Artificial intelligence is fundamentally transforming the methods of scientific discovery, and with it, the pathways through which cities compete in technological advancement. Historically, scientific strength was derived from the concentration of talent, institutions, and capital, but as research activities become increasingly platform-based and networked, the crucial question for cities now is how to convert existing technological advantages into broader, more universally accessible public capabilities.
This analysis argues that Shenzhen should not follow the traditional trajectory of established scientific cities. Instead, it can explore the construction of interdisciplinary, reusable, and globally accessible scientific intelligence infrastructure, thereby forging a new competitive edge in the era of AI-driven science. The city's genuine opportunity in AI for Science (AI4S) does not lie in building dedicated AI platforms for biology, materials, or energy, but rather in organizing computational power, scientific data, robotics, automated laboratories, major research facilities, and standards systems into a unified, cross-disciplinary urban foundation that global research institutions can access through standard interfaces.
AI4S Competition is Shifting from "Who Has the Strongest Model" to "Who Has the Most Complete Research System"
Over the past few years, the focus of AI competition has been primarily on models: parameter scale, training compute, and benchmark scores. However, the next phase of AI for Science will be determined not by a single model, but by a comprehensive closed-loop of scientific discovery that encompasses hypothesis generation, computational invocation, experiment design, equipment operation, data acquisition, and feedback learning.
The U.S. Department of Energy is advancing the Genesis Mission, aiming to connect supercomputers, experimental facilities, AI systems, and unique data across its 17 national laboratories into an integrated scientific platform, with a goal of doubling U.S. research productivity and impact within a decade. Europe's RAISE (Resource for AI Science in Europe) is pursuing a similar direction: rather than building isolated platforms around specific disciplines, it aggregates compute, data, talent, and research funding, using interdisciplinary networks to enable AI capabilities developed in one field to be reused in others.
These developments signal a significant shift: AI4S is moving from being a "disciplinary tool" to becoming "research infrastructure." If this assessment holds, the real prize for cities is no longer leading in a particular scientific model, but rather becoming the infrastructure provider for next-generation scientific discovery.
The Last Thing Shenzhen Should Do is Replicate Another Discipline-Specific Platform Center
Shenzhen should certainly develop AI4S applications in biomedicine, advanced materials, energy, and quantum technologies, but a city-level strategy should not equate to building a separate platform for each discipline. Many of the foundational capabilities of AI4S are inherently reusable: scientific literature and knowledge bases, compute scheduling, scientific data governance, multimodal models, simulation environments, agent orchestration, laboratory instrument interfaces, robot control, experimental data feedback, and traceability and security audits are not exclusive to any single discipline.
What truly requires deep domain specialization is the top layer: professional knowledge, domain-specific models, experimental workflows, and evaluation systems. Therefore, a more sensible architecture for Shenzhen would be "shared foundation at the bottom, specialized at the top": the city builds a cross-disciplinary scientific intelligence base, upon which different disciplines develop their own adaptation layers. Experimental scheduling capabilities developed in materials science can be reused by chemistry; data governance tools from life sciences can be transferred to drug discovery; and the perception-planning-execution loop from embodied intelligence can enter automated laboratories.
In other words, Shenzhen's challenge is not "how each discipline can have its own AI4S," but "how different disciplines can share the same set of scientific intelligence production assets." Only this approach can generate city-level economies of scale.
Why This Path Suits Shenzhen: From Tech to Systems
Shenzhen's greatest strength is not its concentration of individual research resources, but its ability to turn technology into complete systems. First, the city has a substantial innovation investment base. In 2024, Shenzhen's total R&D investment reached 245.31 billion yuan, with an R&D intensity of 6.67%, ranking first among Chinese cities. Notably, enterprise R&D investment accounted for over 93% of the total, reflecting an innovation structure where the distance between research and engineering, technology, and industry is exceptionally short. AI4S precisely requires this capability to rapidly transform algorithms into experimental systems and research tools.
Second, Shenzhen possesses a physical execution layer that is rare and highly valuable in the AI4S era. In 2025, the added value of Shenzhen's strategic emerging industries reached 1.67 trillion yuan, accounting for 43% of its GDP. The city produced 133,900 industrial robots and 7.1129 million service robots during the year. From chips, sensors, machine vision, and motion control to robotic arms, drones, mobile robots, and industrial software, Shenzhen has established a complete industrial chain that applies digital intelligence to the physical world.
