XtalPi puts predictive AI into Fangda Carbon graphite electrode production

XtalPi Holdings said Friday that an AI model built with Fangda Carbon for choosing and blending graphite-electrode raw materials has cleared acceptance testing and is running in Fangda's production workflow to cut cost and speed substitute-material decisions.

Most industrial AI pitches stay in the pilot pen. XtalPi says this one cleared acceptance and is already ranking real graphite-electrode recipes for a major supplier — AI as a procurement and quality lever in steel's power-hungry electrodes, not another chatbot demo.

On Friday, 2 October 2026, XtalPi Holdings Limited said a raw-material selection model it built with Fangda Carbon New Material Co., Ltd. passed acceptance testing and is now in use in Fangda's production workflow. The PR Newswire page is titled “XtalPi and Fangda Carbon Deploy Predictive AI to Optimize Cost and Material Efficiency in Graphite Electrode Manufacturing.” The page stamps Oct. 2, 2026, 10:01 ET, which is 10:01 a.m. Eastern. The dateline is Cambridge, Massachusetts, and Shenzhen, China. The source line is XtalPi Inc. XtalPi's shares trade as 2228.HK. Fangda Carbon's shares trade as 600516.SH. A raw-material selection model, here, is software that ranks which ingredients to use, and in what mix, before a plant runs a physical trial. Those lines are the wire's.

What the lines under the headline say the model does. The jointly developed model has passed comprehensive acceptance testing and is operating inside Fangda Carbon's production workflow. By forecasting performance and screening formulations before physical trials, the release says, the system meets cost-optimization targets and lets the plant evaluate substitute raw materials quickly. This first deployment, the release says, is a base for XtalPi's industrial AI work in new materials: turning manufacturing records into predictive models that can be reused for advanced carbon materials. A formulation is the recipe: which raw materials, and how much of each. Screening means the model sets aside poor recipes before anyone mixes them in the plant. Those lines are the release's. The page does not print the cost target as a percentage, and it does not print a dollar amount saved.

What a graphite electrode is, as the release explains it. Graphite electrodes carry the very large electrical currents used in electric arc furnace steelmaking. An electric arc furnace, often shortened to EAF, melts scrap steel with an electric arc, a spark at the scale of a furnace rather than a welding torch. The electrode is the carbon piece that carries that current into the furnace. The release says key properties, including how well the electrode conducts electricity and how strong it is, depend on the raw materials, the blend ratios, and the manufacturing conditions. A blend ratio is the share of each ingredient in the recipe. Those lines are the release's.

Why the recipe cannot sit still. The release says materials of the same type can differ by supplier and by batch, so a formulation has to be looked at again when the inputs change. Price swings and broken supply, it says, make a fast and accurate swap of materials a commercial necessity. Expert judgment still matters. Trying every combination in the plant is, in the release's words, prohibitively slow and expensive. Those lines are XtalPi's, on the wire. The page does not print a count of combinations, and it does not print a price for one trial.

What the partners say they built. They combined Fangda Carbon's six decades of proprietary manufacturing data with XtalPi's data engineering and algorithms. Six decades is 60 years. Proprietary means the records belong to Fangda and are not a public dataset. XtalPi says it structured decades of scattered production records, engineered predictive features, and deployed a hybrid system. A predictive feature is a number drawn from those records that the model uses to estimate how a recipe will behave. Hybrid, here, means three pieces together: a forecast of performance, optimization algorithms that search for a better blend, and expert rules written into the software. An optimization algorithm is a procedure that hunts for a mix that meets the goal, in this case lower cost at the required quality. Those lines are the release's.

What the model does once the ingredients are fixed, and what it does when they are not. For any fixed set of raw materials, the release says, the model quickly finds blend proportions that lower cost while still meeting quality requirements. When supply or price shifts, it evaluates other inputs and recommends how to change the formulation. The release says that widens Fangda Carbon's room to buy. “Significantly expanding” procurement flexibility is the release's phrase. Procurement is the buying. The page does not print a percentage of money saved, and it does not name a substitute material.

