The next Claude Code in the AI field is likely to be born in a laboratory.

Over the past two years, AI has taken the lead in building a high-value commercial closed loop in the coding sector.

Overseas institutions estimate that in the next 12 months, AI Coding is expected to generate a combined annualized revenue of about 1.2 trillion US dollars for OpenAI and Anthropic. Coding is one of the core scenarios driving its commercial growth.

The reason why Coding has become the first professional scenario for large models to realize value is not only the high cost of programmers, but the fact that code naturally has a set of "verifiers" including compilation, testing and operation results: every piece of code generated by the model can be immediately judged right or wrong, and the capability improvement can be directly converted into productivity.

However, as Coding gradually becomes a market consensus, investors have begun to look for the next high-value entry point for AI commercialization.

Scientific research is likely the direction closest to the answer.

Data from the European Commission shows that the 2,000 companies with the highest global R&D investment in 2024 invested a total of about 1.446 trillion euros. Industries such as life sciences, new materials, chemical engineering and energy invest huge sums of money in R&D every year, but a large amount of resources are still consumed in inefficient trial and error, fragmented processes and repeated experiments.

From improving digital productivity to improving scientific productivity, AI for Science (AI4S) is trying to replicate the commercialization path of AI Coding. The difference is that the verifier of Coding is the compiler, while the verifier of AI for Science is real experiments.

At present, a number of leading technology companies including NVIDIA and Anthropic are accelerating their layout of AI for Science. XtalPi is the relatively front-runner in the domestic market.

On July 29, XtalPi officially released XtalPi Science, an open intelligent R&D platform for scientific research scenarios.

It is a native AI for Science operating system that connects large language models, vertical scientific models and automated robotic laboratories. It attempts to integrate the capabilities previously scattered in models, data, software and experimental equipment into a set of reusable scientific research infrastructure that can be applied across industries.

At the same time, the "Open Ecology Alliance for AI for Science", jointly initiated by 27 enterprises and institutions, was officially established at the press conference. The first batch of members cover pharmaceuticals, materials, embodied intelligence, universities and scientific research institutes, including industrial and academic participants such as Sansheng Pharmaceutical, Junshi Biosciences, JAKK, Baicheng Pharmaceutical, Danaher, MindSpire, Motiv Robotics and The Chinese University of Hong Kong (Shenzhen).

In this industrial wave that is still in its early stage but whose certainty is growing rapidly, XtalPi has undoubtedly seized a key position in advance.

To judge whether an AI4S enterprise is leading, we should not only look at the score of a single model, but also see whether it can be responsible for the final scientific research results.

At present, most participants in the industry still cut in from a certain link of models, scientific software, experimental instruments or R&D services. The difference of XtalPi is that it has begun to connect vertical models, professional workflows, robotic laboratories, real project data and industrial customers into a complete closed loop.

Model capabilities can be purchased, computing power can be invoked, and software can also be developed quickly. However, experimental infrastructure, long-accumulated failure data, and experience formed by continuous operation of real R&D processes are difficult to replicate in the short term.

/ 01 / AI for Science, the next AI Coding

Traditional scientific research is still a production method highly dependent on expert experience.

A large amount of knowledge is scattered in papers, patents, databases and experimental records, and the R&D process relies heavily on the experience of a small number of experts. A single experiment may take weeks or even months, while a large number of failed results and abnormal data are often stored in personal computers or experimental logs, which are not structurally saved, let alone effectively utilized in the next round of R&D.

What AI for Science wants to change is not just a certain link, but the entire scientific research process.

In the past, the exploration of large models in the field of scientific research mainly stayed in the digital world, including searching papers, analyzing data, invoking scientific software, and putting forward seemingly reasonable molecular structures, synthesis routes and material formulas.

But whether these answers are valid ultimately has to be verified in the physical world. This has always been the core bottleneck faced by scientific Agents: they can design experiments, but cannot be responsible for the experimental results. In other words, the scientific field has long lacked a feedback system similar to a code compiler.

The core of AI4S is precisely to enable AI to verify itself.

If the compiler constitutes the feedback system of AI Coding, then the DMTA closed loop of "Design-Manufacture-Test-Analysis" is the "scientific compiler" of AI4S in the physical world: it allows every model hypothesis to be tested by real experiments, and feeds the experimental results back to the next round of prediction.

This is also becoming a direction that all parties in the industry are betting on together.

In June this year, Anthropic launched Claude Science for researchers, which tries to become a new entry point for researchers to access data and software by integrating tools such as literature, genomics and chemical computing. NVIDIA released the BioNeMo Agent Toolkit and cooperated with Thermo Fisher to further extend the Agents to laboratory equipment.

The moves of overseas giants verify the industrial direction of AI4S, but their entry path is still mainly extending from basic models and computing tools to experimental links.

