A new change in this round of innovative drug market rally: "wet experiments" are becoming the core assets in the era of AI drug discovery.
In June, Anthropic launched Claude Science, built its own physical wet lab, and spent 400 million USD to acquire Coefficient Bio, an AI drug R&D startup founded only eight months prior. A company that built its reputation on models went all in on what model developers could achieve, then pivoted to purchase beakers and fume hoods.
In September, Novo Nordisk signed a cooperation deal worth up to 1.4 billion USD with Orbis Medicines. What it acquired was a closed-loop system of "AI design — automated synthesis — experimental feedback", as well as the experimental data generated by the system that does not exist anywhere in the public domain.
The two news items, three months apart, point to the same conclusion: In this wave of AI-driven pharmaceutical R&D, models are consumables, data is fuel, and the place that generates data — the wet lab — is evolving from a cost center into the most tangible asset on the balance sheet.
01
Why Have Wet Labs Become Core Assets?
I. Shift of Bottlenecks: Surplus of Models, Scarcity of Validation
First, let's clarify the terminology. Dry experiments refer to prediction, screening and molecular design completed on computers; wet experiments refer to hands-on synthesis, cultivation, animal testing, analysis and detection carried out in the real world. The ideal closed loop for AI-driven pharmaceuticals is the alternation of dry and wet experiments: machine design, experimental validation, data feedback, and model iteration.
A research note from Industrial Securities last year put it clearly: the rate-limiting step of current AI drug R&D is not how fast the model runs, but how fast wet experimental data can be generated.
The reason is not hard to understand:
The model side is a fully competitive market. Tools like AlphaFold have pushed the marginal cost of structure prediction to nearly zero. Generative models can output millions of candidate molecules every day, and the foundational models adopted by different players are largely similar, leaving no room for real differentiation.
The data side, by contrast, is an extremely scarce market. Experimental data in public literature is inherently biased: all published data are successful cases, while failure data is barely recorded; conditions in different labs are not compatible with each other; public data for cutting-edge directions such as macrocyclic peptides and covalent inhibitors is almost non-existent. The CEO of Orbis put it more bluntly: AI models only work when fed with real experimental data, so we have to generate data by ourselves.
As a result, the entire bottleneck has shifted. You can design 100 million molecules, but a lab can only synthesize 50 of them a week. Models are becoming less valuable, while data that makes models smarter is gaining value. The capability to generate data, including automated labs, high-throughput synthesis and screening, standardized data collection, and more, has become the scarcest link in the entire industrial chain. Anthropic's acquisition of wet labs essentially aims to build independent control over data production.
II. Two Counterintuitive Observations: AI Does Not Reduce Experiments, It Expands Their Scale
The first counterintuitive point is that common sense would assume AI efficiency gains reduce demand for experiments, but the industry's actual accounting shows the opposite result.
AI has brought the marginal cost of molecular design down to nearly zero. Instead of reducing the number of experiments, it has increased the number of candidate molecules entering the validation pipeline by several orders of magnitude. Previously, only 3 to 5 molecules per project were worth testing, but now hundreds or thousands of molecules per project are queuing up for synthesis, activity testing, and safety assessment. The cheaper design becomes, the more expensive validation gets.
This is why when NVIDIA and Eli Lilly co-built a lab, their core metric was not model accuracy, but "one iteration every two hours" — the rotation speed of the dry-wet closed loop, rather than the standalone speed of dry experiments.
The second counterintuitive point relates to the evolution of business models.
AI pharmaceutical companies are shifting from "telling stories to raise financing" to "exchanging platform capabilities for milestone payments". Orbis has raised a total of 133 million USD in financing, and secured a milestone framework worth up to 1.4 billion USD, with a leverage ratio of more than 10 times. But every milestone payment is tied to a molecule that has been validated by wet experiments.
In other words, the monetization capability of an AI platform is physically constrained by the throughput of the experimental pipeline behind it. The company that runs the dry-wet closed loop faster will be the first to turn its AI capabilities into revenue.
III. Three Criteria for Judging High-Quality Wet Lab Assets
Not all labs count as assets. Based on the industrial logic of this wave, we have summarized three judging criteria:
First, throughput. The data output of manual labs cannot feed models. Only automated, standardized production lines can generate data that models can process. Judging indicators include verifiable disclosures such as the number of robotic workstations and weekly experimental throughput.
Second, data ownership. If you run projects for clients and the data belongs to the clients, the lab is just a production capacity asset. Only when you build your own platform to generate proprietary data can the data be precipitated as the company's asset. Even for the same wet experiment work, different legal ownership leads to completely different valuation logic.
Third, closed-loop integration. Dry experiments and wet experiments must operate in the same feedback loop, rather than being two separate silos where "the AI department does design and the experimental department takes orders". The judging criterion is whether the organization has tightly integrated model iteration and experimental scheduling. This is the hardest criterion to meet, and also the key part that Anthropic chose to fill through acquisition.
