Starting in 2026, a data center in Beijing that runs less efficiently than the city’s standard pays an electricity surcharge of 0.2 yuan per kilowatt-hour on the portion above the limit, rising to 0.5 yuan when it is more than double the limit. That surcharge, set out in Beijing’s 2024 to 2027 plan for existing data centers, applies to facilities whose power usage effectiveness (PUE) is above 1.35. Nationally, four agencies led by the National Energy Administration published an action plan on AI and energy on May 8, 2026 with 29 tasks, one of which makes the share of green electricity a reference metric when new computing facilities are planned. China’s plan is to make AI data centers comply on efficiency and clean power; the documents use money, monitoring and consolidation to get there, not closure orders.
PUE is the industry’s standard ratio of a facility’s total energy to the energy that reaches its computing equipment. Google, which publishes its own figure, explains that a PUE of 2.0 means one extra watt goes to cooling and power distribution for every watt of IT load, while a value near 1.0 means almost everything goes to computing.
What the May 2026 AI and energy plan asks of data centers
The Action Plan on Promoting Mutual Empowerment Between Artificial Intelligence and Energy (document Guonengfa Keji [2026] No. 34) was issued jointly by the National Energy Administration, the National Development and Reform Commission (NDRC), the Ministry of Industry and Information Technology (MIIT) and the National Data Administration. According to the full text published by Xinhua, it aims to have a secure, green and economical energy supply system for AI initially in place by 2027, and by 2030 to bring clean energy supply for AI computing facilities to a world-leading level. The data center tasks that matter most:
- Green power as a planning metric. The share of green electricity becomes an important reference indicator in planning computing facilities, and operators are supported in buying green certificates and green power to raise that share.
- Efficiency in project approval. For new and expanded computing facilities, the renewable energy plan, PUE, green power share and waste heat recovery become key items in the energy saving and carbon review.
- Location. Computing facilities are steered to gather in regions rich in wind and solar power.
- Backup power. Operators are encouraged to replace diesel backup generators with clean energy.
- Domestic chips. The plan calls for faster adaptation between Chinese AI chips and Chinese deep learning frameworks.
Pakistan’s The News, which reported the plan on May 9, 2026, read it as a push to move data center growth toward China’s northern and western provinces, where wind and solar are plentiful.
Where the “comply” pressure comes from
The efficiency numbers are older than the AI plan. In July 2024 the NDRC, MIIT, the National Energy Administration and the National Data Administration released the Special Action Plan for the Green and Low-Carbon Development of Data Centers. As Xinhua summarized it, its targets for the end of 2025 were:
- a national average PUE below 1.5 and overall utilization of at least 60%;
- a PUE within 1.25 for new and rebuilt large and very large data centers, and no higher than 1.2 at national hub nodes;
- green electricity above 80% for new data centers at the national hub nodes, with renewable energy use growing 10% a year.
The same plan calls for retrofitting low-efficiency data centers and for consolidating “old, small and scattered” ones. China Daily reported its 2030 goal as average PUE and per-unit computing efficiency at internationally advanced levels.
Beijing layered its own rules on top. Its computing infrastructure plan for 2024 to 2027, announced April 26, 2024, requires every existing data center to reach a PUE no higher than 1.35 by the end of the period. The November 2024 plan for existing data centers repeats that goal for 2027 and adds the teeth: centers above 1.35 become priority monitoring targets, and the surcharge starts in 2026 based on each center’s actual PUE in the previous year.
Does China force data centers to shut down?
Not in the texts we could read. None of them orders closures; the 2024 national plan speaks of retrofitting and consolidating old, small and scattered sites. Beijing’s plan for existing data centers instead encourages large, efficient operators to absorb low-efficiency ones and take over their workloads, and it pays for retrofits: energy saving projects can earn up to 1,200 yuan per ton of standard coal saved, other green projects up to 30% of eligible investment, capped at 30 million yuan per project, according to the Beijing government text. An inefficient facility is not switched off. It pays more for power until it is upgraded, merged into a bigger operator, or its owner decides the numbers no longer work.
What has happened since
On June 25, 2026, The News reported that the NDRC and the National Energy Administration had released a document supporting direct green power connections for data centers, dedicated links from wind and solar projects that bypass the public grid. That is the practical answer to the May plan’s location task: if computing moves to where the wind and sun are, it can be wired to them directly.
Where we land
For most US small businesses this is a distant story, and we will not pretend a Beijing surcharge moves your software bill next month; we found no published estimate of how it flows into AI prices. What it does show is that electricity, not only chips, is now a lever governments pull on the cost of AI. Chinese providers already price compute by the clock: DeepSeek, for one, bills peak and off-peak hours differently. That is one more reason to judge an AI tool by the work it saves your team at today’s price, the arithmetic in our guide to what AI software is worth paying for, rather than assuming compute only ever gets cheaper.
The rules also aim the pressure at engineering: better cooling, reused waste heat, cleaner backup power. That is efficiency won through better equipment and the people who run it, the kind we would rather see rewarded than efficiency measured in headcount.
The figure that would show how hard all this bites is one Beijing has not published, as far as we could find: how many data centers paid the surcharge on their 2025 numbers. When that count appears, it will tell us whether “comply” in practice meant retrofits or simply bigger power bills.
Questions about China’s AI data center rules
Does China force AI data centers to shut down?
Not in the official documents we reviewed. Beijing puts existing data centers with a PUE above 1.35 under priority monitoring, charges them an electricity surcharge from 2026, and encourages large efficient operators to absorb low-efficiency ones, but it does not order closures. (Source: Beijing municipal government)
What PUE do China’s data center rules require?
The 2024 national plan set end-2025 targets of an average PUE below 1.5, a PUE within 1.25 for new and rebuilt large and very large data centers, and no higher than 1.2 at national hub nodes. Beijing requires existing data centers to reach 1.35 or lower by 2027. (Sources: Xinhua, gov.cn)
What is China’s 2026 AI and energy action plan?
It is a plan published May 8, 2026 by the National Energy Administration, the NDRC, MIIT and the National Data Administration. Its 29 tasks include using green electricity share as a reference metric when planning computing facilities, with goals set for 2027 and 2030. (Source: National Data Administration)
