AI for Science Investment and Entrepreneurship: Where Are the Opportunities in the Next Decade?
"That failed data is valuable."
AI for Science, in simple terms, is about making artificial intelligence a "super assistant" and "insight engine" for scientists. On November 25, 2025, Google DeepMind released AlphaFold: Five Years of Impact, using the revolutionary breakthrough in protein structure prediction to put an exclamation point on the "scientific value" of AI for Science.
AlphaFold: Five Years of Impact
Image | deepmind.google
A little over two weeks later, on December 18, XtalPi — a leading Chinese AI + robotics company — was officially included in the Hong Kong Stock Exchange Tech 100 Index. This move not only propelled it into the top tier of Hong Kong's tech sector, but also marked the leap of AI for Science from "technical concept" to "industrial hard power," giving it a tangible "benchmark sample" in China.
As a pioneer in both AI drug discovery and AI for Science, XtalPi has long validated its ability to translate technology into real-world results: from its debut as the "first AI drug discovery stock" on the Hong Kong Stock Exchange, to landing a nearly $6 billion mega-order with DoveTree, to a $345 million collaboration with Eli Lilly and Company — each step has been rewriting the speed at which we move from "AI-accelerated research" to "research feeding back into industry." Not long ago, at FreeS Fund's 2025 Annual Investor Summit, FreeS partner Rui Ma and XtalPi co-founder and Chief Innovation Officer Lipeng Lai sat down for an in-depth conversation. They explored AI for Science, the integration of technological innovation with industry, and the application of large models in biomedicine — unpacking the core opportunities of the next decade for investors and founders alike. The main topics they covered:
- AI drug discovery has reached its harvest season, while AI for Science is still in a from-zero-to-one phase. How do we truly fuse scientific innovation with industry?
- What is the underlying logic driving the development of AI for Science? And around this, what strategic adjustments is XtalPi making for its future business?
- Where are the capability boundaries of large models in biomedicine? Which problems have been reasonably solved, and which remain stubbornly difficult?
- What are the core challenges in migrating AI+ technology from pharmaceuticals to materials, energy, and other fields? What are the keys to making it land?
We've edited portions of the conversation, hoping to offer fresh angles for thinking about these questions. This piece is part of our "AI Industry Watch" series, which will continue to share firsthand practices and observations from AI entrepreneurs. If you're building something in the AI for Science direction, feel free to reach out at marui@freesvc.com.
"AI Industry Watch" Series
From Large Models to AI Companions: What Cyclical Patterns Lie Behind the Rotating AI Hype?
How to Bridge the "Last Centimeter" Between Robots and the Physical World?
"In AI Hardware, There's Always a New Opportunity"
Community Giveaway
Beyond playing chess, chatting, and writing poetry, AI can help scientists push into the frontiers of drug, energy, and materials R&D. If you were calling the shots, which everyday scientific puzzle would you want it to tackle first?
Drop your ideas in the comments! By 5:00 PM on January 16, 2026, the two most creative commenters will each receive a copy of The FreeS Fund Industry Research Handbook.
01. AI Drug Discovery Enters Its Harvest Season
Rui Ma: Let me set the stage by sharing how we think about AI for Science. First, AI drug discovery has reached a point where it's bearing real fruit. Three years ago, plenty of people were skeptical — can AI actually produce drugs? Looking at where we are today, the answer is unambiguous.
Take XtalPi as an example. In November 2025, its wholly-owned subsidiary Ailux reached a $345 million AI-driven macromolecule R&D collaboration with Eli Lilly and Company, a top global pharma company. In August 2025, XtalPi and US biopharma company DoveTree announced an AI drug pipeline partnership with a total order value approaching $6 billion.
Beyond that, Insilico Medicine, another AI drug discovery company, has already listed on the Hong Kong Stock Exchange's main board. Its AI-generated small molecule for idiopathic pulmonary fibrosis (IPF) achieved strong results in Phase II clinical trials.
Then there's METiS Pharmaceuticals, an early FreeS portfolio company. Its self-developed MTS-004 orally disintegrating tablet has reached the primary endpoint of Phase III clinical study, becoming the first AI-enabled formulation new drug in China to complete Phase III trials.
