Yuanshi Autonomy, an industrial physical-AI company based in Nantong, has completed Pre-A and Pre-A+ financing rounds worth tens of millions of yuan, according to a September 13 36Kr financing report. Dachen Caizhi backed the Pre-A round, while Yuanhe Puhua invested in the Pre-A+ round. The report named both investors but disclosed neither a precise round total nor a spending plan for the newly raised capital.
The company is registered as Yuanshi Intelligence Technology (Nantong) Co., Ltd. English coverage from 36Kr calls it Primordium Intelligence, but the company’s preferred technical branding is Yuanshi Autonomy. That distinction matters because China’s physical-AI sector is filling quickly with young companies whose Chinese and English names can vary across financial reporting, corporate materials, and technical publications.
Yuanshi’s pitch is centered on its World Mechanism Model, or WMM. The company says the system introduces physical constraints into AI learning so that models can operate more reliably in industrial settings where training data is scarce and mistakes carry a high cost. That is a credible problem to address, but it remains a company claim rather than an independently verified benchmark.
The industrial case for physical AI
Physical AI has become a catchall term for systems that connect models to robots, machines, sensors, and real-world processes. In a factory, an apparently capable model must deal with the awkward realities of equipment wear, sensor noise, edge cases, safety rules, and processes that cannot be paused for a speculative experiment.
Yuanshi says it is trying to meet that challenge by building mechanism knowledge into a model instead of relying only on correlations in large datasets. The idea is not unique to this startup. Engineers have long used physics-based simulation and digital twins. The newer question is whether models can combine those constraints with learned representations in a way that is useful on the factory floor.
The financing report does not provide a customer list, deployment scale, or independent performance test. Those omissions matter because industrial AI systems must prove reliability in operating conditions, not only in a technical description. Industrial buyers will ask for evidence from real factory settings.
EastFrontier’s recent coverage of PHYMI’s large seed round for physical AI showed how investors are funding a broad range of companies that want to make AI more useful beyond screens. The Xiaoyu industrial-embodied-AI round reflected the same investor interest in factory settings. Yuanshi represents a narrower industrial route: use models to understand the mechanisms of machines and processes rather than begin with a general-purpose consumer robot.
Why investors are financing the data-scarcity argument
Data scarcity is a recurring problem in manufacturing AI. A consumer platform can gather feedback from millions of users. A high-value industrial process may have only a small number of machines and few acceptable failure cases. Collecting more examples can be slow, and the most valuable examples may be precisely the rare conditions a factory wants to avoid.
That is why Yuanshi’s emphasis on physical rules is likely to appeal to industrial investors. If a model can start with known constraints, it may need fewer real-world examples to make a useful recommendation. But that proposition has to be tested case by case. A system that works in one mechanical process may not transfer to another without careful modeling, calibration, and validation.
The funding announcement also underscores a shift in the Chinese AI market. The most visible rounds often go to humanoid robots or general foundation-model companies. Yet industrial automation remains a large target for startups that can demonstrate reliability, integration skill, and measurable payback. Those are less glamorous metrics, but they often determine whether a factory renews a contract.
Reliability will decide whether the model matters
The central test for Yuanshi is not whether WMM sounds different from a conventional AI model. It is whether the approach reduces downtime, improves process control, or makes an industrial decision more trustworthy without adding unacceptable complexity. Customers will also want to know how the system behaves when sensors fail, when a process shifts, or when a recommendation conflicts with established engineering practice.
The company’s funding gives it more room to develop that evidence. 36Kr reported that the rounds were aimed at a company already positioning itself around industrial scenarios, but did not disclose specific spending plans. The most sensible uses would include product engineering, deployment pilots, and the difficult work of converting research logic into tools that operators can use.
China’s physical-AI race will not be won by a single model architecture. It will be won in deployments where models must coexist with machinery, engineers, and commercial constraints. Yuanshi’s financing is a sign that investors see value in a physics-aware approach. The next question is whether the company can prove that approach in production environments.