Why is AI4S replicating coding from three years ago?
There has never been a shortage of “trillion space” stories in the secondary market. What is really scarce is a definitive path that can be transferred from technological progress to orders, revenue, and cash flow.
If we set the clock back to 2022 to 2023, the commercialization of GitHub Copilot proved a critical point: the place where generative AI first achieved large-scale commercial implementation and made customers willing to pay was often not open, borderless general chat tasks, but professional scenarios where workflows have been digitized, results can be verified, and unit labor costs are extremely high. The reason why software engineering can become the first stop for AI commercialization is precisely because its input and output are highly structured, and the compiler provides perfect “instant verification”.
Today, capital markets are looking for the next plain of generative AI. When big models face ROI (input-output ratio) questions in generalized copywriting, R&D pools that can easily reach 100 billion US dollars in fields such as life sciences, new materials, and energy have become new soil for carrying the ambitions of giants. From digital productivity to scientific productivity, AI for Science (AI4S, scientific intelligence) is rapidly replicating the commercial trajectory of coding three years ago.
Strategic inflection point: upgrading from a scientific research topic to “national and giant expenses”
This round of AI4S heating up is not a pure industrial narrative, but is based on a high degree of resonance between the top strategies of China and the US and the capital of giants.
A shift in policy signals often precedes budget migration. The US government recently launched the “Genesis Mission” in a high-profile manner and released the official report “Science: A New Golden Age”, which directly proposed integrating the National Laboratory's supercomputing, scientific research data, and automation facilities to double the productivity of scientific research within ten years. Meanwhile, in the domestic WAIC (World Artificial Intelligence Conference) and the “Artificial Intelligence+Technological Innovation” policy introduced in many places, the establishment of high-quality scientific data sets and independent laboratories has gone from discussions in academia to actual construction of a new type of infrastructure. When the government and national laboratories began to jointly lead this process, AI4S ceased to be a cutting-edge exploration in papers, but began to form huge capital expenses that can be carried out by the industrial chain.
The actions of top tech giants are the most straightforward weather vane. In June of this year, Anthropic launched Claude Science for researchers. Instead of being a simple dialog box, it presets dozens of scientific research connectors in an attempt to directly occupy the work portal for researchers to access literature and process genomics and chemical calculations. At the same time, Nvidia released the BioNemo Agent Toolkit and collaborated with Thermo Fly to extend the smart device to laboratory hardware; DeepMind's IsoMorphic Labs frequently signed major orders with global pharmaceutical companies; and major domestic manufacturers such as ByteDance and Tencent are also racing on underlying facilities such as protein calculation and molecular dynamics. The giants are not only providing computing power, but are also competing for the “operating system” of the scientific computing era.
Commercial downsizing: Physical AI's first stop wasn't in the living room, but in the lab
In this grand story, Physical AI (physical intelligence) plays the role of “hands and feet” that push AI into the physical world. However, the market's fervor for physical intelligence often mismatches the commercialization timeline, and many people still expect domestic robots to quickly enter thousands of households.
This expectation overlooks the long-standing complexity of the real world. Consumer-grade robots face an open world without borders. The lighting, debris, and lifestyle habits of different households are very different. In a household scenario where fault tolerance is extremely low, even a 1% edge case (Edge Case) is enough to crash the entire business model.
In contrast, science labs are the ideal “sandbox” for physical AI. This is a highly structured microcosm in the physical world: hundreds of test tubes, microplates, and pipettes have uniform specifications, and instrument interfaces, operating procedures, and constant temperature and humidity environments can be accurately recorded. More importantly, scientific exploration itself is a high-value, highly standardized, highly controllable, and high-value process. If an experimental verification fails, the cause of the error can be accurately captured and fed back by the sensor to feed the model for the next round of learning iterations; once successful, its commercial value is immeasurable. Therefore, while artificial intelligence is still struggling in the quagmire of the living room and showing “too much milk” level operation performance from time to time, it can take the lead in breaking through the closed loop of “intelligent decision-making and physical execution” in a highly controllable laboratory. Whoever can let robots take over the tedious tasks in the lab can get the first large commercial check in the field of Physical AI.
Core barriers: scientific research “compilers” and real data wheels
The compiler for software engineering is code, and the compiler for scientific research is the law of physics. AI4S competition is often loosely summarized as “who has more research data,” but this is an extremely poor perception.
Published academic papers and patents are only the first layer of static data. Not only are they accessible to everyone, but they also have serious “positive bias.” Numerous negative test results are permanently locked in laboratory drawers, but for training a large model that understands the laws of physics and chemistry, “why it failed” often has a higher information entropy than “happened to succeed.”
On this track, the real moat is to have a closed loop of verifiable dynamic data. It requires AI agents to propose molecular hypotheses, automatically complete drug dispensing and synthesis through laboratory robotic arms, and finally the testing instrument sends back real physical data, whether positive or negative, in real time for the next round of reinforcement learning for the model. Being able to establish this kind of closed-loop enterprise not only crosses the industrial divide where AI experts don't understand “wet experiments” and scientists don't understand model architectures, but it can also continuously generate exclusive high-precision data in the daily workflow that competitors cannot buy with money. There is no feedback based on real physical verification, and even a huge static data set may just repeatedly amplify existing cognitive biases.
Investment logic: Beta comes from budget migration, Alpha comes from closed-loop ownership
The capital market needs to clearly recognize that AI4S will not re-evaluate the entire industrial chain at the same speed. The closer to cash flow, the stronger the beta attribute; the link that relies more on a single scientific breakthrough, the stronger the alpha attribute.
The first phase of beta implementation is likely to appear in computing power, scientific research software, experimental equipment, and data management platforms. These “operating system” providers don't need to wait for a new drug or material to go through lengthy regulatory approval and clinical verification. As soon as research institutions and companies' R&D budgets begin to migrate to digitalization and automation, they can obtain real large-scale orders and deferred revenue.
However, the excess revenue from the second and third stages belongs to AI-driven enterprises that have mastered the vertical scientific model, brought AI native assets to the market, and gained a share from innovation. This type of asset contains extremely high valuation flexibility, but investors must distinguish the difference between “total potential contract amount” and “current revenue” and understand the risks behind milestone payments. The combination of large-scale operating systems and explosive expectations of single-point innovation is expected to unlock a new blue ocean market and become another important aspect of AI after overcoming programming
Three years ago, the capital market was looking for a programming entrance in the AI era; today, the commercialization of AI4S does not need to wait for a scientific miracle to arrive; its explosion is based on the intersection of the collapse of computing power costs, maturity of automation equipment, and the immediate needs of the industry. The mapping of A/H shares on this new main line has been extended from scientific research computing power to the world of experiments and physics: Zhongke Shuguang (603019.SH), Inspur Information (000977.SZ), and Industrial Wealth Connect (601138.SH) correspond to computing power and servers, and Jingtai Holdings (02228) is positioned as an AI4S platform for algorithm+robot, Preferred Choice (09880), Central Control Technology (688777.SH), Huichuan Technology (300124.SZ), and Yuejiang Technology (02432), covering robot bodies, intelligent manufacturing, and industrial automation. There are many designers in related fields. In the end, who can take the lead in turning technical exposure into repeatable orders and truly grasp the “model-experiment-data” closed-loop pricing power?