02
What Are China's Core Wet Lab Assets?
They can be divided into three categories by capability scope: platform closed-loop type, experimental capacity type, and upstream supporting type.
I. Platform Closed-Loop Type: Generate Data Independently
XtalPi is the asset with the highest platform purity among public companies in A-share and Hong Kong stock markets. It has two core foundations: quantum physics computing and AI models for dry experiments; a robotic lab with more than 300 automated workstations for wet experiments, covering synthesis, crystallization, and formulation testing. XtalPi has compressed the traditional 3 to 4 week experimental iteration cycle to about 6 days, independently running more than 10,000 experiments per week, and precipitating more than 200,000 reaction data entries per month, 80% of which are negative failure samples that cannot be found in public literature.
Its clients cover 17 of the world's top 20 pharmaceutical companies. 2025 was the first year of scaled revenue: total annual revenue reached 803 million RMB, up 200% year on year; revenue from drug discovery solutions hit 538 million RMB, up more than 300% year on year.
We also need to point out the risks: the total potential amount of its cooperation with DoveTree is about 6 billion USD, but that is the total framework value including R&D, clinical and sales milestones, and no amount has been recorded in the income statement yet. When tracking this company, the key metrics to watch are client repurchase rate and actual milestone payment arrivals.
Chengdu LeadGene takes a different closed-loop path: instead of relying on a legion of robots, it relies on chemical libraries. Its core asset is DEL technology, which attaches a DNA barcode to each of a massive number of compounds, forming a trillion-scale physical molecular library that can be screened in parallel, supplemented by FBDD/SBDD structural biology screening. This model naturally generates data: every screening produces a complete map of interactions between the target and trillions of compounds, and the company has accumulated data for hundreds of targets.
In 2025, it increased its capital and took a controlling stake in Moshine Intelligence through a fund it participated in setting up — the latter integrated more than 100,000 data sources to build a structured pharmaceutical data cluster; in April 2026, it further built a computing power base in partnership with Chengdu Supercomputing Center.
According to its financial report, 2025 revenue hit 526 million RMB, attributable net profit reached 110 million RMB, gross margin stood at 54.24%, DEL and structural screening together accounted for about 77% of revenue, and milestone revenue began to be recognized. Its capability boundaries need continuous observation: its self-developed pipeline is still in the early stage, and the renewal rate of orders from large overseas pharmaceutical companies will be the key to its success.
Hopax Pharma sits between platform providers and CROs. It has self-developed the DiOrion large drug design model and PR-GPT tool, claiming to realize the closed loop of "AI design to synthesis and testing in its own lab". It has delivered 65 candidate drug molecules in total (57 of which are first-in-class), serving 53 CADD/AIDD clients. Its wet experiment scale is not at the same level as the previous two companies, and the company itself acknowledges that AI currently acts as "an enabling tool to improve service efficiency and delivery capabilities".
We classify it into the platform category because it has the organizational structure for closed-loop operation; what it needs to prove is whether AI can grow from an efficiency tool to an independent, identifiable revenue line. Tracking its development is equivalent to observing whether the "AI-augmented CRO" business model is viable.
II. Experimental Capacity Type: Turn Lab Scale Into a Moat
Platform-type companies generate data to feed their own models, while the logic of capacity-type players is more straightforward: when the flood of AI-designed molecules arrives, experimental capacity itself is a right to collect rent.
WuXi AppTec is the giant in this segment. Its 2025 revenue from continuing operations hit 43.4 billion RMB, with hand-held orders of 58 billion RMB at the end of the year, up 28.8% year on year — no matter whether the pipelines of AI companies succeed or not, these orders will turn into revenue from synthesis, screening, testing and production. Two parts of its wet lab assets are often underestimated: first, one of the world's largest private chemical databases, with more than 6.5 million compound structures and their reaction records, naturally precipitated from decades of experimental capacity accumulation; second, the "AI design — robotic synthesis — intelligent validation" production line at its Wuhan base, which can complete thousands of chemical reactions per day. Its TIDES business revenue reached 11.37 billion RMB, up 96% year on year, perfectly positioned in the GLP-1 and nucleic acid drug tracks, the two sectors that consume the most wet lab capacity.
It has also invested in XtalPi and Insilico Medicine, and previously established a joint venture with Schrodinger. Its capability boundary: AI does not have an independent line item in its financial statements, and attributing order growth to AI is overinterpretation. Its correct positioning is the "utility infrastructure for the AI era".
Pharmaron has a similar structure to WuXi AppTec but smaller in scale. Its differentiated move is to take a controlling stake in Haixinzhihui, an oncology AI company, to integrate high-quality compliant patient data into its clinical service system. More than 20,000 R&D personnel are another form of wet lab asset. Key observation keyword: overseas biotech client reserves — the first group of payers for AI pharmaceuticals are exactly those algorithm companies that do not have their own labs.