The difference of XtalPi is that when this direction begins to become the consensus of the global technology industry, it has accumulated nearly ten years around vertical scientific models, robotic laboratories and real R&D projects.

According to the company, this is the world's first native AI for Science operating system that deeply couples large language models (LLM), professional vertical models and large-scale automated robotic laboratories.

Among them, Genius Agents undertakes the orchestration of the entire R&D process. It can make plans around specific scientific research goals, break down complex tasks into different steps, invoke corresponding professional models, scientific software and experimental capabilities, then summarize the results and precipitate project knowledge.

In order to enable Agents to truly enter scientific research organizations rather than just stay at the model demonstration level, XtalPi further breaks down the DMTA process into hundreds of specific R&D intents, and precipitates the tacit experience accumulated by scientists over a long period of time into professional workflows that can be executed standardly and invoked repeatedly.

In this system, AI is responsible for large-scale search, scheme generation and process execution, while scientists retain the final judgment right over scientific research goals, technical routes and key decisions.

This collaborative paradigm of "AI large-scale exploration + scientist key decision-making" can reduce handover stagnation and decision-making waiting in the R&D process, compress some verification cycles that originally took weeks or months to "days", and independently connect long-range tasks such as literature study, structure-activity relationship analysis, molecular design, ADMET prediction, synthesis route planning and automated experimental testing, extremely compressing the R&D verification cycle that often takes weeks or even months to "days".

At present, this system has been running intensively inside XtalPi and has undergone multiple rounds of real R&D verification, becoming an indispensable productivity tool in scientists' daily R&D work.

More important than efficiency improvement is that the system begins to continuously produce the data that AI really needs.

Every successful, failed and abnormal experiment can be completely recorded and become new training data. Compared with public papers and patents, these data generated by the real R&D process are scarcer and more difficult for competitors to obtain.

Public literature usually tells AI "what succeeded", while the failure data in real projects helps AI identify "why it failed". What is even scarcer is not the final result, but the full-process data from scientific research goals, model prediction, route selection, to experimental process, abnormal situation and result feedback.

At present, XtalPi's intelligent autonomous laboratory generates more than 50,000 reaction yield data and 300,000 process data every month, and has accumulated more than 500,000 real experimental records, including a large number of failure samples that have been missing in public databases for a long time.

These data have already been reflected in model capabilities. In the evaluation covering 350 real industrial molecules, the chemical hallucination rate of SureRXN dropped to 4.6%, the accuracy of the first synthesis route reached 51.7%, and the prediction accuracy of failed experiments reached 81% to 89%. In real service projects, the 5 to 10 experimental iterations required for target molecule synthesis have been shortened to an average of 1.19 times.

The industrial significance of these figures lies in the ability to reduce the number of expensive physical experiment rounds, and shift R&D from experience-dependent repeated trial and error to accurate prediction supported by real data.

The more experiments there are, the richer the data; the richer the data, the more accurate the judgment of models and Agents; the more accurate the system is, the more scientific research tasks will be attracted to the platform. This is the data flywheel that XtalPi is really trying to build.

XtalPi's moat is therefore not a static data set, but a system that continuously produces high-quality data. Models can be purchased and invoked, but failure data and physical closed loops cannot be built overnight.

/ 02 / Re-recognize XtalPi: What has changed is not just the business boundary

In the past, the market used to regard XtalPi as an AI pharmaceutical company.

This classification is fully justified: the company's technical starting point, customer resources and revenue sources largely come from drug discovery.

However, after the release of XtalPi Science, it is difficult to explain what XtalPi is doing now by simply summarizing it as an "AI pharmaceutical company".

The coverage of the platform is no longer limited to drug molecules, but also includes proteins, chemical reactions, material formulas, solar cells and industrial experiments. The more important change is not the increase in the number of industries, but that XtalPi has begun to reorganize its capabilities.

In the past, customers usually purchased an R&D project, a set of robotic equipment or the research results of a certain stage; now, XtalPi tries to put models, professional tools, Agents and automated experimental resources into XtalPi Science uniformly, and schedule and measure them through Science Token.

In other words, the company is changing its "delivery unit": from delivering a single project to gradually continuously supplying scientific research capabilities.

The on-site demonstration of Genius Agents shows that it can continuously integrate experimental data, patents, SAR relationships and project context to provide suggestions for the next round of experiments; for materials, it can carry out intelligent search around high-dimensional variables such as composition, formula, structure and process, so that AI can form a self-closed loop of "reading, thinking and doing", and continuously adjust candidate schemes according to experimental results.

Although these research objects are different, the underlying processes are highly consistent: understand the scientific research goals, invoke data and models, design and execute experiments, and then enter the next round of optimization according to the results. What XtalPi Science replicates is not a single application, but a set of scientific research closed loops that can be migrated across industries.

In addition to the above applications, at this press conference, ecological partners also demonstrated applications in directions such as virtual cells, organoids and perovskite solar cells.

