Factory-Native AI: Why PowerPoint Demos Still Can't Run a Production Line

WAIC 2026 just wrapped up in Shanghai, and the headlines were all about agent smartphones, humanoid robots, and Kimi K3's "Sputnik moment." But in a quiet conference room away from the packed exhibit halls, a far more grounded question was being debated: how far is AI from actually working on the factory floor?

The gap is wider than you might think.

Factories Are Not the Internet

Wei Jie, chairman of Shanghai Jingzhi, introduced a concept called Factory-Native AI.

To understand what it is, you first need to understand what it isn't. It's not about bolting a quality inspection module onto a production line or hanging a data dashboard on the wall. Wei defines it as an intelligent system that understands industrial constraints, participates in production decisions, and takes responsibility for outcomes.

The difference from internet AI? Here's the sharpest way to put it: when a chatbot says the wrong thing, it apologizes. When a factory AI recommends the wrong process parameter, parts get scrapped and production lines shut down. Real money burns.

Wei identified three systemic barriers to industrial AI deployment:

First: data is hard to get. Factories generate mountains of data daily—orders, blueprints, process parameters, equipment telemetry, quality reports—but they're scattered across different systems, machines, and people. Worse, this data lacks complete production semantics. The AI sees a parameter value and a work order but can't read the hidden causal chains connecting them.

Second: models struggle with expertise. A process plan isn't a single correct answer. It must simultaneously satisfy tolerances, material properties, tooling constraints, heat treatment requirements, quality standards, and delivery deadlines. A production schedule must account for worker skills, material availability, process sequencing, and order priority—none of which a general-purpose LLM can "guess."

Third: there's no room to experiment. Internet products ship fast and iterate faster. Factories don't have that luxury. A single bad recommendation can scrap an entire batch. A poorly timed schedule change can idle a whole line. The margin for error is near zero.

"Fertile Soil": An Open-Source Process Model

At the roundtable, Shanghai Jingzhi open-sourced a process-focused AI model called "Wo Tu" (Fertile Soil).

The training data comes from real factory floors: nearly 100,000 original engineering drawings and 300,000 records of non-standard processes. The goal isn't a general-purpose chatbot—it's an AI that grows directly from real manufacturing data, where the factory serves as data source, training ground, and validation environment all at once.

Shanghai Jingzhi's solutions already cover nearly 300 manufacturers across machining, die-casting, sheet metal, and injection molding. Their clients include Tesla's motor suppliers and component makers for China's Fuxing high-speed trains.

By open-sourcing Wo Tu, Wei is betting on making it the intelligent foundation for Chinese manufacturing processes—so that small and medium factories can actually deploy AI instead of just watching PowerPoint demos.

Physical AI: No Room for "Sorry"

If Wei represents the industrial software camp, Chai Jian, VP of Hicend Intelligent, represents a different approach: Physical AI.

Chai draws a hard line between two kinds of AI. Digital AI generates text and images from language corpora. Physical AI perceives the real world through edge devices and makes real-time decisions.

The critical difference: digital AI can retry when it fails. Physical AI steering a vehicle the wrong way doesn't get a second chance. Chai calls it "non-rollbackable."

Hicend's core product is a multispectral AI fire detection system, already deployed in data centers, power grids, and energy facilities. A telling signal: their biggest customer has shifted from industrial sites to data centers—meaning the expansion of AI compute infrastructure itself is now driving demand for industrial AI.

The Consensus: Follow Profit, Not Hype

The roundtable's closing remarks came from Cao Lili of Spacetime Quadrant, and her advice was brutally simple: "Follow the money, not the noise."

She cited her team's AI collectible toy project as an example: launch a 59-yuan entry product to validate the market, hit a million units in sales, then apply embodied intelligence technology to the IP. "Profit first, iterate second, reinvest stable revenue into R&D. That's long-termism."

Song Renjian, AI & Computation Director at Synbio Tech, offered a deeper observation: how fast an industry gets transformed by AI depends entirely on the speed of its data flywheel.

"Coding is already fully automated because developers can iterate models fast enough. Bio-manufacturing is the same—the speed at which physical experiment data flows back into the model determines how far AI can go."

This is precisely the core problem Factory-Native AI aims to solve: can the data loop from production line to model actually close? Can the model continuously evolve under real industrial constraints?

Final Thoughts

The flashiest exhibits at WAIC 2026 were humanoid robots and agent smartphones. But what truly determines AI's trajectory might be unfolding in these quieter corners—real factories, real hospitals, real retail spaces, where the question isn't "can it impress a demo audience" but "can it run stably, sustainably, and profitably."

When an entire industry shifts from "look at our parameters" to "look at our P&L," that might be the real beginning of AI maturity. The party tricks are over.