastern demands for autonomous driving, cloud rendering and financial risk control. Ulanqab has already contracted more than 5 million standard server racks and aims to brand itself China’s Token Capital within three years.

Tech giants are also experimenting with such synergy. In June 2025, Tencent deployed a small-scale computing experiment in a zero-carbon industrial park in Chifeng, Inner Mongolia. Operating entirely on green power, the facility explores which computing tasks can be seamlessly ported to take advantage of the park’s 40% cheaper electricity. Tencent aims to run its entire data center portfolio on 100% green energy by 2030.

Strains on the grid

Despite these successes, the AI boom is placing unprecedented localized strain on traditional infrastructure.

In Datong, a northern city historically known as China’s Coal Capital, the industrial landscape is fracturing. In 2025, the city’s computing industry consumed more than 6 billion kilowatt-hours of electricity — a roughly 40% year-on-year jump that officially surpassed the power use of its traditional coal sector.

The computing sector now accounts for 26.2% of the local grid’s load. Similar surges are visible in Zhangjiakou, where data centers now consume approaching 30% of the city’s power, and in the Gui’an New Area in southwestern China’s Guizhou province, where local power authorities had to build a 500-kilovolt substation simply to double the area’s grid capacity.

The challenge lies in the physical constraints of the grid. While AI workloads seem smooth on a macro level, at the micro level of chips and servers, they are volatile.

Visitors photograph a China Telecom data center in Gui’an, Guizhou province, in April 2025. The southwestern region is one of China’s largest computing clusters. Photo: VCG

Inference workloads — the computing required to answer real-time user prompts — peak during the day alongside general commercial electricity use, creating a “peak upon a peak” effect for the grid. Furthermore, AI servers consist entirely of power-electronic devices. Their power demand can fluctuate wildly within 0.5 to 1 millisecond. A gigawatt-scale computing center could see instantaneous fluctuations of 10,000 kilowatts.

This creates a perilous “dual-high” scenario in regions such as northern Hebei province. There, renewables — which access the grid via power-electronic inverters — account for more than 81% of installed capacity, while data center hubs simultaneously draw massive power-electronic loads. When power generation and power consumption are both high volatility, the resilience of the local grid degrades.

The risks are not theoretical. In July 2024, a thunderstorm in northern Virginia in the U.S. triggered a minor voltage disturbance. Protective mechanisms kicked in, causing around 60 data centers to instantly switch to backup power and drop off the grid. Some 1,500 megawatts of load vanished in seconds, dangerously spiking the grid’s frequency. A similar event occurred there in early 2025. These incidents serve as stark warnings to Chinese regulators that AI load management is fundamentally a grid-security issue.

Exacerbating this is a severe mismatch in infrastructure timelines. A large computing center can be built in 8 to 24 months, whereas the planning, approval and building of a traditional power substation takes three to five years. Furthermore, actual computing workloads are hard to predict. Some centers hit 80% utilization immediately, while others languish at 10% to 20% for years. Grid operators, lacking a clear view of the opaque operations in private data centers, struggle to plan future capacity accurately, risking either wasted utility investments or stranded computing assets.

The synergy struggles

To mitigate these risks, China is exploring sophisticated virtual power plants (VPPs) and dynamic energy markets.

In December 2025, the State Grid’s Shanghai branch and China Telecom executed a historic test of rapid, cross-provincial computing transfer. Facing a simulated local power peak, a VPP platform in Shanghai identified low-latency inference tasks — such as language models and video recognition algorithms — running on 104 AI processors. Within three minutes, the tasks were seamlessly migrated to a computing node in Fujian province, shedding 50 kilowatts of local load without disrupting business operations.

Similarly, in May 2026, Guangdong province integrated three large telecom data centers into its electricity spot market via a VPP. When local electricity prices drop, the VPP tells the data centers to increase non-urgent computing tasks; when prices spike, they dial back. As a China Mobile expert said, this turns every watt of data center flexibility into tangible financial savings.

True “computing and power synergy” remains in its infancy, however. Moving data is vastly more complicated than moving electricity. Water and electricity flow uniformly in one direction, but computing tasks require a two-way flow of data with zero tolerance for latency.

“The datasets required for actual model training are incredibly large. If we relied entirely on networks to transmit from east to west, it might take months,” an executive at a top computing service provider said. “Companies sometimes have to physically transport hard drives.” For inference tasks, which require instant responses to consumers, moving operations out of the densely populated East is nearly impossible.

Network transmission costs also threaten economic logic. If a company saves one dollar on western electricity but spends two dollars on cross-country bandwidth, the model collapses.

The path forward requires sweeping structural reforms to unify China’s fragmented power, computing, telecommunications and carbon markets. In the long term, tech giants are already looking beyond traditional grids. Recognizing AI’s insatiable thirst for stable, clean baseload power, companies such as Alibaba Group Holding Ltd. and Tencent are actively investing in next-generation nuclear technology, acquiring stakes in small modular reactor projects and nuclear fusion startups.

“Currently, the coordination of computing and power is still in its early stages,” said Wang Yongzhen at Beijing Institute of Technology, who expects the experimental phase to last another three years.

“But the ultimate goal is clear: to integrate security, green energy and economic efficiency.” For China, mastering this balance is no longer just an infrastructure upgrade — it is the prerequisite for winning the global AI race.

Contact reporter Han Wei (weihan@caixin.com)