Second, can AI replicate its pharmaceutical success and extend its capabilities outward to chemistry, materials, physics, and other fields? We've found that AI may be the biggest variable driving fundamental innovation across these scientific disciplines.
This is a crucial part of today's conversation — beyond AI drug discovery, can we still capture listed or pre-listed companies of this caliber over the next five years?
Third, since AI drug discovery is already bearing fruit while AI for Science remains in a from-zero-to-one phase, can we actually fuse these technological innovations with industry?
The "15th Five-Year Plan" essentially gives us the answer. It proposes that over the next decade, China should build another high-tech industry around quantum technology, biomanufacturing, brain-computer interfaces, embodied intelligence, and nuclear fusion.
So what we want to explore today is whether AI for Science can ultimately combine technological innovation with industrial innovation, creating a multiplier effect from one to one hundred.
Let me introduce Dr. Lipeng Lai. Lai is a co-founder of XtalPi and its Chief Innovation Officer. He has been one of the three founders since day one, and has been with XtalPi all the way to its IPO. He currently leads XtalPi's Innovation Center, pushing from-zero-to-one technological innovation and translation across peptides, gene and cell therapy, and other fields based on XtalPi's foundational capabilities — doing a tremendous amount of AI+ and AI for Science work. Many of the new drug modalities XtalPi is advancing internally fall under his purview.
My first question for Lai: XtalPi has had an excellent run in the capital markets over the past year. Could you walk us through XtalPi's current business practices and strategic positioning? And what new moves are coming on the strategic and business fronts?
Lipeng Lai: Building on what Ma just said, let me start with why the core logic of AI for Science holds up, and how that connects to the commercial viability of AI for Science and XtalPi's future business planning.
People often note that compared to AI, the human brain runs on remarkably low power. AlphaGo needed massive supercomputers; a person just needs a bowl of rice. The reason for such low consumption is that we're good at simplifying problems. Kepler observed celestial motion and reduced it to Kepler's elliptical equations. It looks simple, but the natural world doesn't exist for human convenience. There are two categories of things that cannot be simplified.
One is the vast macroscopic realm — complex celestial motion, for instance, where neither humans nor computers can adequately solve the chaotic three-body problem. The other is the microscopic world, encompassing biology, chemistry, materials science, and so on — domains whose complexity also exceeds what the human brain can readily handle.
Take biology as an example. Humanity's pursuit of health dates back thousands of years to Qin Shi Huang's quest for immortality. In modern industry, this is a trillion-dollar annual market. What this means is, first, the existing market is enormous, and second, human intelligence simply cannot adequately comprehend this complex system. So AI is bound to create value within it — this is the underlying logic for why AI for Science can develop going forward.
Looking at concrete implementation: over the past decade, AI has accelerated, to varying degrees, every critical individual step in drug R&D. Based on XtalPi's own experience, in preclinical drug discovery, AI improves efficiency by an average of 20% to 80%.
Additionally, biological data remains extremely scarce compared to internet data. First, data collection costs are high. Second, data quality is relatively low — experiments conducted in different labs, documented in different papers, or even by the same person at different times can yield data of varying quality. Based on these two factors, we believe that over the next three to five years, data will become a critical asset in the AI biopharmaceutical space. Accordingly, XtalPi's business strategy focuses on several directions:
First, greater emphasis on internationalization. We are very bullish on the development of China's innovative drug sector. From a short- to medium-term perspective, our strategy will definitely be a hybrid approach combining "internationalization + China." Our R&D collaboration with Eli Lilly and Company, for instance, represents an important step in XtalPi's international expansion.
Second, diversification of drug modalities. Over the past decade, AI has primarily focused on small-molecule compounds, which account for roughly 70% of the pharmaceutical market. Of the remaining 30% comprising biologics, 25% are antibodies and 5% are other types such as peptides, gene and cell therapies, and vaccines.
