The infrastructure competition for AI-driven pharmaceutical R&D has kicked off.

When AI pushes drug design to the "minute-level", laboratories capable of validating such outputs are equally scarce.

This trend is unfolding simultaneously across the globe. On the other side of the Pacific, Chai Discovery, founded less than two and a half years ago, has reached a valuation of 3.8 billion US dollars. Its CEO holds a counterintuitive judgment: AI will not reduce the number of wet experiments, but instead may increase them, just as the demand for programmers rises rather than falls after their work efficiency improves.

Swiss startup Adaptyv Bio completed wet experimental validation of 1320 protein sequences for Anthropic's Claude. Its revenue has increased nearly 10 times in the past year, and it has just closed a 40 million US dollar Series A round. This further declares to the outside world that the smarter the model is, the more explosive the demand for validation will be.

On September 11, GenScript Biotech joined this competition with an official announcement: it plans to place 77.126 million new shares at an issue price of HK$30.50 per share, with an estimated net proceeds of about HK$2.327 billion.

Among them, 70% of the net proceeds, about HK$1.629 billion, will be used to expand the production capacity and infrastructure of the group's AIDD platform, including automated high-throughput wet laboratory facilities, related equipment and laboratory upgrades; 20% will be used for AIDD-related R&D, digital workflow integration and global business expansion; the remaining 10% will be used for general corporate purposes and operations.

Why would a company whose main business focuses on gene synthesis, protein expression and life science services invest more than HK$1.6 billion at one time in an AI drug discovery platform?

The answer may be that the competition in AI pharmaceuticals is shifting from "who can design more molecules" to "who can validate these molecules faster and more accurately".

As a result, an infrastructure competition dedicated to AI pharmaceuticals has begun.

/ 01 / AI can design more molecules, but laboratories are not yet ready

In the past few years, the most concerned link in AI pharmaceuticals has been computational design.

By inputting target, protein structure and disease information, the model can generate a large number of candidate molecules in a shorter time, and can also search for structures that were difficult to discover in the past in a larger chemical space.

But design is not the end point.

Can the candidate molecule be synthesized? Can it be expressed? Can it bind to the target? Does it have cellular activity? Is it druggable? All these questions ultimately need to be answered in wet laboratories.

GenScript mentioned at its 2026 interim performance meeting that AI models can now generate thousands of candidate molecules within a few hours, but traditional experimental processes are not designed for such scale and speed. The validation that used to take several weeks often goes through multiple disconnected steps, and experimental data may not be fed back to the AI model in a timely and standardized manner.

This creates a new mismatch.

The model is getting faster and faster, but the speed of experiments has not improved synchronously; the model generates thousands of answers at a time, while laboratories still validate them one by one according to processes from decades ago; algorithms can be iterated continuously, but experimental data cannot flow back quickly, making it difficult for the model to truly become better.

The bottleneck of AI pharmaceuticals is shifting from "whether a good idea can be proposed" to "whether a large number of ideas can be turned into high-quality data".

This is also why the infrastructure of AIDD no longer only refers to GPUs and algorithm models.

It also includes gene synthesis, protein expression, antibody discovery, automated experiments, screening and detection, data management and digital processes. Only when computational design and wet experimental validation are connected can AI pharmaceuticals form a real closed loop.

/ 02 / First develop the "shoveling-selling" business

GenScript's this fundraising provides us with a window to observe the infrastructure of AI pharmaceuticals.

According to the announcement, what the company plans to expand is not a single experimental device, but a set of automated high-throughput synthesis, screening and detection platforms.

What it aims to solve is not the single-point demand of a certain customer, but whether the platform can stably undertake, quickly validate and return standardized results when AI pharmaceutical customers bring hundreds or thousands of candidate molecules at one time.

This is not completely the same as the logic of traditional life science services.

Traditional R&D services are usually carried out around a single project: the customer puts forward demands, the service provider completes a section of experiments, and then delivers the results. The demands brought by AI pharmaceutical customers may be high-throughput, high-frequency and continuously iterative: the model generates a batch of designs, the experiment validates a batch of results, the data flows back to the model, and the model continues to generate the next batch of designs.

