AI4S Doesn't Lack Stories; It Lacks Real Commercial Value | Linear Voice

As AI for Science moves from concept toward real-world deployment, "industrial capability" has become the unavoidable benchmark for the entire field. Every scientific question must eventually be validated in the real world — it has to enter the fermenter, the factory, the supply chain, the customer's production line.

But the central question is: what does the commercialization path for AI4S actually look like? From the first lab result to the first recurring purchase order, what real barriers must be crossed, and what trade-offs must be made along the way?

Recently, at an event co-hosted by the World Artificial Intelligence Conference's Future Tech, Linear Capital, and Shanghai Jiao Tong University's Industrial Research Institute, Linear Capital partner Yingzhe Zeng, together with Linear portfolio founders Rui Su of SynMetabio and Luyang Zhang of Deep Principle, shared how they use AI to compress trial-and-error costs, build data flywheels, and find commercial footholds for AI4S in real industrial scenarios.

For a long time, discussions about AI4S focused mainly on whether it could "discover new things." AlphaFold predicting protein structures, large models screening for new materials, AI uncovering previously overlooked molecular combinations — these were the headline examples.

But for people actually working in industry, "discovery" is often not the hardest part.

The real difficulties come after: once a molecule is discovered, how do you reliably mass-produce it in a 20-ton fermenter? How does a lab result navigate the constraints of process, equipment, and real-world conditions to become a product customers will keep paying for?

In biomanufacturing, this complexity is especially pronounced. Today, AI can already make extensive predictions at the molecular level. But once a strain enters an industrial fermentation system, variables explode: pressure differentials, dissolved oxygen, feed rates, impeller design, even micro-scale changes that defy complete explanation — any of them can cause the entire fermentation to fail.

Traditionally, these problems relied heavily on experienced engineers' "feel" and prolonged trial and error. Now, AI4S is beginning to change that.

When the test for AI for Science expands from "model capability" to "industrial capability," those scientific questions must return to the real world for validation — entering fermenters, factories, supply chains, and customers' production workflows. The core question remains: what does AI4S commercialization actually look like? From first lab result to first recurring order, what real obstacles must be crossed, and what trade-offs must be made?

In this process, China's complete manufacturing system, industrial scenarios, and engineering capabilities are becoming an increasingly important — and increasingly hard to replicate — advantage in global AI4S competition. Below are highlights from the conversation:

Rui Su, SynMetabio

SynMetabio was founded in 2021. We're a biomanufacturing company that uses microorganisms to produce elastomer materials. When we first started, AI wasn't as buzzy as it is now, but since we're in biomanufacturing, we've been gradually exploring how to use AI to empower traditional industry end-to-end since ChatGPT emerged.

What I'll mainly share today is our understanding of AI, and the opportunities and explorations around leapfrogging — particularly in the transition from lab to factory scale in manufacturing.

Scale-up fermentation is a widely acknowledged extreme challenge in synthetic biology. Tools like AlphaFold mostly solve molecular-level problems. DNA itself is a language; after translation into proteins, the subsequent processes still belong to a language system. That's relatively straightforward in this industry. But in biomanufacturing, the situation is completely different.

In biomanufacturing, the hardest thing is never discovering some valuable product. For our current product selection, our backup library probably already contains over 60 materials and molecules, all worth pursuing. But the core pain point lies in actually achieving industrial-scale production.

To give a simple example: from our initial bio-elastic condensed molecules all the way to mass production in a 20-ton fermenter, it took us nearly four years. We abandoned countless approaches along the way.

The biggest problem here is: biology itself is a highly complex "black box." We often joke that today the fermentation tank might fail simply because "we didn't pray to the right god." There are too many variables — you can never predict what chain reactions will occur when organisms reach extremely high densities of 10^12 to 10^13.

Using existing techniques like controlling pH, dissolved oxygen, or fine-tuned feed control can solve some problems, but when scaling up to fermenters above 20 tons, higher-dimensional challenges emerge.

A classic case: yeast might not withstand the liquid level's pressure differential, or impeller design might create excessive linear velocity that directly "kills" the cells or bacteria.

