A data center hums day and night, and somewhere a meter keeps spinning. China now wants that hum to cost far less, and it is writing the rules to make it happen. The fight is no longer about whether AI should grow — it is about where the machines live and how much the lights cost.
A plan with a number on the door
In July 2024, four central agencies — the National Development and Reform Commission (NDRC, 国家发展改革委), the Ministry of Industry and Information Technology (MIIT, 工业和信息化部), the National Energy Administration (NEA, 国家能源局) and the National Data Administration (国家数据局) — jointly issued the Special Action Plan for Green and Low-Carbon Development of Data Centers (《数据中心绿色低碳发展专项行动计划》, document no. 发改环资〔2024〕970号). It took effect the day it was published, on July 3, 2024.
The headline is a single efficiency figure: PUE, or power usage effectiveness (电能利用效率), the ratio of total facility power to the power that actually reaches the servers. The lower, the better. A PUE of 1.5 means one third of the electricity is spent on cooling, conversion losses and overhead rather than computing.
The targets, in plain terms
The plan sets a ladder of deadlines:
- By the end of 2025, the national average data-center PUE should fall below 1.5.
- New and expanded large or extra-large data centers must hit PUE ≤ 1.25 by end-2025.
- Inside the national hub nodes, new projects are capped even tighter at PUE ≤ 1.2.
- By the end of 2030, average PUE, per-unit-compute energy efficiency and carbon efficiency should reach "internationally advanced" levels.
- By end-2025, overall rack utilization (上架率) should be at least 60%, and national hub nodes should host more than 60% of all new compute added nationwide.
- New data centers in the hub nodes should run on more than 80% green electricity by end-2025.
These are not soft aspirations. The plan tells provincial energy regulators to diagnose inefficient facilities, and it signals tiered electricity pricing (阶梯电价) for data centers that fall below the efficiency line — a direct cost penalty rather than a slap on the wrist.
Why it matters for AI, specifically
The document names artificial intelligence outright. It calls for "strengthening the major productive-capacity layout for AI" and guiding intelligent computing centers (智算中心) toward standardized, clustered development. Translation: the explosion of GPU clusters behind large model (大模型) training and inference is exactly what this rule is built to contain.
Two background numbers explain the urgency. According to the China Academy of Information and Communications Technology (中国信通院, CAICT), national data-center electricity use is projected to pass 380 billion kWh by 2030. MIIT has reported that by the end of 2023, China's in-use data-center racks already exceeded 8.1 million standard racks, growing more than 20% year on year. Every new large model (大模型) training run stacks more racks onto that curve.
The "send it west" mechanism
The plan hardens the "East Data West Computing" (东数西算) constraint. New large and extra-large data centers should sit inside the national integrated computing-network hub clusters. Non-real-time workloads are explicitly encouraged to migrate to the western hubs, where renewable power and land are cheaper. Cities that already have a data center running below 50% utilization for over a year are told, in principle, not to approve another big build.
This is the quiet geography lesson inside the energy rule: latency-sensitive AI inference stays near users on the coast, while the energy-hungry training and batch jobs move inland to provinces like Guizhou, Gansu and Ningxia.
What actually changes on the ground
Six task areas frame the rollout:
- Layout discipline — hub clusters first, scattered builds consolidated.
- Entry controls — stricter energy reviews for any new project; only efficient servers allowed.
- Retrofit — "old, small, scattered" data centers merged or upgraded.
- Renewables — green-power share baked into the approval, with pilots for direct green-power supply.
- Resource recycling — waste heat recovery, regenerated water, equipment reuse.
- Cooling tech — liquid cooling, evaporative cooling and AI-driven operations promoted.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the plan turns energy from a background accounting line into a front-door gate for any new AI compute project — builders now design the power bill before they design the cluster.
The honest tension
The rule pushes efficiency without capping total AI growth, and that is deliberate. China wants both more compute and a smaller footprint, which only works if renewable supply and cooling technology improve fast enough. There is also a real risk that tight hub-node PUE caps slow local AI ambition in regions that cannot hit 1.2, concentrating capability in a few corridors.
Honest limitations
This article relies on the official plan text (gov.cn), Xinhua and People's Daily reporting, and CAICT/MIIT figures cited in trade coverage. It does not independently audit whether provinces have met the 2025 PUE or 80% green-power targets — those are forward commitments, and 2025 year-end compliance data was not re-verified here. The "380 billion kWh by 2030" figure is a CAICT model projection, not a measured outcome, and is sensitive to large model (大模型) adoption speed. Provincial implementation rules and tiered-pricing details vary and were not surveyed province by province. The piece covers energy and siting policy only; it does not address AI chip supply, water use beyond a brief mention, or carbon-offset accounting.
What readers can do now
- If you build or buy AI compute in China, price the PUE 1.25 / hub-node 1.2 ceiling and the 80% green-power expectation into any 2025+ site plan — they are approval thresholds, not suggestions.
- Track the "East Data West Computing" (东数西算) hub clusters as the default location for training and batch workloads; coastal sites will favor low-latency inference.
- For global comparison, watch China's national-average PUE as a public metric of how fast liquid cooling and green-power procurement are actually scaling — it is one of the few public, recurring energy numbers tied directly to AI infrastructure.