This means that the value of the platform is no longer just "completing the experiment", but whether the experiment can be made into a replicable, scalable and sustainably deliverable production line.

The 2026 first-half performance of GenScript has shown early signals of this trend. In the first half of the year, the company's AIDD-related business doubled year-on-year, maintaining rapid growth for three consecutive half-year periods. In addition, it stated that the Gene-to-Protein platform accounts for about 66% of the revenue of the Life Science Business Group, and has become an important pillar of business growth.

This set of data shows that AI pharmaceuticals not only brings orders to model companies, but also may transmit to upstream experimental services, reagents and consumables, automated equipment and data services.

The faster the model company designs, the larger the amount of experiments that the validation platform needs to process; the more frequently the model iterates, the more important the speed of data feedback becomes; the more candidate molecules a single project generates, the less likely customers are to rely on scattered suppliers to complete them one by one.

This is where the opportunities for infrastructure-focused companies emerge.

/ 03 / "Wet experimental validation", the next high ground for competition?

In the past, AI pharmaceuticals was more like a model competition.

Whoever has stronger algorithms, greater computing power and more data will have the opportunity to take the lead in the drug design link.

However, when more and more companies enter this field, the model itself is becoming more accessible. What is truly difficult to replicate has begun to turn to things beyond the model: where does high-quality data come from? How to standardize experiments? Are the results of different batches comparable? Can it be completed within a few days instead of several weeks?

This also explains why the "infrastructure" of AI pharmaceuticals must integrate dry and wet capabilities.

The "dry" end is responsible for putting forward hypotheses, generating designs and optimizing candidate molecules; the "wet" end is responsible for synthesis, expression, screening and validation; the digital system is responsible for converting experimental results into data that the model can continue to use.

Without any link, the closed loop cannot be formed.

If there are only algorithms without fast enough experimental validation, the candidate molecules generated by the model will accumulate in the virtual space; if there is only experimental capability without standardized data return, the laboratory is only an automated upgrade of traditional CRO; if there is only data without a continuously iterative model, the data will hardly be converted into R&D efficiency.

Therefore, the competition for AI pharmaceutical infrastructure ultimately boils down to four indicators: scale, speed, reliability and data quality.

GenScript disclosed during its 2026 first-half performance exchange that about 60% of laboratories worldwide have been equipped with AI-driven automated workstations, and the Gene-to-Protein platform can shorten the delivery cycle from digital sequences to model-available data to as fast as 4 days.

What does 4 days mean?

It does not mean that drug R&D can be completed in 4 days, nor does it mean that all experiments can be finished within 4 days. But it at least shows that the AI pharmaceutical industry is redefining the "validation cycle" — shifting from the past week-based calculation to a shorter delivery cycle.

This is also why Anthropic's protein binder experiment, the wet experimental validation of 1320 candidate sequences, was entrusted to Adaptyv Bio, a startup company. It rebuilt the laboratory into "the physical backend of AI Agent": the cost of each binding detection starts at 79 US dollars, supports direct API connection, returns machine-readable structured data, and also operates the open Proteinbase database containing negative results (which has included more than 3000 validated proteins).

Traditional wet laboratories are designed for humans and deliver results via emails and PDF reports. Adaptyv is designed for AI Agents and delivers results via APIs and structured JSON. It is not a faster and cheaper CRO, but a kind of reconstruction, which exactly catches the outbreak window of validation demand for AI protein design models. In the past year, the company's laboratory has tested more than 10,000 proteins, the number of customers has exceeded 100, and its revenue has increased by nearly 10 times; now it expresses thousands of proteins every month, with an expression success rate of over 90%.

Back in China, GenScript's HK$1.6 billion bet is exactly the local response in the same competition.

For AI-native pharmaceutical companies, the faster the experimental validation is, the faster the model iteration will be; the faster the model iteration is, the higher the efficiency of candidate molecule elimination and optimization will be. For service providers, shortening the delivery cycle is not only to improve customer experience, but also the basis for seizing long-term cooperative relationships.

The next high ground for competition arises from this.

This article is from the WeChat official account "Amino Insight" (ID: anjiguancha), author: Amino Jun, published with authorization from 36Kr.