XtalPi initially focused on small molecules and began laying out antibody capabilities in 2019. Beyond technologies like ADCs and molecular glues, in recent years we've also been deploying AI for peptide and nucleic acid drugs, as well as combinations with cell therapies such as in vivo CAR-T. This is driven partly by technical feasibility and partly by the broader trend in biomedical R&D toward modality diversification, with greater emphasis on clinical needs rather than simply pursuing technology for its own sake.
Whether it's small molecules or antibodies, as long as it can treat disease, people will use it. I think this also aligns with commercialization trends.
Third, exploring adjacent areas beyond biopharmaceuticals. We've begun branching into consumer products, cosmetic active ingredients, food and health supplements — any area where AI and molecular design can intersect, we're actively exploring.
Fourth, building data moats. Since 2019, XtalPi has been deploying its own automated experiment clusters. We will further expand the scale of data collection to establish a sustainable data competitive advantage.
Additionally, in chemistry, AI-assisted prediction accuracy for common pharmaceutical chemical reactions has already reached 80% to 90%. We're also extending our medicinal chemistry capabilities into chemicals, materials, new energy, and other fields.
Ma Rui: What proportion of services versus products will XtalPi's business comprise going forward? Will we eventually develop our own drugs?
Lai Lipeng: Not in the foreseeable future. XtalPi will remain fundamentally a technology platform. Unlike traditional drug service companies, beyond providing standardized services, we're particularly adept at solving the "hard problems" — the kinds of challenges that everyone in the industry currently finds extremely difficult.
Take solid powder transfer in chemical automation as an example. Those of you with chemistry backgrounds probably know that liquid transfer is relatively straightforward. But solid powder transfer, especially in laboratory settings that demand high flexibility and adaptability, is actually very difficult to handle.
Experiments involve hundreds of thousands of different powder components. They're expensive, and their properties — particle size, flow characteristics, morphology — vary enormously. Yet this is often a critical step before high-throughput experimentation. Having humans perform solid powder weighing and transfer represents not only a major efficiency bottleneck but also precision challenges. The pain point is clear.
Because XtalPi has very strong R&D capabilities in robotic automation, we've solved this problem well, and the corresponding technology has gained recognition from major pharmaceutical companies.
On the product side, our main strategy is co-development with partners. We contribute early-stage technology and exchange our R&D output capabilities for future product rights. Subsequent steps like clinical trials are handed over to pharmaceutical partners more experienced in that type of research.
02. Market Prospects and Capability Boundaries of Large Models in Biomedicine
Ma Rui: Next I'd like to discuss models. Over the past two years, AI's spillover into biomedicine has significantly advanced models in this space — for instance, introducing Transformer and Diffusion models into the field, which contributed substantially to the AlphaFold series' progress. I'd like to ask Dr. Lai: what phenomenon-level models from the past two years do you think deserve our attention? And where roughly are the capability boundaries of large models now? Which problems have been relatively well-solved, and which remain beyond what models can currently address?
Lai Lipeng: Overall, I believe the space for AI and large models in biomedicine remains enormous, with correspondingly large commercial opportunities.
Returning to what we discussed earlier: human comprehension of the microscopic world is extremely limited. At least for me, I can imagine how tens of thousands of people might run around a plaza, but I cannot imagine how tens of thousands of proteins interact within a cell. So AI models have primarily played a role in drug design over the past decade — I think this has been relatively well-utilized.
This includes the small-molecule design and protein design I mentioned earlier. After AlphaFold emerged in 2018, people believed AI could perform well in proteins, this biological "natural language," and reality has borne that out. But there's still vast room in this domain.
Let me use my own experience playing pool as an analogy. There are roughly three stages: at first you know nothing, can't even hold the cue properly, can't hit the ball accurately at all. At the intermediate stage, it's "brute force miracle" — you can mostly aim correctly, and if you hit hard enough, something will go in. But to truly excel in the future, you need more precise practice and cue ball control.
This is exactly like AI models in drug design. Compared to David Baker's general protein design algorithms, our internally developed algorithms targeting specific disease domains show very significant improvements. This isn't because we've done substantially more work on data — we use the same datasets — but because the algorithms are more refined, with more biological knowledge embedded, so performance and applicability improve greatly.
So even though people have felt that AI has already performed very well in molecular design over the past decade, with many commercialization cases, this direction still holds growth opportunities.
