A blast furnace does not forgive mistakes. For a century the hottest calls on a steel line came from veteran operators who could read a furnace by sound and smell. This year, at China's largest steel group, some of those calls started coming from a model.
On a Spring Festival eve in 2025, at a Baosteel (宝钢) silicon-steel unit, a cleaning-section temperature drifted from 80°C to 76°C. Before a human would have noticed, the AI large model flagged it, an operator acted, and a potential fault dissolved. Nobody rushed. The "co-worker" on duty was software.
From "AI+钢铁" slogans to the floor
China Baowu (中国宝武) is the country's flagship steel combine. In 2025 its "AI + steel" (AI+钢铁) push moved from brochures to production. The headline result: Baoxin Software (宝信软件) developed a 钢铁大模型 (steel large model) whose core technology and application plan won the SAIL Award — the highest honor at the 2025 World AI Conference (WAIC).
The model is not a chatbot bolted onto a mill. Baowu describes a full stack: a Baoxin-built industry model (宝联登钢铁行业大模型), a "predictive manufacturing" smart factory for high-end green silicon steel, and a "问数2526" finance agent that answers data questions in seconds. At the group level, Baosteel committed to a three-year build of four foundations — compute, models, talent, and organization — targeting "thousand-card compute, thousand-model capability, thousand-person team," with a goal of 1,200-plus AI scenarios and 25-plus benchmark AI production lines.
The furnace gets an AI brain
The most telling deployment is the 高炉大模型 (blast furnace large model), built with Huawei. The furnace has always been the "black box" most dependent on veteran instinct. Baowu says the model's prediction hit-rate and control-adoption rate both exceed 90%, and that furnace-heat control precision improved markedly. In plain terms: the system suggests adjustments, operators accept them most of the time, and the process stays steadier than before.
At the silicon-steel smart factory, Baosteel reported 100% equipment numerical control and 28 smart scenarios as of mid-2024, then in October 2025 joined the first national batch of 15 "领航级智能工厂" (pilot-level smart factories) — the top tier of a six-ministry cultivation program. A sister site, Nanjing Steel's special-steel personalized-customization plant, made the same list.
One site, 324 robots, 186 agents
The rollout is not only at the flagship. At TISCO (太钢, Taiyuan Iron & Steel), a 2025 "2526 project" pushed AI across ironmaking, steelmaking, and cold-rolling. By the company's account, 85% of main-process steps now run under intelligent centralized control, key-process numerical-control rate exceeds 86%, 324 industrial robots are deployed, and 186 intelligent agents sit on 368 knowledge bases — 15 of them highly active across the business chain.
There is a sovereignty angle too. Baoxin's Tianxing PLC (天行可编程逻辑控制器) achieved full-stack autonomy, breaking a long reliance on imported high-end controllers; Baowu says it has landed in 100-plus projects covering about 60% of steel processes.
Green is the other engine
AI and decarbonization travel together here. On December 23, 2025, Baosteel Zhanjiang's million-ton-class near-zero-carbon high-grade plate demonstration line was fully connected — a marker for hydrogen-based direct-reduction steelmaking in China. Across the group, "AI + energy" scheduling and low-carbon product lines (BeyondECO) are presented as the same transformation, not a side project.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, steel is the most conservative industry to trust a model with live production — "when a blast furnace takes the model's advice, that is AI leaving the lab for good."
Honest limitations
Most figures come from Baowu/Wugang group releases, China Metallurgical News reporting, and industry summaries; they are company-disclosed, not independently audited. "Exceeds 90%" for the blast-furnace model is stated without a public methodology, and "1,200 AI scenarios" is a target-and-progress claim, not a verified count of production systems. Steel AI also depends on rich historical process data that newer or smaller mills lack, so the Baowu playbook does not copy trivially.
What readers can do now
- If you buy or specify steel, ask suppliers about low-carbon and traceable grades; Baowu's BeyondECO line signals that certifiable green steel is becoming a purchaseable product, not a promise.
- If you work in process manufacturing, copy the "predictive manufacturing" pattern: start with one black-box step (furnace, kiln, reactor) where a model can suggest, not replace, the operator.
- If you follow industrial software, watch Baoxin's Tianxing PLC and industry models — domestic controllers plus domain models are the quieter, higher-moat story than the flashy agents.