The scientific problems they face are different, but the underlying processes are highly similar: understand the scientific research goals, call up data and models, generate experimental schemes, execute real experiments, and then enter the next round of optimization according to the results.

These cases jointly illustrate one thing: What XtalPi Science replicates is not a single product, but a set of R&D processes that can be migrated across industries. Every time a new scenario is entered, the platform will add new models, workflows and experimental data; these capabilities can be reused by subsequent customers.

For investors, this may bring three different sources of value.

The first is the value of existing business, including drug discovery, robotic laboratories and professional R&D solutions. This part of the business provides revenue, and also brings real projects, industrial customers and high-quality data to the platform.

The second is the value of platform invocation. With the gradual implementation of Science Token, automated experimental services and private deployment, the same set of models, tools and experimental facilities can be reused by different customers, which is expected to bring stronger revenue continuity and improve the utilization rate of R&D assets.

The third is the option value of R&D achievements. For drug pipelines, new materials and new energy projects, XtalPi may also share long-term achievements through milestone payments, intellectual property licensing, joint venture equity and commercial revenue sharing.

Existing business provides certainty, platform invocation contributes compound interest, and R&D achievements bring non-linear elasticity. These three layers of value constitute a complete framework for re-understanding XtalPi.

/ 03 / Science Token, bringing scientific research into the era of Token economy

The new product form also means changes in the business model.

Science Token is a mechanism for XtalPi Science to uniformly schedule and measure models, data, professional tools, Agents and automated experimental resources.

A complex scientific research task may need to invoke large models, special models for molecules or proteins, scientific databases, simulation tools, automated experimental equipment and testing instruments at the same time. In the past, these resources belonged to different suppliers, and adopted different procurement, management and charging methods.

Users need to purchase software separately, coordinate laboratories, apply for equipment, and then researchers manually connect the entire process. What Science Token wants to do is to abstract these complex resources into a unified capability.

What users purchase is no longer a certain tool, but a scientific research task that can deliver complete results.

The first change brought by Science Token is the pricing method of scientific research capabilities.

In the past, it was difficult to form a standardized price for a molecular design, a round of experimental verification or a set of data analysis. After Science Token incorporates the invocation of different resources into a unified system, scientific research services can be available on demand, billed by usage, and even further charged by task or result, just like cloud computing.

Second, it establishes an ecological scheduling mechanism for AI for Science. External scientific research institutions, laboratories, model companies and tool manufacturers can connect their algorithms, data, equipment and experimental capabilities to XtalPi Science, which will be organized and scheduled uniformly by the platform.

At this point, XtalPi will gradually build an open scientific research service network.

/ 04 / What evidence is needed for the valuation system shift

The value realization of AI4S needs to complete two verifications: the first is to prove that AI can connect the whole process of scientific research, and the second is to prove that this capability can be standardized and scaled for repeated use by more customers.

XtalPi has completed the previous stage of verification in internal R&D and real projects, and now it is starting the second stage through XtalPi Science.

As of noon on July 29, nearly 200 enterprises, universities and scientific research institutions have submitted applications for platform trial use. XtalPi also announced that it will provide a total of 100 million Science Token platform trial quotas to alliance members and a group of in-depth cooperating universities and scientific research institutes.

Once scientific research capabilities can be invoked repeatedly, XtalPi's growth will no longer only depend on how many new projects are added every year, but also on how many customers the same set of models, workflows and experimental facilities can serve and how many tasks they can complete. As a result, the company is expected to gradually obtain the platform attributes of AI for Science infrastructure from a project-based R&D service provider.

This shift of valuation system will not be completed by a single press conference, but requires a series of quantifiable commercial indicators for continuous confirmation. However, in China's public market, enterprises that simultaneously have vertical scientific models, Agent orchestration, large-scale robotic laboratories, real project data and cross-industry customers are still scarce.

XtalPi's forward-looking vision is not only that it saw AI4S earlier, but that before the industry reached a consensus, it had completed the most time-consuming, capital-intensive and most difficult infrastructure construction for latecomers to quickly catch up.

In the past, AI was good at generating answers, and AI in the next stage needs to deliver evidence. What is truly scarce in AI for Science is not to put forward more hypotheses, but to continuously transform hypotheses into results verified by real experiments.

Whether XtalPi Science can promote the shift of the valuation system will ultimately be proved by paying customers, Science Token invocation volume, experimental facility utilization rate and platform revenue. But from the current point of view, short-term profit expectations are being reset, the endogenous business of AI4S is still maintaining rapid growth, and the platform-based business model has just entered the external verification stage.

Existing business provides the value base, platform invocation opens up growth space, real data constitutes a long-term moat, and drug and material R&D results bring upward elasticity — this may be the complete framework for re-understanding XtalPi.

This article is from the WeChat official account "Silicon-based Observation Pro", written by A Qi, published with authorization from 36Kr.