What I think is harder going forward is the biology portion, including processing large amounts of biological data and using AI to help clinical success.
If you're familiar with this field, the clinical success rate for drug development — multiplying the probabilities of Phase I, Phase II, and Phase III — may be less than 10%. Over 90% fail. Before market launch, 75% of R&D costs are in the clinical stage, so this represents a potentially much larger market.
But this is difficult for two reasons: first, relatively limited historical data accumulation; second, and more fundamentally, the iteration speed of data is too slow. The further you are from clinical stages, the faster the data feedback; the closer to clinical stages, the slower the feedback.
This is also how we at XtalPi internally assess whether AI can be applied in new domains: it's not about how much data already exists, but how fast data feedback loops are. This is why AI applications at the clinical end remain relatively limited.
So going forward, whether in terms of our own resource allocation or from an investment perspective, on one hand we need to look at preclinical sub-modalities — cyclic peptides, oral peptide drugs, small nucleic acids, and so on. You have to break problems down very specifically while also having distinctive specialized teams; then I think there are commercial opportunities.
The second is the larger unsolved problems. In bio-AI + biology, if someone can use experimental methods, data channels, or algorithmic iteration to increase the speed of data collection and feedback, then I think such technologies and products have tremendous value.
Ma Rui: Dr. Lai's point is that more vertical, more refined models may outperform general models. Additionally, we can't look only at molecular design — AI + biology involves target selection, clinical strategy, and so on. If this domain sees more and faster data iteration, new opportunities may emerge.
I see roughly three paths for models:
The first is AlphaFold-related models that can predict molecular or intermolecular conformations and structures.
The second is all-atom models, such as the current RF Diffusion3, designed with extreme granularity down to the atomic level. When AI learns at this level, it can genuinely learn some physics; it's not purely data-driven.
The third is large models that directly combine genomic sequencing data, transcriptomics data, and so on, to make certain predictions after training.
So I'd like to ask Dr. Lai: where do you think models are heading next? For future models, what will be particularly important?
Lai Lipeng: I should note that everything I'm saying today represents personal views. Here's how I see it, in two phases:
In the near term, due to data scarcity, there will definitely be development of AI models reinforced with specialized domain knowledge. In molecular design, these are physics-informed models, including all-atom models and even those guided by quantum chemistry calculations. In biology, these are models enhanced with biological knowledge.
For example, three years ago XtalPi developed an antibody design model using methods that now look somewhat dated — similar to BERT. As you know, BERT masks roughly 30% of information during language model training, then has AI predict and reconstruct the full language. Through this process it learns the language's structure.
But we found this approach didn't work well for antibody design. Without changing the model architecture much, we simply increased the masking ratio from 30% to 70%, and changed random masking to targeted masking in the most important antibody regions. This uses biological knowledge to tell AI: this is the hardest part, you should focus more learning effort here. The model's performance improved dramatically.
So I think when data volume is still insufficient, this kind of expertise-guided model training may be more important.
On the other hand, long-term, Cathie Wood once predicted in a report that biological data will eventually exceed internet corpus data. I believe there will be an inflection point: when data becomes sufficiently abundant and we have appropriate models to observe the scaling law of this data growth, we can extract underlying rules through unsupervised large model learning rather than relying on humans to tell the model what the rules are.
So large models will definitely play a major role in the future. Something similar happened over a decade ago: in 2012, AI surpassed human capability in image recognition. One of the key technical points was that before 2012, humans told AI what was most important in images. But in 2012, with the development of convolutional neural networks, humans no longer needed to define rules — AI could extract key information from images just by looking at pixels.
Returning to what I said earlier: human capacity for processing complex systems is indeed limited. So the transformation brought by AI and AI for Science may far exceed in value the 2012 milestone of convolutional neural networks surpassing humans in image recognition.
03. "Failed Data Is Valuable"
Ma Rui: For AI, the three most important things are compute, algorithms, and data. On the data side, XtalPi has AI-powered autonomous experimental platforms with many robots that can generate massive amounts of data — data that's particularly important for biology, chemistry, materials science, and especially AI+biology or AI+drug discovery. So I'd like to ask Dr. Lai: is there room in this space for something like a Scale AI opportunity? Also, what kind of data do we most need right now? And in the near future, can we actually get it?