How the release says the model was checked, and who still decides. Validation against independent test datasets and against production trials confirmed that the model reached the required predictive accuracy across multiple key performance indicators. A key performance indicator, often shortened to KPI, is a measured result the plant already tracks, such as strength or conductivity. The release says the model hit the required accuracy on those measures. It does not print the accuracy as a percentage, and it does not name the indicators. It says the calculation is fast enough to keep up with the pace of production, so it fits Fangda's workflow. The model screens out unsuitable options and ranks the ones worth trying, which narrows the work for human experts. Final formulations are still set by those experts, through an experiment or a production check. The release calls that “compute first, verify later.” The computer proposes a recipe. People still check it before it is final. The model does not take that last decision away from them. Those lines are the release's.

Where this sits in the longer project. The accepted model is the first core module of a graphite-electrode formulation project that comes from a strategic agreement signed in 2025. The release says the partners are connecting XtalPi's AI, its robotic experimentation, and its quantum-chemistry work with Fangda's industrial position, and turning scattered historical know-how into records a computer can trace. Robotic experimentation, in that sentence, is lab robots that run physical tests. Quantum chemistry is computation from the physics of atoms and molecules. The release names those capabilities as part of the partnership. It does not say Friday's raw-material model is itself a robot run or a quantum-chemistry calculation. Going forward, the two companies plan to expand the system into fuller formulation design and into process optimization. Process optimization means adjusting how the electrode is made, not only which ingredients go in. For XtalPi, the release says this job extends industrial delivery: reusable data standards, model designs, and a way of handing the work over. Adapting that setup to other advanced carbon materials, including graphene and carbon nanotubes, is the release's stated path toward a wider industrial AI business and toward making next-generation materials at commercial scale. Graphene is a sheet of carbon one atom thick. A carbon nanotube is a tube rolled from that same carbon. Those materials are named as a path. The release does not say this model is already running on a graphene line or a nanotube line. Those lines are XtalPi's, on the wire.

Who the about box says XtalPi is. XtalPi Holdings Limited, ticker 2228.HK, was founded in 2015 by three physicists from the Massachusetts Institute of Technology. The box calls it an R&D platform powered by quantum physics, artificial intelligence, and robotics. R&D is research and development. It says the company combines first-principles calculations, AI algorithms, high-performance cloud computing, and standardized automation. A first-principles calculation starts from physics rather than from a fit to yesterday's plant data alone. The box says XtalPi provides digital and intelligent R&D tools to companies in pharmaceuticals, materials science, agricultural technology, energy, new chemicals, and cosmetics. “A global leader in graphite electrodes” is the release's description of Fangda Carbon. The page does not print a ranking table under that phrase. The phone line on the page is PR Newswire's, 888-776-0942, from 8 a.m. to 9 p.m. Eastern. The release does not print a price for the model, and it does not describe a regulator's clearance.

The picture is a split graphic. The left side is a light field with the XtalPi logo. A teal band separates that field from a dark navy panel. The panel is labeled XTALPI and SOFTWARE. The lines read: Predictive AI live at Fangda Carbon; Raw-material selection model; passed acceptance — production; Graphite electrodes, EAF steel; Cost + substitute-material screen; and XtalPi Holdings (2228.HK). EAF is the electric arc furnace named in the release. The frame does not print a calendar date. It is not a photograph of a furnace or a plant. It does not print a savings percentage or an accuracy score.

In plain terms, XtalPi said on Friday, from Cambridge and Shenzhen, that a model built with Fangda Carbon for choosing and blending graphite-electrode raw materials has passed acceptance testing and is running in Fangda's production workflow. For a fixed set of ingredients it looks for a cheaper blend that still meets quality. When supply or price moves, it ranks substitutes. The company says checks on held-out data and on production trials met the accuracy the project required, and that the math keeps up with the plant. It does not print that accuracy as a percentage, and it does not print a dollar saving. People still set the final recipe. The release calls the pattern compute first, verify later. The model is the first module of a 2025 agreement. Graphene and carbon nanotubes are named as a later reuse, not as lines already running this model.