the AI+ energy action plan: how is it shifting policy?
nudging data-centre demand to flex: workload scheduling, storage, green-power contracts
steering computing towards renewables: siting hubs with wind, solar and dedicated energy
enlisting AI to run the grid: better forecasting, better dispatch
AI data centres want three things from the grid: power that never fails, is clean, and is low-cost. No grid can meet all three; Beijing’s AI+ energy Action Plan does not try to solve that trade-off. It asks data centres to choose, and to pay for their choice.
Until now, data centres have run flat out around the clock, drawing power whatever the price and whatever the state of the grid. The plan wants them to behave like large customers already do: use more when power is cheaper and clean, less when it is scarce and expensive. Centres that can shift their work around get cheaper, greener power. Those that do not must pay for steady supply. The Action Plan aims to price that trade-off rather than solving it. Whether the price signals arrive depends less on AI than on the unfinished business of power-market reform.
None of this is conceptually new. Time-of-use pricing and demand response for large industrial users have run in the US, the UK and elsewhere for decades, as cement plants and steel mills schedule grinding and rolling around cheap night-time rates. What is new is China building that market in a power sector that has been administratively priced for most of its history, doing so as AI demand grows quickly, and under a green mandate most of those earlier schemes never had to meet. New data centres at national hub nodes must already source 80 percent of their power from renewables, ratcheted up from 50 percent in 2021.
the impossible triangle
Launched nationwide in 2022, Eastern Data, Western Compute designated eight national computing hubs and ten data-centre clusters, using western China’s land and renewable energy to host electricity-intensive computing, while eastern hubs served demand requiring fast network response. Background processing, offline analysis and data storage could move west; latency-sensitive workloads stayed near users.
AI has strained this spatial fix. National computing-centre consumption reached 170 TWh in 2025, just 1.6 percent of national demand, but growing near 40 percent a year. Wang Hongzhi 王宏志 NEA (National Energy Administration) expects 800 TWh, around six percent, by 2030. This is double what CAICT (China Academy of Information and Communications Technology) modelled as its middle scenario only a year earlier.
The share remains small beside industry, but no other load is growing as fast, and AI clusters differ in kind from conventional cloud computing: they need continuous, high-quality electricity, and their loads run denser and more volatile. AI hardware draws far more power per rack than conventional servers, and training in particular causes sharp, synchronised swings in demand as thousands of chips ramp up and idle together. Western wind and solar are low-cost and clean but variable; moving computing west adds latency; keeping it east burdens already overstretched grids.
Yue Hao 岳昊 Jibei Electric Power Company calls the result an ‘impossible triangle’: data centres need electricity that is at once secure, green and low-cost, and improving one side undermines another. Each corner has teeth. A millisecond voltage dip can corrupt a training run’s model parameters; Yue cites an Amazon estimate that one minute of interruption costs US$5 million. Green power trades at a premium that data centres, whose overriding priority is lower cost, are reluctant to pay.
The triangle binds three parties with different interests (grid companies, computing operators and their customers), each pulling towards its own optimum.
load as a market asset
The plan works on three fronts: how data centre demand behaves, where the centres are built and how the grid plans for them.
First, demand is to become responsive rather than fixed. Electricity-market prices are to guide data-centre energy management and the scheduling of workloads across regions and grids; new facilities are pushed towards multi-year green-electricity contracts with renewable generators. At data-centre operator Zhongjin Data’s zero-carbon computing base in Ulanqab, Inner Mongolia, this already happens: AI training shifts to midday, when solar output peaks, and data backup to night, when wind is strong, releasing more than 30 percent of capacity as flexible load.
Other reforms do the work: green-power direct connections link renewable projects to data centres, storage sustains power quality and shifts demand, and the developing national power market could pay facilities for moving tasks out of expensive or constrained hours. The logic applies most readily to delay-tolerant training; latency-sensitive inference (running a trained model to answer real-world requests) remains hard to move.
Second, joint infrastructure planning. Large renewable-energy bases and national computing hubs are to be coordinated, with computing facilities and backbone-network access points concentrated in renewable-rich regions whose energy, water and land can support them. Pilots will pair million-kilowatt-scale AI computing facilities with dedicated energy systems, alongside a campaign to raise power-supply quality. This extends Eastern Data, Western Compute from a siting policy into joint planning: workloads that can be relocated move towards wind and solar, while separate measures shore up supply to facilities that must stay east.
Third rests on information: two-way empowerment 双向赋能, advertised in the policy’s name, enlists AI to lighten the load AI creates, feeding better forecasts to both siting decisions and demand response. The first 51 high-value ‘AI+ energy’ scenarios include renewable forecasting, grid dispatch, virtual-power-plant coordination, storage management and intelligent operation of direct connections. Better forecasts of renewable output and computing demand tell operators when to shift workloads, charge storage or draw from the grid.
Huang Zhen 黄振 Alibaba Cloud chief energy architect pushes the idea further: as AI agents become large consumers of compute, the grid need not schedule physical servers at all. A ‘translation’ layer could instead ask the agents themselves whether they can tolerate delay, turning incremental computing demand into a dispatchable resource by design.
Siting, contracts and forecasting each price one side of the triangle against another: siting prices latency against low-cost power; contracts and scheduling price greenness against firmness; forecasting narrows the risk premium on all three.
no price signals yet
Stripped of its market language, the plan is doing something more basic: getting power to where the AI is. The pricing sophistication remains an aspiration.
The plan can widen what gets traded, but it cannot set the price. That is fixed by geography, cost and timelines beyond its control.
Geography sets the first limit. Delay-tolerant training migrates west, but inference must stay near users; Wang Chaoyang 王朝阳 Alibaba Cloud global data centre sees deployment settling on ‘large western bases for training, with eastern inference computing arranged according to local conditions’. Eastern centres will continue to need local generation, grid capacity and green power access rather than being displaced by western hubs.
Reliability carries a price. Direct connections improve renewable access, but Wang Yongzhen 王永真 Beijing Institute of Technology notes their per-kWh cost still runs above subsidised grid power or power-purchase routes, and multi-user projects face unresolved questions over operating responsibility and cost allocation.
Even the Ulanqab flagship computing base shows the ceiling: its dedicated 300 MW wind-solar-storage station covers around 40 percent of the base’s designed 2.1 TWh annual consumption, with market-traded green power filling part of the gap and the public grid as the final backstop in its four-layer supply scheme. Variable output means data centres need storage, grid backup or other firm supply to cover the gaps. Covering those gaps costs money.
Timelines diverge in both directions. Wang Zesen 王泽森 State Grid Jibei Electric Power Research Institute notes a computing centre completes in around eight months while a supporting substation takes three to five years. But the opposite error is already visible: Yue Hao reports that north-western data centres reach only 30 percent of planned load by their fifth year, stranding supply built for demand that never arrived, and warns that high and low computing-demand scenarios differ by a factor of one to two. Overbuilding the grid wastes assets and raises costs as surely as underbuilding constrains AI.
incentives aren’t there either
The binding constraint, though, is commercial. Wang Chunhui 王春晖 Nanjing University of Posts and Telecommunications argues that storage, direct connections and market participation all raise upfront costs, while the revenues meant to reward flexibility (time-of-use tariffs, ancillary-service payments and green-certificate income) still await clear local rules.
Yue Hao is blunter, arguing most flexibility trials remain experiments because the sums do not work; the income from virtual power plants or peak-valley tariffs rarely covers the risk that flexibility poses to computing reliability; what moves operators today is ESG pressure. He draws a distinction the plan elides: allocating computing among users is a mature business, but adjusting it in time and space to suit the grid is a different and far harder one. Liu Zhi 刘志 Tsinghua Energy Internet Research Institute adds an obstacle no tariff reaches: inside the hyperscalers, energy teams lack the authority to redirect computing, and operations teams see no reason to accept risk for electricity savings that barely register against their revenues.
Until those rules exist, prices have no way to reach data centres: operators will not make workloads flexible where the price does not cover the risk. AI can sharpen forecasting and dispatch, but it cannot conjure firm capacity, waive latency requirements or compress construction timelines.
voices from the triangle
Yue Hao 岳昊 | Jibei Electric Power Company senior engineer
Yue reads the Action Plan as system-level management of the impossible triangle’s tensions: coordinated planning to place suitable workloads near renewable bases; dedicated energy systems, grid-forming storage and possibly direct nuclear or hydrogen supply for reliability; and, on the demand side, telling flexible training apart from latency-sensitive inference. His broader argument is that green electricity can be converted into higher-value digital output along a chain running from electricity to computing power, tokens and services, but only with supporting markets, standards and scheduling. The plan creates more ways to trade among its sides.
Yue is a senior engineer and senior expert at State Grid Jibei Electric Power’s Economic and Technical Research Institute. He has worked on NEA energy strategy and monitoring teams and researches grid planning and energy economics.
Wang Chaoyang 王朝阳 | Alibaba Cloud general manager for global data centres
Wang speaks as an operator. Deployment will follow ‘large western bases for training, with eastern inference computing arranged according to local conditions’: training chases lower-cost, cleaner western power; latency-sensitive services stay near users and commercial centres. His core worry is tempo. Computing demand grows exponentially and data centres build quickly, but generation, grids and transmission do not—and AI clusters impose loads volatile enough to stress local grids even where annual supply looks adequate. He also prices the stakes: siting a 1 GW data centre in a low-price region rather than a developed coastal one can save some C¥5bn (~US$740 million) a year in electricity. Yet low-cost power is not the only consideration; latency, token quality, carbon intensity and the value generated per unit of electricity all matter. His conclusion: computing and power must be integrated technologically, operationally, spatially, financially and through policy, not connected after facilities are built.
Wang is Alibaba Cloud’s global data centre general manager. He oversees data-centre deployment and has discussed how training, inference, electricity prices and grid construction timelines shape computing infrastructure.
Wang Chunhui 王春晖 | Ministry of Industry and IT Information Communication Economy Expert Committee member
Wang sees a five-part framework linking energy, computing power, application scenarios, data and models, whose central innovation is treating computing facilities as adjustable resources rather than mere consumers. His emphasis is on rollout. Storage, direct connections and market participation need capital; operators will defer workloads or provide grid services only if time-of-use prices, ancillary-service payments and green-certificate revenues cover the operational risk. Constraints run beyond electricity economics: energy data is commercially valuable but often belongs to critical-infrastructure operators, so opening it for AI training needs clear rules on classification, sharing and local-versus-cloud deployment, while engineers fluent in both AI and power systems remain scarce. The national framework is directionally sound but incomplete. Success rides on local pricing rules, replicable business models, data-governance standards and sustained cooperation among grid companies, data-centre operators and AI firms.
Wang is a professor at Nanjing University of Posts and Telecommunications and a member of the MIIT Information and Communications Economy Expert Committee. A long-time expert in economic law, he specialises in the digital economy, telecommunications regulation and information security.
context
26 Jun 2026: NEA announces over C¥20 tn in new energy system investments
15 Jun 2026: People’s Daily reports that the nationally unified electricity market system is ‘basically built’
4 Jun 2026: NEA director estimates 800 tWh of annual data centre electricity usage by 2030
27 May 2026: NEA releases China AI+ Energy Development Report
26 May 2026: NEA releases 51 high-value ‘AI+ energy’ scenarios
20 May 2026: multi-user green power direct connections introduced; computing facilities priority
9 May 2026: AI+ energy two-way empowerment action plan released
28 Apr 2026: Politburo calls for accelerated planning and construction of ‘six infrastructure networks’
14 Mar 2026: 15th five-year plan calls for coordinated green-power and computing layouts and stronger nationwide computing-resource dispatch
10 Mar 2026: annual Government Work Report includes ‘six networks’ including power grids and computing networks as priority areas
11 Feb 2026: State Council releases new opinions on unified power market construction
30 Jan 2026: NDRC plan calls for improving power capacity pricing mechanisms
24 Jun 2025: NEA announces national power market to be initially established within 2025
11 Dec 2025: State Grid issuesAI integration plan
14 Sep 2025: renewable energy guaranteed price floor released in Shandong
10 Feb 2025: new pricing policy for new energy released
13 Nov 2024: Energy Law marks PRC’s first comprehensive energy legislation
25 Oct 2024: China Electricity Council calls for public comments on blueprint for a national power market
15 Oct 2024: inter-provincial power spot market officially launched
23 Aug 2024: city computing network announced
20 May 2024: State Council research reaffirms new energy as the engine of China’s new economy
25 Aug 2023: first national guidance for power market regulator released
18 Aug 2023: Central Commission for Comprehensively Deepening Reforms new power system construction acceleration plan
2 Jun 2023: final blue paper on new power released
12 Feb 2019: MIIT calls for ramped up green data centre efforts