Ant's EnergyTS 3.0 forecasts weather 45 days out for power traders
The AI-to-energy story in the West is about how much power the models need. In China it is about who sells the grid a forecast — and Ant Group's enterprise arm just pushed the horizon out to six weeks.
Ant Digital Technology (蚂蚁数科) used the Bund Summit in Shanghai on September 10 to ship EnergyTS 3.0, its time-series foundation model for the energy and carbon sector, with a forecast horizon of up to 45 days. The company claims weather-element forecasting error that beats comparable models across multiple public benchmarks, and a prediction-interval coverage ratio above 95%, on training data covering more than 100 million electricity trading records. The upgrades are multi-factor learning, retrieval of analogous historical conditions, probabilistic forecasting, and long-horizon adaptation — meaning the model was rebuilt around how far out it can go and how honest it is about uncertainty.
The number that matters is the coverage ratio, not the 45 days. Point forecasts are useless to a trading desk; a calibrated distribution is what you can bid and hedge against. A model that says "I am 95% confident the answer is inside this range" and is right 95% of the time can be wired into a risk system, which is exactly what Ant is aiming at. This is the same problem Google is attacking on the open-science side with its own weather models — we covered that trade-off in Google's WeatherNext 3 trades the six-hour lag for hourly forecasts — but Ant is selling it as a line item in a procurement, not a research release.
The model is the engine for something bigger: Agentar Energy Edition, an "agent super factory" for the sector. Ant Digital packaged energy and carbon know-how into ready-made skill libraries and "digital expert" agents that coordinate sub-agents against vertical SaaS systems. In the electricity-trading workflow the company describes, one orchestrating agent dispatches sub-agents through forecast, sensing, analysis, decision, bid submission, risk control and post-trade review — the whole loop, not a copilot for one step. In EV and charging-infrastructure operations it does the same across site selection, operations, battery services and predictive maintenance. Ant says its energy "digital employees" now cover more than 20 sub-scenarios across generation, grid, load, storage and carbon, and are deployed at GCL Technology and Linyang Energy.
Why it matters: this is the other half of the AI-energy story happening this week. While Samsung and SK Hynix were refusing to prepay $18.7 billion of AI power bills, Ant was selling the grid's own operators the tools to run it. American labs are selling compute; Chinese enterprise AI vendors are selling into the constraint that compute runs into. Ant's claims are its own and no independent benchmark run of EnergyTS has been published, so treat the "leads comparable models" line as vendor reporting.
What to watch: whether anyone publishes a third-party evaluation of the 45-day horizon, and whether Ant discloses how its probabilistic calibration degrades week four to week six — that is where the 95% figure earns or loses its keep.
If an agent filed the bid, took the position and wrote the post-mortem, would your risk team sign off on it — tell us in the comments.