Solving these problems in the past relied purely on traditional engineers' experience and intuition — they were like "old masters" controlling industrial details, something that couldn't be simulated with conventional models.

To address this, we developed a strain-specific system capable of simulating fermentation performance from 5-liter R&D scale all the way to 20-ton maximum industrial scale. We use various algorithms for composite modeling, because AI Agents have tremendous autonomy — they can write hundreds of mathematical models to composite together. Through this modeling, the system eventually forms an approximately predictable pattern.

So we believe using AI to discover and accelerate the laws of fermentation scale-up is highly meaningful.

To give another simple example: previously, scaling up to 20-ton production might require 7 to 8 consecutive batches of trial and error to find the pattern. In industrial fermentation, the cost of one experimental batch is extremely high.

But with AI, we can simulate on a 200-liter miniature system. One experiment might cost only a few thousand yuan, yet simulate the large-system process, precisely guiding our future actual fermentation production. This is the first part — genuinely reducing costs and improving R&D efficiency through real cost savings.

The second logic is that it forms a new commercial闭环 for biotech companies. We've been thinking: where will AI companies' truly valuable assets lie in the future? We believe it's in vertical industrial scenario data.

Now, after each fermentation completion, the system's built-in adaptive module inputs newly generated data for continuous model learning. As data accumulates, the model's predictions become increasingly stable. In a highly complex field like biology, this kind of scenario-specific AI加持 carries extremely high commercial value, capable of genuinely empowering many application scenarios.

Luyang Zhang, Deep Principle

Deep Principle was founded in 2024. We mainly use deep learning combined with first-principles calculations to unlock new materials and scientific discoveries. Before becoming Deep Principle's COO, I also worked at Horizon Robotics.

I studied semiconductors myself, and in this field there's Moore's Law. Moore's Law means computing power grows extremely fast. But looking at drug and materials discovery, R&D efficiency is actually declining.

We realized there was a great opportunity here: using exponentially growing computing power to compensate for an industry with increasingly declining R&D efficiency.

So we targeted this direction from the start. For materials, the entire chemical space contains 10^60 possible arrangements, making data extremely sparse — it's unrealistic to try solving everything with one large model. So we firmly believed that whatever we did had to be落地, had to solve concrete problems.

Based on this thinking, we built two products: one called Agent Mira™, another called SciClaw, targeting different users. Agent Mira™ is more enterprise-facing; SciClaw serves professional individual users, helping them improve research operation efficiency.

Over the past two years, we've accumulated many customers. In AI computing infrastructure, new energy, beauty R&D, and other directions, they've achieved many results through our products (including Agents and specific models).

We've also basically closed the commercial loop. At first, we might just help customers solve one specific small problem, in project service mode; now, we've gradually shifted to enterprise deployment subscription models, and potentially value-based partnership models after achieving results with our tools. So the company's commercial progress over these two years has been quite smooth.

The whole process is continuous iteration between frontier research, technology development, and commercialization.

Rui Su, SynMetabio

This actually nests with what I just said about AI. There are two commercial paths: one is selling services, like some early AI biotech companies mainly did services; the other is selling products. We've always felt we should sell products, because if something truly has value, completing the full commercial闭环 matters more.

But we discovered one major problem in this process: there are too many industrial steps involved. From biological culture, genetic modification, to fermentation parameter determination, to industrial scale-up, to downstream leather finished products including injection molding, blow molding and other processes — the complexity is extremely high.

So we chose to reduce complexity. The core logic was: we don't manage the leather manufacturing process part, others can do whatever they want with that, we just focus on raw materials.

Looking at it today, this step has played an enormous role. Recently orders have surged, and one very important reason is geopolitical impacts — crude oil supply problems in Taiwan, Japan and other places have caused prices of many plastic products, especially high-end rubber elastomers as substitutes, to skyrocket. More critically, it's not about being expensive, it's about being completely unavailable — you simply can't buy them.

In this context, by becoming a raw material supplier, we can very smoothly enter the market, directly supplying Taiwan and Japanese customers who can take it and produce. Higher focus, stronger supply flexibility.

