NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

China's Seven Central Bodies Want AI Firms to Disclose How Much Energy Models Use

A jointly issued national plan links AI growth to energy efficiency, asking generative-AI providers to assess and disclose resource use and pushing data centres toward greener power — a stated direction, but with the details still to be written.

2026-11-20 · 1099 words · NeuroAI
China's Seven Central Bodies Want AI Firms to Disclose How Much Energy Models Use

A data centre hums through the night, its racks drawing power that grid planners must match. Somewhere inside, a large model (大模型) runs a training job whose electricity cost almost nobody outside the company can see. A new national plan wants both halves of that picture brought into the open.

Released in September 2026, the document does not ban anything or impose a fine. It does something quieter and arguably more durable: it makes energy efficiency a stated priority of China's AI build-out, and puts named ministries behind the task.

What the Plan Is — and Is Not

The file is the Implementation Plan for Coordinated Digital and Green Transformation (促进数字化绿色化协同转型发展实施方案(2026—2030年)), jointly issued by seven bodies: the Cyberspace Administration of China, the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Ecology and Environment, the Ministry of Housing and Urban-Rural Development, the Ministry of Agriculture and Rural Affairs, and the Ministry of Commerce.

It matters to be precise about its status. This is a formally issued plan, published through official channels and reported by the state news agency — not a draft and not a consultation paper. Comments and drafts that may have circulated earlier are not what is described here. At the same time, it is a guiding plan aimed at 2030, not a regulation with an immediate penalty clause. Treat it as a settled direction with implementation details still to be filled in.

The structure is four areas and fourteen tasks. In plain terms: fuse digital and green technologies, cut energy use in new sectors such as computing and telecoms, apply the approach across agriculture, manufacturing, trade, consumption and city operations, and build the supporting foundation.

The AI-Specific Clauses

What makes this plan relevant to AI readers is how directly it names the industry.

  • Green and efficient AI. The plan calls for research into storage-compute integration, data compression, distributed training frameworks and database-for-AI techniques, plus faster interconnects, high-bandwidth memory and low-power computing. It points to mixture-of-experts and dynamic sparsity as architecture directions and to quantisation and pruning as ways to shrink models.
  • Resource disclosure. It proposes a lifecycle resource-consumption assessment system for generative AI, guiding providers to disclose how their large models use resources. In the source text this is framed as guidance and encouragement, not a mandatory reporting rule. The details of what gets disclosed, to whom and how often were not in the material reviewed.
  • Greener data centres. The plan mentions liquid cooling, waste-heat recovery and low-power chips, explores 800-volt high-voltage direct-current power architecture and supports ultra-high-density cabinets above one hundred kilowatts. It also pushes operators toward renewable power, green-certificate trading and siting compute near wind, solar and hydro resources.
  • AI for carbon accounting. It proposes exploring an AI model for the national carbon market, building databases of greenhouse-gas and product carbon-footprint factors, and improving carbon-footprint labelling and disclosure.

The headline target, repeated across official summaries, is that by 2030 the low-cost, high-efficiency advantages of AI should be more visible and should support green and low-carbon innovation. For compute facilities and 5G base stations, the plan says renewable electricity consumption should reach the share required by each province's renewable-energy consumption responsibility weight.

Why This Goes Beyond Data Centres

Energy use by AI has usually been discussed as a private engineering problem: a company wants cheaper training and better cooling. This plan reframes it as a public planning question, linking where compute is built to where clean power is available, and linking model efficiency to national carbon goals.

That shift has consequences. If disclosure becomes routine, buyers and regulators gain a way to compare models on something other than benchmark scores. If compute is steered toward regions rich in renewables, the geography of AI infrastructure changes. And if carbon-accounting models are built into the national system, AI becomes both the object of regulation and a tool of it.

None of that happens automatically. The plan sets direction and asks ministries to coordinate. Its effect will depend on the standards, measurement methods and enforcement rules that follow — the unglamorous documents that turn a policy statement into something a company must actually do.

Reading the Fine Print Honestly

Three cautions are in order. First, a plan that runs to 2030 sets a horizon, not a deadline for next quarter; progress should be measured over years. Second, several of the technical items are research directions rather than requirements, and "explore" and "guide" are not the same as "must." Third, a target tied to provincial renewable-consumption weights leaves room for wide regional variation.

For readers outside China, the plan is also a signal about how the country is choosing to compete: not only by building more compute, but by arguing that efficient compute is the competitive advantage.

Key takeaways

  • The plan was jointly issued by seven central bodies and covers the period to 2030; it is a formally released guiding plan, not a draft or consultation document.
  • It sets fourteen tasks across four areas, including green and efficient AI and lower energy use in computing and telecoms.
  • It proposes a lifecycle resource-consumption assessment system for generative AI and encourages providers to disclose how large models use resources.
  • Data-centre measures include liquid cooling, waste-heat recovery, 800-volt DC power and a push toward renewable electricity.
  • By 2030, compute facilities and 5G base stations are expected to meet provincial renewable-energy consumption responsibility weights.

Honest limitations

This article is based on the official summary released through the Chinese government website and on reports from financial and provincial government outlets that quote the plan's text. I have not read the full fourteen-task document, only the official summaries and the clauses those sources reproduce. Where a measure is described as "guidance" or "exploration," I have kept that wording rather than upgrading it to a requirement.

The resource-disclosure provisions come from a report quoting the plan via an official party channel; I could not verify implementation timetables, reporting templates or any compliance mechanism in the sources reviewed, so none are asserted here.

No currency figures appear in this story, so no RMB-to-US$/HK$ conversion applies. The article describes the plan's stated aims; it is not an assessment of whether those aims will be met.

What readers can do now

  • Watch for the standards, not the headline. The measurement and disclosure rules that follow this plan will matter more than the plan's broad language.
  • If you build with AI, track efficiency metrics. Model cost per task and energy per task are likely to become part of procurement conversations.
  • Read "guide" as "coming." Check whether implementing rules appear; guidance alone imposes no obligation.

Related coverage

Cite this article

NeuroAI (脑机网), "China's Seven Central Bodies Want AI Firms to Disclose How Much Energy Models Use," 20 November 2026, https://neuroai.site/a/na-policy-digital-green-ai-plan

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Policy & Governance” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories