At a dispatch center in Changsha, a transfer decision that used to take a human half an hour now appears in about sixty seconds. No dramatic alarm, no heroics — just a recommended plan on a screen, generated by a model nicknamed 光明 (Guangming, "Light").
This is the quiet center of China's AI-in-energy story. Not a flashy robot, but a model sitting inside the control room of the world's largest power network, helping decide how electricity flows.
A trillion-parameter model for electrons
State Grid Corporation of China (国家电网) operates the grid serving the most customers and the most renewable-connected capacity of any utility on earth. In 2025 it released 光明电力大模型 (Guangming Power Large Model), described as the country's first trillion-parameter (千亿级) multimodal large model for the power sector. It passed dual evaluations by the China Academy of Information and Communications Technology and the China Electronics Standardization Institute at the top "卓越级" (Excellent) grade.
The deployment is broad, not a toy. Under a "6541" framework — six business domains, five core capabilities, four-level application, one model foundation — it spans planning, grid operation, equipment management, work control, customer service, and operations, with 600-plus scenarios. As of 2025 it covered headquarters and all 27 provincial power companies, and State Grid said it had built 100-plus typical scenarios, 1,700-plus large and small models, with daily model-service calls above 10 billion.
From thirty minutes to one
The clearest before-and-after is dispatch. At State Grid Hunan's Changsha control center, operators describe the old world: a single load-transfer decision could take 30 minutes or longer. With 光明, the recommended plan arrives in about a minute. The model also cut the assessment time for key HVDC equipment status from a week to minutes, and lifted smart power-supply-plan compilation efficiency more than sevenfold.
Elsewhere the gains are about prediction. At State Grid Hangzhou, the 光擎 (Guangming-engine) "smart brain" covers more than 1,800 bus nodes and 1,400-plus 35 kV-and-above lines, and pushed new-energy ultra-short-term (15-minute) forecast accuracy to 98.92%. When the wind and sun decide supply, that precision is the difference between a stable grid and a curtailment headache.
The award and the virtual power plant
At WAIC 2025, State Grid's "'AI + Guangming' builds a megacity smart energy management master" project won a SAIL Star (SAIL之星) — the only energy-sector project among the honorees. The work behind it, run by State Grid Shanghai, uses Guangming's reasoning to coordinate a city-level virtual power plant: shifting user-side load, tracing carbon flows, and responding in seconds. State Grid says Shanghai's virtual power plant delivered cumulative economic benefit above 80 billion yuan (≈ US$11.3 billion / HK$88 billion) and cut about 510,000 tons of carbon a year.
The same logic reaches into green certificates, electric-vehicle charging, and "source-grid-load-storage" coordination. State Grid Qinghai built the country's first clean-energy and green-compute dispatch center, lifting green-power consumption for compute centers by about 10%.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the grid is where AI must be right every time — "a model that can hallucinate is a toy; a model that keeps the lights on is infrastructure."
AI and the grid, in both directions
The relationship cuts both ways. The grid feeds AI: State Grid runs about 500 million smart meters (约5亿只电表) producing precise, high-volume data that fixes the thin-sample problem generic models suffer. And AI feeds the grid: as data-center loads swing two to three times normal, source-grid-load-storage coordination keeps the system stable. In other words, the AI industry and the power industry are now each other's constraint and each other's cure.
Honest limitations
Figures are from State Grid releases, Xinhua Energy, and China Energy News; they are company- and state-media disclosed, not independently audited. "10 billion daily calls" and "600-plus scenarios" are aggregate deployment claims without public per-scenario verification. Power models must meet a near-zero-tolerance bar, so production use is concentrated in decision-support and forecasting, not autonomous control of critical assets. The 80-billion-yuan virtual-plant benefit is a cumulative, modeled economic figure, not cash on a balance sheet.
What readers can do now
- If you run a data center, ask your utility about green-power and compute coordination — Qinghai-style dispatch is spreading and can lower both cost and carbon.
- If you work in critical infrastructure, copy the "decision-support, not autonomy" pattern: let the model recommend, keep the human's sign-off on the load transfer.
- If you follow energy, track virtual power plants as the proving ground where AI meets real megawatts — Shanghai's numbers show the economics are already positive.