Medicilon is a sample in the pre-clinical validation segment. Pharmacology, toxicology and safety assessment are the statutory checkpoints for AI-designed molecules to move forward, and every batch of AI-designed molecules must pass through these checks. The company has built the MAIDD platform, launched an AI pre-clinical research service platform in Shanghai, supplemented by new methodologies such as organoids. More than 650 R&D projects it has participated in have obtained IND approvals. When AI pharmaceuticals enter the clinical validation stage, its order structure will directly expand.
Joinn Laboratories deserves a separate mention: this year it signed a strategic cooperation agreement with Insilico Medicine, to build a collaborative ecosystem covering non-clinical research, experimental animal resource support, and AI life science model development. Safety assessment plus model animals is the most asset-heavy segment in wet experiments, and also the segment hardest to be replaced by AI. After all, there is no shortcut to measuring toxicity, it can only be observed in animal subjects.
III. Upstream Supporting Type: The Shovel-Sellers' Shovels
Pharmablock is a global leader in molecular building blocks. The vast majority of new AI-designed molecules require custom blocks to build their skeletons, making building blocks a rigid-demand consumable in the AI synthesis chain. Its 2025 revenue hit 1.974 billion RMB, of which CDMO revenue reached 1.124 billion RMB, as its business expanded from selling building blocks to integrated services.
HuaAn Pharmaceutical follows a similar path: it has a reserve of nearly 100,000 molecular building blocks and tool compounds, operates the MedChemAI platform, and adopts a dual-wheel model of front-end consumables plus back-end small-scale CRDMO services.
Moving further to the biological side, recombinant protein and antibody expression reagents from ACROBiosystems and Sino Biological are validation materials for AI macromolecule design — models can generate antibody sequences, but expression and affinity testing still rely on protein factories.
The common feature of this category: small transaction value per order, high repurchase rate, no exposure to the success or failure of specific pipelines. It is the segment with the highest performance certainty and the lowest growth elasticity across the entire industrial chain.
03
Three Deductions
Deduction 1: Value distribution shifts to the validation end, and "experimental throughput" will become a valuation parameter.
With model homogenization, data scarcity, and validation as the rate-limiting step, profit distribution will inevitably shift to the segments that hold experimental capacity and data assets. The primary market has already voted with their feet, and the re-rating in the secondary market is only a matter of time.
In terms of valuation methodology, "experimental throughput" will gradually be incorporated into valuation models, just like how "production capacity" was a core metric for CDMOs and "user duration" was for the internet industry in the past. In the future, investors will ask three questions: how many proprietary data entries does this company generate per month? How long does one round of dry-wet closed loop take? Who owns the data ownership? The valuation premium corresponding to these three questions will be larger than the current market pricing difference based on whether a company has the "AI concept" label.
Conversely, companies that claim to be AI pharmaceutical players just because they "deployed CADD tools" will lose their valuation premium in this wave, because tools are purchased assets, not proprietary assets.
Deduction 2: China's competitive advantage needs to be re-priced, shifting from the model narrative to the chemical capacity narrative.
In this global competition, China's real bargaining chip is not models — generative model diffusion has no national borders, and the foundational model capabilities of top labs in China and the US are rapidly converging. The real bargaining chip is chemical capacity: the place with the lowest molecular synthesis cost, the highest throughput, and the highest density of engineers in the world is China.
The "two-hour iteration" that NVIDIA and Eli Lilly pursue, and the "delivery of thousands of molecules in several weeks" that companies like Orbis aim for, all physically point to the Chinese supply chain. The extreme compression of R&D costs driven by payer price pressure leads to the optimal solution of placing wet experiments in China.
The global division of labor for AI pharmaceuticals may repeat the story of the consumer electronics industry: design in Silicon Valley, manufacturing in the Yangtze River Delta.
Deduction 3: The verification window falls between 2027 and 2029.
The logic of wet lab assets will ultimately be verified by the clinical success or failure of AI-designed drugs. BCG statistics show that the phase I success rate of AI-designed drugs exceeds 80%, significantly higher than the historical level; the phase II success rate is about 40%, no different from the industry average.
Put in plain language: models are good at "finding molecules that can enter clinical trials", but have not yet proven their capability to "find drugs that can be launched on the market". From 2027 to 2029, the first batch of AI-native pipelines will release concentrated data, which will be the coming-of-age ceremony for the entire track, including the wet lab narrative.
Before that, three leading indicators need to be tracked.
First, order quality: in the contracts of platform companies, whether the proportion of milestone recognition is rising or the proportion of service fees is rising. The former represents deep binding with shared risks, while the latter is still shallow cooperation based on outsourcing relationships.
Second, data capitalization: whether any company turns its proprietary experimental data from a cost line item into a tradable asset, through data licensing, data equity contribution, or joint training, which will be a signal of valuation system transition.
Third, talent flow: whether top computational biologists are moving to labs or to