On the afternoon of September 19th in Shanghai, the first session of SVTR Frontier League brought four industries with almost no overlap to the same table: power and water utilities for industrial digital twins, manufacturing for physical modeling, drug development for virtual experiments, and engineering design for blade optimization.
Three entrepreneurs were present, and one scholar joined via a video call from Hong Kong. By the end of the discussion, all four were stuck on the same issue.
One-third of the attendees were investors who came to listen to physical constraints
The registration structure of this event itself was a signal: a closed-door session discussing simulation accuracy and physical equations, with more than one-third of the registrants coming from investment institutions, and the rest mostly entrepreneurs and corporate R&D personnel.
What they were looking for was clearly not another model, but a sample that could clearly explain the moat.
First layer: Sparse data is the default state for these industries
Yan Chen, founder of Shanghai Xuntu Digital Technology, came from the distribution network background. He recalled that the transmission network could be managed with full sensor coverage and a manual rule base, but once the number of distribution network equipment and user connections increased, this approach no longer worked. The team therefore turned to using limited node data to replicate the entire system's behavior.
He gave the example of a watershed: for hydrodynamic prediction, the focus is on replicating the overall movement of the watershed using node sensor data, not calculating the trajectory of each drop of water. The Q-MAP engine also makes similar trade-offs: its rendering capability aligns with game-level standards, but it actively cuts out features like hair and fabric that are not used in industrial scenarios, allocating all resources to real-time rendering and control of large-scale dynamic objects. According to the SVTR AI venture database, the company was founded in 2003 and was spun off from the State Grid Research Institute. Its graphics engine has been developed for over 20 years.
He wrote the endpoint of this path on the projection: the premise of industrial AI is to first create a synchronized closed-loop digital twin that is visible to everyone, before accumulating an industrial model with real-world structure and physical constraints, and then obtaining an industrial intelligence agent that does not generate hallucinations.
Shen Chongzhi's founder Fan Xihong encountered the same constraint in manufacturing, but his explanation was more direct: many key physical quantities are naturally difficult to measure, or the cost of deploying sensors is not recoverable, and the total amount of data itself does not constitute a bottleneck. His approach is to let physical equations such as heat conduction and diffusion enter first as constraints, with on-site data only responsible for calibrating parameters. According to his on-site statement, this approach still works when data is sparse, and the reasoning process can also be connected to process and equipment control logic.
Second layer: The enterprise system stores results, not the reasons
What truly brought the four together was the next layer.
Jiang Shuo, a researcher from the Department of Systems Engineering at Hong Kong City University, works in engineering design. He said that existing industrial software records documents, approvals, supply chains, and management information, but AI needs the intent behind each designer's choice: why a plan was rejected, when constraints were changed, and how performance, cost, and manufacturing processes were balanced. These are usually scattered in meetings and personal experience.
Xu Ke, founder of Foxu Virtual Lab (Foxu.AI), gave a nearly identical list in the pharmaceutical field. He believes that even large pharmaceutical companies still operate in a sparse data environment, because what is valuable is not just the result of a single experiment, but also whether the project was successful, failed, or terminated, key decisions at each stage, and the context in which experts made changes. Their strategy therefore focuses on continuously engaging with ongoing research: providing products and services to CROs and biotechs, with data growing from the service, not through procurement.
Yan Chen added the sharpest point: data in industrial and social management systems is not only missing, but may also be biased, inconsistent in definition, or even manipulated. If the underlying data is untrustworthy, no matter how complex the model or how powerful the computing power, the output cannot be used as a basis for decision-making.
Three statements together point to the same thing: enterprise systems are good at preserving results, while AI needs the choices made during the process.
Third layer: Products need to know when to stop
Xu Ke's case was the most specific. In clinical statistics, if the patient enrollment date has only incomplete information, the model should not fill in a date and proceed, but the system should identify the incomplete data and hand the judgment over to clinical, sponsor, or regulatory personnel.
He broke down reliability into four things: knowing the applicability boundary of its conclusions, being able to reproduce the calculation, having the input, rules, versions, and intermediate processes retained, and refusing to answer and reporting when uncertain.
Jiang Shuo gave another reason from the engineering side. AI that has long self-optimized in a purely virtual environment tends to over-converge on a single metric, leading to theoretical optimal solutions that are not feasible in engineering. Therefore, engineers' judgment remains irreplaceable: when to use low-cost proxy models, when to switch to high-fidelity simulation, and when to proceed to physical testing.
Fourth layer: Make the deliverer and the client stand on the same side
Fan Xihong's part on business model was very practical. He projected a single sentence on the screen: sustainable delivery relies on aligned interests.
Specifically, it is two things that are closely linked. One is charging by volume, with revenue directly tied to the client's output: the more stable the client's production and the higher the yield, the stronger the willingness to pay, and the deliverer must therefore continuously care about long-term operational results, rather than selling and walking away. The second is what he calls dCOA, a digital process and quality certificate. Each process improvement brought by the model is recorded, verified, and issued as a certificate, making process certainty traceable, verifiable, and tradable. Quality certification moves from post-hoc sampling and static tables to data-driven proof throughout the process, giving the client a basis for payment decisions and the deliverer a basis for continuous investment.
He concluded that this logic means AI is no longer an auxiliary tool on the production line, but a productive unit that the client is willing to pay for continuously.
But he believes the biggest obstacle to implementation is not the model. According to his on-site statement, most projects get stuck on whether the organization is willing to change its workflow. In traditional enterprises, even changing a data page requires going through several layers: requirements, approval, development, and testing. He also reminded one thing: deploying a large number of industrial software does not mean that AI can be implemented. These systems accumulate management data; AI needs process data, intent data, and anomaly data, which are usually not present at all.
After this event, how do we view this sector
The judgment is clear: whether AI for Science projects can be replicated across customers depends first on whether the customer has the ability to generate data, and then on the model.
The four guests come from completely different industries, but their moats point to the same place: whether the company is embedded in the customer's data generation process. Companies that only buy static databases and those that integrate experiments, equipment, simulations, and customer operations every day will have completely different assets after a few years.
This aligns with our previous judgment. As of September 9, 2026, the AI4S tag in the SVTR AI venture database has 118 confirmed financing rounds in the past 12 months, with 13 companies securing over $100 million in a single round, accounting for 73% of the disclosed funds. Capital is pricing in "things that others cannot do in the short term"; and the event told us that the most difficult-to-replicate capabilities often do not lie in model parameters, but in the position of data.
For funds watching this sector, the next due diligence question is clear: is the company's data bought or grown?
The complete financing data for AI4S can be found in "The Valuation of AI4S Begins to Diverge Before the Results Emerge", and the details of this event can be found in SVTR Frontier League #001｜AI for Science: From Model to Discovery.
The next time we look at these companies, we will focus on three things: who is continuously getting new data, who has started to enter the customer's daily workflow, and who can turn experimental and engineering results into long-term revenue.
SVTR Frontier League discusses one frontier direction per session, targeting global AI entrepreneurs and investors. Welcome to join us and have a conversation.