NISCO and Huawei say the YuanYe steel model is now in commercial use from ironmaking to quality inspection, with company-reported savings of about US$15 million a year. The individual gains are modest, which is the point: value comes from connected equipment, shared data and frontline staff who can build tools themselves. Executives should count models in daily production, not pilots.
What NISCO Says It Has Built
Nanjing Iron & Steel (NISCO) announced this week that, with Huawei and other partners, it has put its YuanYe Steel Large Model into commercial use across several production scenarios. Coverage published on 27 and 28 October describes AI embedded in ironmaking, steelmaking, rolling and quality inspection, rather than in one showcase line. The headline figures are about breadth. NISCO reports more than 300 AI models deployed across its plants and 25 business agents launched by October 2026, covering visual inspection, process optimisation, production scheduling and operations management. Frontline employees have built more than 1,000 low-code agents and lightweight applications. The foundation is data. The company says all key equipment is connected and digitally controlled, and that millions of operational data points from 26 business domains feed a real-time digital twin in its operations centre.
Read the Numbers as Claims, Not Benchmarks
These results come from company and partner material, and one of the two reports is labelled partner content. They have not been independently verified, and the same release gives an 80% efficiency gain for metallographic inspection while its subheadline speaks of efficiency doubling. Executives should treat the figures as a direction of travel, not a benchmark to copy into a business case. With that caveat, the claims are specific. In ironmaking, blast furnace temperature prediction is said to give operators a two-hour warning with more than 90% accuracy, and prime-grade hot metal yield is reported up by more than 10%. In rolling, the rolling-force hit rate is reported to have risen from 88% to 98%, with overall yield up 0.5% on more than 2.5 million tonnes of plate a year. In the laboratory, automated metallographic inspection is said to cut the time per sample from 10 minutes to 3 to 5 minutes. A sulfur-control agent working from 602 process parameters is credited with one-hour risk warnings and annual benefits above CNY600,000, about US$90,000.
Small Gains Only Pay When They Stack
Look at the size of the individual wins. Half a point of yield on a plate mill, a faster lab test, an earlier warning on a furnace, a sulfur agent worth less than US$100,000 a year. None of these would justify a transformation programme on its own, and most would struggle to survive a capital committee as a standalone pilot. The value appears when the gains run at the same time, on the same data, across the whole route. On output above 2.5 million tonnes, half a point of yield is more than 12,000 tonnes of plate. NISCO puts the combined effect, including R&D cycles reported to be 50% shorter, at more than CNY100 million a year, about US$15 million. Huawei's steel lead, Cheng Yongfu, put it directly: the value of industrial AI "can only be realised when it expands from single-point applications". That is the uncomfortable message for companies that have run one impressive pilot and stopped there.
The Frontline Is the Multiplier
The most important number in the announcement is not 300 models. It is the 1,000 low-code agents and applications built by frontline employees. That points to an operating model in which the people who run the furnace and the mill can turn a known problem into a working tool without waiting for a central project. NISCO's AI lead, Li Jinyan, framed the purpose as "not to replace people, but to unleash human potential". Whatever one thinks of the phrasing, the design choice matters. A central team can build a few dozen models. Only a trained workforce with access to clean data can build hundreds of small tools. Others are reaching the same conclusion through policy. In South Korea, a 23.5 billion won programme led by RIST with 15 partners will run from 2026 to 2029 to spread AI across the Pohang steel complex, including a support centre for small and mid-sized firms. The stated aim is to move AI from a few companies to the whole complex.
What Executives Should Ask Before the Next Pilot
Start with connectivity. If key equipment is not connected and its data is not clean, the second, third and tenth use case will each start from zero. That cost, not the model, is what kills scale. Then count what is actually in production. Industry examples already show the pattern: GMK Center's review cites Tata Steel Kalinganagar running more than 260 AI algorithms for real-time decisions. The useful question for a board is how many models run every shift, who owns each one, and what each is measured against. Finally, decide who is allowed to build. A company that keeps every AI tool inside a central team will get a few good projects. A company that gives trained operators safe tools and data will get many small ones, and the many small ones are where the tonnes and the savings add up.
Sources
Intelligent CIO APAC, 27 October 2026: Steel manufacturer deploys world's first swarm-intelligence model at scale Developing Telecoms, 28 October 2026: From Expertise to Intelligence: NISCO Deploys World's First Swarm-Intelligence Steel Model at Scale Seoul Economic Daily, 18 October 2026: Pohang Steel Complex Launches AI Project With 15 Partners GMK Center, 31 December 2025: Robots near the blast furnace: artificial intelligence in steel industry