Lai Lipeng: This is something we've been thinking about too. As I mentioned earlier, data will likely be the most critical strategic resource in the next three to five years. My personal view is that the development of new productive forces driven by industrial and technological revolutions usually stems from changes in the means of production.
Lai Lipeng, co-founder and Chief Innovation Officer of XtalPi
The shift in means of production for AI actually began back in the 1950s and 60s when machine learning was first proposed. There was an idea that everyone saw at the time — and in retrospect, it was correct — that failed data is valuable.
Before machine learning developed, much research only published correct or successful results. But in science, successful data is actually quite rare. Vast amounts of failed data were discarded, with no good tools to make use of them. When new production tools emerged — machine learning and deep learning methods — the value of this failed data was finally recognized.
Every productivity revolution is essentially about rediscovering the latent potential of existing means of production. The steam engine discovered new value in thermal energy, which had always existed in nature. Then came petroleum energy — oil had always been there too. Then semiconductors — silicon had always existed. Now, in the information age or AI era, failed data from past research has always been there; we're just in the process of rediscovering it.
There's a common saying: "garbage in, garbage out." Personally, I don't fully agree. Joking aside, even in investing, waste recycling is a very hot sector. So with good technology, data we previously considered garbage can be put to excellent use.
So how might we utilize data in the future? There are a few key points:
First is what XtalPi is doing itself: building on existing data processing methods, using automation and robotics to unify data standards and reduce the cost of data collection.
Second is what XtalPi hopes to do through partnerships or investments: exploring new tools and methods for data acquisition — such as multi-omics technologies, rapid DNA/RNA synthesis, and novel clinical testing technologies. These can bring about "data dimensional upgrading," surfacing incremental data that was previously invisible.
The third opportunity, I think China has a major advantage in. We have massive numbers of hospitals and huge patient populations, with reasonably good accumulation of clinical data and biological samples. I believe that through reasonable, compliant mechanisms for collaboration among research institutions, hospitals, government, and enterprises, these clinical resources can be fully utilized.
So back to the core question: how exactly do we use data? What kind of data is most valuable?
There's no consensus yet, because AI for Science — especially AI for Life Sciences — is still so new. We haven't yet figured out "what data is most useful for AI."
Take image recognition: we used to manually define "distance between eyes, length of nose" as inputs, then later discovered that raw pixel inputs worked better. Later, infrared detection data might also become inputs — our understanding keeps evolving.
Biology is even more like this. Previously in drug development, whether a drug actually works in animals or humans still depended on biologists' professional expertise.
So in the future, could we rely less on experts and let AI extract indicators more correlated with clinical outcomes or post-market drug effects? I think this has enormous commercial potential, but we're still not sure what data would enable AI to achieve this.
Against this backdrop, the most valuable data right now is data with high consistency, high standardization, and the ability to enable rapid collection and feedback.
Because when AI is learning, humans are also figuring out what data to feed it, and feedback speed is critical in that learning process. For now, the most important, the best data is high-quality data with fast feedback loops.
Currently, relatively accessible data includes imaging data, transcriptomics data, and the chemical synthesis data we're generating ourselves. Next, once proteomics data costs come down and throughput goes up, it may become the next generation of core data. Those are my rough thoughts.
04. Breaking Consensus: How AI Enables Drug R&D
Ma Rui: Dr. Lai just mentioned three keys to data collection: first, fast, cheap, and standardized; second, data most directly related to human disease; and third, failed data.
Digitalization is extremely important. Actually, Feng Shu (Li Feng of FreeS Fund) keeps asking us internally: AGI is starved for data — what data do you think might be most useful?
I've summarized it into three categories:
First, data related to the real physical world, including embodied interactive manipulation, spatial intelligence, and 3D reconstruction data.
Second, highly relevant to our conversation today: microscopic data needed for AI for Science, such as molecular interactions.