Returning to AI commercialization logic, we're actually considering two paths now. One is using AI capabilities internally to continuously discover new things. For example, our R&D director recently used AI to uncover a new molecular metabolic pathway, very valuable. But we haven't hit the ceiling in our current elastomer market yet, so it's not suitable to invest immediately.

Could we then package the entire "process package" and sell the pipeline to other companies? But who to sell to is a problem, because there aren't many players in domestic biomanufacturing who can take it over.

Thinking from another angle, could biomanufacturing possibly follow a business model similar to new drugs, licensing the process package as an asset? This is a direction I'm still exploring and thinking about.

Luyang Zhang, Deep Principle

I think our thinking was simple from the start: keep running, keep touching real problems, solve problems when you encounter them, don't overthink the process itself.

Our company set two directions: first, unlock new materials; second, the unlocked materials must be industrially viable. Meanwhile, our technical route is also machine learning-based, inherently considering subsequent industrial scale-up possibilities at the early screening stage — this tone matches the problems we want to solve.

In serving different customers, we also realized that facing customers at different stages, they have different needs — some need a result, some need a "shovel."

We also hesitated: are we a "shovel" company, or a company selling material pipelines? In the end we decided not to limit ourselves according to traditional R&D-commercial divisions. As long as we can solve customers' concrete problems at different stages, that's a good business model.

Therefore, Deep Principle is now a company that "both develops its own products and platforms, while also independently developing material formulations in certain specific domains."

These two businesses form a positive循环 — if our own platform isn't something we'd want to use ourselves, the product isn't good; if we achieve good material results using our own platform, that naturally proves the product is good. In this model, we've built strong momentum.

Our method is set a big goal, while using continuous small goals as signals, constantly making choices and corrections.

Rui Su, SynMetabio

We've been at this since our campus startup days five years ago, and have completed a relatively full industrial闭环 with revenue, orders, and scale, still growing continuously. This is the foundation from our first product.

For us, AI is continuous optimization叠加 on top of China's manufacturing base. This also explains why so many American companies doing bio-based leather and bio-based elastomers haven't succeeded, and ultimately companies like ours have to capture this market — what it really comes down to is industrialization fundamentals.

So I believe that in the long run, China's most valuable scenarios for AI for Science will definitely be in China's strongest industrial verticals.

For example, in fermentation where we're positioned, China accounts for about 70% of global capacity; leather production is close to 80% of global output. In these two dominant domains, AI4S has enormous room to further widen the gap with other regions. China has excellent soil for this, there's no doubt.

Second, responding to the cycle question raised earlier: I've been entrepreneuring through years when the metaverse and Web3 were still trendy, and looking back now, much has changed.

I always tell our investors: don't rush. Because compared to other companies, our biggest advantage is: the first product is already running, the second product can be pushed forward slowly,抗周期. In China, valuation doubling speed in most industries may not be as extreme as Silicon Valley, which反而 lets you settle down, find real scenarios, real demand, create real value.

In the end everyone must answer the same question: how to create real commercial value? This is the underlying logic.

So we're constantly asking the team several questions: how much efficiency has AI actually improved? What new things has it actually discovered? How do we evaluate value? These are things we continuously discuss.

Luyang Zhang, Deep Principle

I'll speak from two levels — I've always felt that in doing things, knowledge and action should be united.

First, people matter most. China has the best scientists and the best engineers. I personally worked in the US for a long time, and one day I figured it out: why stay in a place that doesn't really welcome us, where work isn't enjoyable. Coming back to China, work is actually much happier. When people are happy, they do better work.

Second, more importantly, AI4S can form a闭环 in China. This isn't as simple as "there are companies that will buy your stuff."

For example, suppose we develop a coolant formulation — of course we can screen through machine learning, set specifications and start making it. But our starting point isn't like that. Beyond the lab side, we also run these specification items through a series of server simulations, and collaborate with our downstream partners including data centers and chip companies, obtaining final application scenario planning data before we even start.

This means the formulations we develop are co-optimized results integrating "optimal scientific formulation, optimal end-application, and optimal process scale-up" from birth.

From this perspective, China has the most complete data ecosystem. In the US, you simply can't get this set of things — the US might still need to send things to China for validation. This isn't just commercial resources, but a complete system advantage from science to engineering technology.