Chinese Steelmaker NISCO Says 300 AI Models Save It More Than $15 Million a Year

Nanjing Iron & Steel (NISCO) says its AI models cut annual costs by more than 100 million yuan, about $15 million, in work it did with Huawei.

NISCO has capacity of more than 10 million tons of steel a year. If output came close to that, 100 million yuan works out to about 10 yuan a ton by The Denominator’s arithmetic. At that size, the saving is small per ton, and NISCO did not say what the models and chips cost to run.

What did the models change?

Energy use per unit of output fell 12%, according to a published report. The report tied the drop to the furnace work, and NISCO did not say whether the 12% covers the whole mill or ironmaking alone, or what it is measured against.

In ironmaking, NISCO says its models forecast blast furnace temperature two hours ahead with accuracy above 90%, and the yield of prime-grade molten iron rose more than 10%. In rolling, the share of correct rolling-force predictions climbed to 98% from 88%, and the overall yield rate rose 0.5%. In inspection, a robot grinds and photographs each steel sample while AI grades it, and the time per sample fell to three to five minutes from 10.

Who built the applications?

NISCO says it has deployed more than 300 AI models across its plants and launched 25 business agents, which cover quality inspection, process optimization, production scheduling and operations management. Separately, frontline employees built more than 1,000 low-code agents and lightweight applications on their own.

Huawei supplies the chips, cloud and models underneath: Ascend and Kunpeng processors, Huawei Cloud Stack and Pangu models. As deployment grows and computing demand climbs, NISCO and Huawei plan to explore dedicated data centers for smart steel plants.

What does it cost to run?

NISCO did not break down how the 100 million yuan adds up across applications, and it did not say what the Huawei hardware, cloud and model work costs. Until that cost appears beside the saving, the net return stays unknown.

What would the compute behind 300 models cost a mill half NISCO’s size?