Third, brain-related data. If brain-computer interface technology matures, enabling high-density collection of neural data correlated with human behavior, there will likely be many new discoveries — after all, technologies like CNNs and Transformers were initially inspired by advances in neuroscience.
So for AGI, models are indeed currently constrained by data, and we may need to source data from these areas I just mentioned.
The next question: I'd like Dr. Lai to introduce XtalPi's pharmaceutical practice, especially regarding new modalities. If you could also touch on how AI helps with peptides, small nucleic acids, or RNA therapeutics, even better. After all, AI drug discovery ultimately needs to produce actual drugs.
Lai Lipeng: At the end of the day, this is about making money — AI has to land somewhere. I mentioned XtalPi's own layout earlier. In Beijing, we have a separate division called XtalPi Innovation Center. Each year, XtalPi allocates part of its budget to exploratory projects that we believe have significant commercial potential. One very important direction is more modalities.
On one hand, this is market demand, commercially driven: large MNCs tend toward diversifying their modalities in pipeline selection. In 2022, Pfizer published an article summarizing their clinical development strategy progress from 2016-2021, noting that "diversifying drug modality selection" was a key factor.
On the other hand, there's technological support: AI's powerful imagination and molecular design capabilities can help us build multi-modality design capabilities. Because the further you go toward biologics (from antibodies to peptides, then proteins, small nucleic acids), the algorithmic logic is actually consistent. They're all essentially linear sequences. So with good AI infrastructure, expanding into new modalities is quite feasible.
For example, when people used to evaluate or invest in drugs, they'd say small molecules, antibodies, and peptides each have distinct characteristics: small molecules have good PK (pharmacokinetics), easily penetrate membranes and enter cells, can target intracellular targets, with strong delivery capability and stability; antibodies have high precision, don't easily go off-target, hit exactly what you aim at; peptides fall in between — smaller than antibodies, potentially better cell penetration, safer than small molecules, but less stable than small molecules. This is industry consensus — but I think AI's emergence can break this consensus.
For instance, internally at XtalPi, we've used AI to analyze virtually all protein-protein and small molecule-protein interactions. Peptides are unstable because they're composed of 20 natural amino acids; the industry often relies on manual modifications to optimize peptide properties toward small molecule advantages. After analyzing all these interactions, we used AI to generate roughly 2,000 non-natural amino acids — 100 times more than natural amino acids. By introducing these non-natural amino acids, we retain peptide drugs' advantages while absorbing small molecule drugs' strengths, with application potential in areas like brain drug delivery and oral administration.
Take current weight-loss drugs like semaglutide and tirzepatide — mostly injectable formulations. Not just XtalPi, but the entire industry is watching whether we can make better once-weekly or once-every-two-weeks oral weight control medications. This will depend heavily on AI breakthroughs in this area.
When our team looks at many drug R&D problems through an AI lens and re-examines them, there are many new discoveries. So I think the commercial opportunities are truly abundant.
Another example with small nucleic acids. Standard small nucleic acid design has two steps: one is sequence design — based on the target mRNA, selecting appropriate targets and designing roughly 21-23 base small nucleic acids. But small nucleic acids are unstable in vivo and decompose quickly, so they don't achieve therapeutic effect. Thus the second step is chemical modification.
Traditionally, sequence design and modification design are done in two separate steps. Step one: find the best sequence. Step two: match the best modification to that sequence.
This seems fine, but to use a simple analogy — marriage. The best man and the best woman don't necessarily make a happy marriage.
So to get a good nucleic acid drug, these two optimization steps should actually be combined into one, not done separately. But traditional workflows are linear, and relying on human expertise makes it very difficult to balance both simultaneously.
So XtalPi did something very direct internally: put sequence design and modification design into the same generative model, completing both in one step.
This brings two benefits: one, we found nucleic acid molecular sequences superior to those in existing clinical research; two, both sequences and modifications have excellent novelty. As you know, many modification-related patents are dominated by large companies like Alnylam. Our approach both breaks through patent constraints and finds better molecules.
Therefore, I think bringing an AI perspective to re-examine existing drug R&D workflows yields many new discoveries, especially in new modalities where AI applications aren't yet particularly hot — there will be excellent opportunities for real impact.
