The secondary market is never short of "trillion-dollar market" stories, but what is truly scarce is a clear path from technological progress to orders, revenue, and cash flow.
If we rewind to 2022 and 2023, the commercialization of GitHub Copilot proved one crucial thing: the places where generative AI achieves scaled commercial adoption and gets customers to pay willingly are not open-ended, general-purpose chat tasks. Instead, they are professional scenarios where workflows are already digitized, results can be verified, and unit labor costs are extremely high.
Software engineering became AI's first commercial stop because its inputs and outputs are highly structured, and the compiler provides a perfect "instant verification" mechanism. Today, capital markets are searching for the next frontier for generative AI. As large language models face ROI scrutiny in generic copywriting, the multi-hundred-billion-dollar R&D pools in life sciences, new materials, and energy have become the new fertile ground for corporate ambition. Moving from digital productivity to scientific productivity, AI for Science (AI4S) is rapidly replicating the commercial trajectory of Coding from three years ago.
Strategic Inflection Point: From Academic Topic to "National and Corporate Spending"
The current surge in AI4S is not purely an industrial narrative; it is built on the high resonance between the top-level strategies of China and the US and the capital spending of major corporations. A shift in policy signals is often a precursor to a shift in budgets. The US government recently launched the "Genesis Mission" and published the official report "Science: A New Golden Age," which directly proposes integrating national laboratory supercomputing, research data, and automated facilities to double scientific research productivity within a decade.
In China, at the WAIC (World Artificial Intelligence Conference) and through various "AI + Tech Innovation" policies, building high-quality scientific datasets and autonomous laboratories has moved from academic discussion to tangible new infrastructure construction. As governments and national laboratories begin to lead this process, AI4S is no longer a frontier exploration confined to academic papers but is forming a massive capital expenditure that can be absorbed by the industrial chain.
The actions of leading tech giants are the most straightforward indicators. In June this year, Anthropic launched Claude Science, aimed at researchers. It is not a simple chat interface but is pre-loaded with dozens of scientific connectors, attempting to directly occupy the workflow of researchers using literature, processing genomics, and performing chemical calculations.
Meanwhile, NVIDIA released the BioNeMo Agent Toolkit and partnered with Thermo Fisher Scientific to extend agents to lab hardware. DeepMind's Isomorphic Labs is frequently signing major contracts with global pharmaceutical companies. In China, giants like ByteDance and Tencent are also aggressively building foundational infrastructure for protein computing and molecular dynamics. These giants are not just providing computing power; they are competing to become the "operating system" for the era of scientific computing.
Commercial Downshifting: Physical AI's First Stop is Not the Living Room, But the Lab
Within this grand narrative, Physical AI (Embodied AI) plays the role of the "hands and feet" pushing AI into the physical world. However, the market's enthusiasm for Embodied AI often misaligns with the commercialization timeline. Many still expect domestic service robots to quickly enter millions of households, ignoring the long-tail complexity of the real world.
Consumer-grade robots face a completely unbounded open world, where lighting, clutter, and living habits vary wildly between homes. In the low-tolerance environment of a home, even a 1% occurrence of edge cases can cause the entire business model to collapse. In contrast, the scientific laboratory is the ideal "dimensionality-reduction sandbox" for Physical AI.
It is a highly structured microcosm within the physical world: hundreds of types of test tubes, microplates, and pipettes have unified specifications, and instrument interfaces, operating steps, and constant temperature and humidity environments can be precisely recorded. More importantly, scientific exploration itself is a high-value, highly standardized, and highly controllable process. When an experimental validation fails, the cause of the error can be precisely captured by sensors and fed back, becoming fodder for the model's next learning iteration. Conversely, if it succeeds, its commercial value is immeasurable.
Therefore, while Embodied AI is still struggling in the quagmire of the living room, occasionally displaying clumsy operation performance, it can first successfully close the "intelligent decision-making and physical execution" loop within a highly controllable laboratory. Whoever can get robots to take over the tedious, fine-grained operations in the lab will secure the first large commercial check in the field of Physical AI.
Core Moat: The "Compiler" of Science and the Real Data Flywheel
The compiler for software engineering is code, while the compiler for scientific research is the laws of physics. Competition in AI4S is often oversimplified as "who has more research data," but this is an extremely incomplete view. Publicly available academic papers and patents are only the first layer of static data. They are not only accessible to everyone but also suffer from severe "publication bias."
Countless failed negative experimental results are permanently locked away in laboratory drawers. Yet, for training a large model that understands the laws of physics and chemistry, "why it failed" often has higher information entropy than "coincidentally succeeding." In this domain, the true moat is possessing a verifiable dynamic data closed-loop. This requires an AI agent to propose a molecular hypothesis, a robotic arm in the lab to automatically complete the dispensing and synthesis, and finally, detection instruments to feed back the real physical data—whether positive or negative—in real-time for the model's next round of reinforcement learning.
Companies that can establish this closed-loop not only bridge the industry gap where AI experts don't understand "wet lab" work and scientists don't understand model architecture, but they can also continuously generate proprietary, high-precision data within their daily workflows that competitors cannot buy. Without feedback based on real-world physical verification, no matter how massive the static dataset, it may only repeatedly amplify existing cognitive biases.
Investment Logic: Beta Comes from Budget Migration, Alpha Comes from Closed-Loop Ownership
Capital markets need to clearly recognize that AI4S will not revalue the entire industrial chain at the same speed. The closer a link is to cash flow, the stronger its Beta characteristics. The greater the dependence on a single scientific breakthrough, the stronger its Alpha characteristics.
The first phase of Beta payoff is likely to appear in computing power, scientific software, laboratory instrumentation, and data management platforms. These "operating system" providers do not need to wait for a new drug or material to pass through lengthy regulatory approval and clinical validation. As long as the R&D budgets of research institutions and companies start shifting towards digitalization and automation, these providers will receive tangible, scaled orders and deferred revenue.
The second and third phases of excess returns belong to those AI-driven enterprises that have mastered vertical scientific models, brought AI-native assets to market, and can share in the gains from innovation. These types of assets carry extremely high valuation elasticity, but investors must distinguish between "potential total contract value" and "current revenue," and recognize the risks behind milestone payments.
The combination of a scalable operating system and the explosive expectations of point innovation has the potential to unlock a new blue ocean market. This could become AI's next major domain after conquering programming.
Three years ago, capital markets were looking for the programming gateway in the AI era. Today, the commercialization of AI4S does not need to wait for a scientific miracle to occur. Its explosion is built on the convergence of collapsing computing costs, mature automation equipment, and industrial rigid demand. The mapping of this new main line in A/H shares has already extended from research computing power to the experimental and physical world.
This includes Dawning Information Industry Co.,Ltd. (603019.SH), Ieit Systems Co.,Ltd. (000977.SZ), and Foxconn Industrial Internet Co.,Ltd. (601138.SH) for computing power and servers; XTALPI (02228) positioned as an algorithm+robotics AI4S platform; and UBTECH ROBOTICS (09880), Supcon Technology Co.,Ltd. (688777.SH), Shenzhen Inovance Technology Co.,Ltd. (300124.SZ), and other companies covering robot bodies, smart manufacturing, and industrial automation.
Many players are laying out their positions in related fields. Ultimately, who will be the first to convert technological exposure into repeatable orders and truly control the pricing power of the "model—experiment—data" closed-loop?
Comments