05. Finding New Opportunities at the Intersection of AI and Other Fields
Ma Rui: Very well. Dr. Lai's remarks have given us a lot of confidence. The marriage analogy was vivid — optimizing a joint probability with the ultimate goal as the objective function, rather than simply optimizing a single parameter. Next, I'd like to ask: what is XtalPi's layout in AI + materials? How do materials differ from pharmaceuticals and biology? What trends do you see here?
Lai Lipeng: Materials, energy, and some agriculture-related directions are all new fields for XtalPi. We can see clear similarities and differences.
The common ground is technical universality. At least in the microscopic world, XtalPi has identified three areas where we can quickly deploy and demonstrate value.
First, new molecular structure design — this is familiar territory, designing novel compounds and new proteins.
Second, formulation optimization. Nothing in the real world is a single substance; everything is made from different formulations. In pharmaceuticals, this corresponds to drug formulations; in materials, it corresponds to formulation selection.
Third, process development. Once you have the core material and formulation, you need to solve the lab-scale manufacturing problem — turning something into a physically testable object in the lab, and scaling up from lab bench to pilot to production. Our technology can migrate relatively quickly across these stages.
The differences lie in validation speed and data feedback loops across industries. Each industry has its own R&D chain and commercialization process. How you break down the R&D chain into commercialization-relevant stage milestones, and then build rapid data feedback loops — this varies enormously by field.
For example, lab validation in materials tends to be faster. We previously worked on some photovoltaic materials projects where we could achieve feedback on over 100 samples per day in the lab. In pharmaceuticals, one reason XtalPi invested in chemical automation is that synthesizing a single compound used to take a month and a half.
So from a technology deployment perspective, the definition of stage milestones and the speed of data feedback differ across fields. When XtalPi enters new areas like materials, we pay particular attention to data iteration efficiency and feedback speed.
Ma Rui: So AI is critically important in materials, synthetic biology, and related fields. Given time constraints, let me offer a brief closing.
As we've heard today, AI-driven drug discovery is already bearing fruit, and there are enormous opportunities in materials, chemistry, physics, and beyond. Combining these innovation opportunities with China's strong industrial chains could further amplify productive capacity.
For instance, AI pharmaceuticals developed on the back of China's robust innovative drug industry chain. Two years ago, people might not have recognized China's innovative drug capabilities — but no one questions it now.
In fusion energy, AI can assist with plasma control, prediction, and reactor design. China has the potential to lead the fusion industry, given its advantages in manufacturing capacity, materials, components, and power electronics.
In biomanufacturing, we have the world's largest fermentation capacity. Combined with AI design and synthetic biology capabilities, this industry will also rise in China.
In quantum computing, AI is likewise an important variable, with promising future opportunities.
In short, finding innovation opportunities at the intersection of AI with biology, chemistry, materials, and energy — and seeing these innovations diffuse into the future industrial directions outlined in the 15th Five-Year Plan — this is the kind of opportunity, like XtalPi, that FreeS Fund aims to capture over the next decade.
"AI Industry Observations" Series
From Foundation Models to AI Companionship: What Cyclical Patterns Lie Behind the Rotating AI Hype?
How to Bridge the "Last Centimeter" of Robot-Real World Interaction?
"In AI Hardware, There's Always a New Opportunity"
Interactive Giveaway
Beyond playing chess, chatting, and writing poetry, AI can also help scientists push the frontiers of drug, energy, and materials research. If you were calling the shots, which everyday scientific challenge would you most want it to tackle first?
Share your ideas in the comments! By 17:00 on January 16, 2026, the two most creative commenters will each receive a copy of The FreeS Fund Industry Research Handbook.
FreeS Fund's Li Feng 2025 Year-End Share: The Logic and Outlook for AI Investment AI for Science: Standing at the Turning Point of a New R&D Paradigm Looking Ahead to 2026, What Innovation Opportunities Exist in the AI Industry? Is AI Healthcare Having Its DeepSeek Moment?
Virtual Cells: Can Life Be Digitally Simulated?
Star the FreeS Fund WeChat Official Account — timely business insights delivered to you