The importance of this is often underestimated. AI can predict a new material, design a molecule, or propose an experimental plan, but scientific discovery ultimately requires verification in the physical world. The U.S. Genesis Mission also specifically incorporates robotics, edge AI, real-time analysis, and intelligent feedback into its autonomous laboratory systems. For Shenzhen, the significance of its robotics industry extends beyond "developing robots"; it represents an opportunity to become the experimental execution infrastructure for AI4S. Research in robot computing has shown that future robotic systems will require a complete computing ecosystem spanning perception, computation, control, and application software, rather than just a single algorithmic module.
Third, Shenzhen already operates major scientific facilities that can be reorganized by AI4S. The first batch of major scientific infrastructure at Guangming Science City, including synthetic biology, brain analysis and simulation, and materials genome facilities, is already operational, having served over 200 universities, research institutions, and enterprises, with more than 330,000 hours of effective equipment time. Facilities like Pengcheng Cloud Brain III and the free-electron laser are also under continuous development. What Shenzhen truly needs now is not another facility, but the ability for compute, models, robots, and scientific instruments to call each other through unified interfaces.
What Shenzhen Should Build: Not a Building, But a Scientific Operating System
If this path is made concrete, Shenzhen's city-level AI4S foundation should comprise at least five layers. The first is the computing foundation: organizing heterogeneous resources such as intelligent computing, supercomputing, and future quantum computing into a unified scheduling system oriented toward scientific tasks, allowing research teams to invoke resources by task rather than navigating the boundaries of different computing centers.
The second layer is the data and model foundation: establishing scientific data standards, knowledge bases, foundation models, simulation tools, agent orchestration, and traceability mechanisms so that different disciplines can share common AI capabilities.
The third layer is the experimental execution foundation. This should be Shenzhen's most distinctive layer: connecting robots, automated laboratories, scientific instruments, and major research facilities through standardized interfaces, enabling AI not only to propose answers but also to initiate experiments, obtain feedback, and proceed to the next round of reasoning.
The fourth layer is the disciplinary adaptation layer: biomedicine, advanced materials, energy, and quantum fields retain their own professional data, models, and experimental protocols, but no longer duplicate the construction of foundational shared capabilities.
The fifth layer is the standards and governance layer: data formats, instrument interfaces, robot operation protocols, model evaluation, security audits, intellectual property, and cross-border scientific data rules should all be designed in tandem with the platform. Only by achieving this layer can Shenzhen's foundation transcend being a local IT project and become a replicable combination of institutional and technological frameworks.
Beyond Serving Shenzhen: Becoming the Default Interface for Global Scientific Discovery Networks
If this foundation only serves local research institutions, it remains merely an excellent regional platform. Shenzhen's true ambition should be to leverage its global capabilities. In 2025, Shenzhen's total foreign trade import and export value reached 4.55 trillion yuan, continuing to rank first among mainland Chinese cities, with exports of 2.74 trillion yuan, maintaining the top position for 33 consecutive years. In the same year, the World Intellectual Property Organization ranked the "Shenzhen-Hong Kong-Guangzhou" cluster as the world's number one innovation cluster.
What AI4S can learn from these achievements is Shenzhen's long-established ability to turn technology into products, build supply chains around those products, and enter global markets through standards and networks. AI4S can follow the same logic, but the output must shift from products to scientific discovery capabilities. Achieving global coverage does not mean centralizing all the world's research data in Shenzhen. A more realistic system is a "Shenzhen core foundation + global research nodes" model: public models, toolchains, experimental protocols, and evaluation standards can be exported from Shenzhen; sensitive data remains local; overseas laboratories connect through unified interfaces; and experimental results are standardized for cross-institutional replication and validation.
Hong Kong can serve as the connector for global universities, capital, professional services, and international rules, while Shenzhen provides engineering implementation, robotics, computing power, hardware, and productization capabilities. This is where Shenzhen can build a long-term moat compared to many traditional scientific cities. Beijing can possess the strongest basic research institutions, and Shanghai can develop deep disciplinary advantages in several scientific and industrial fields, but Shenzhen has the option to choose a different track: becoming the city that connects AI, computing, and the experimental world, and engineering, standardizing, and internationalizing that connection.
From Manufacturing Global Products to Defining Global Experimental Interfaces
Shenzhen's most successful historical experience has never been inventing all the world's technologies, but rather establishing a highly efficient engineering system that allows global innovation to be rapidly transformed into products in Shenzhen. AI4S presents an opportunity to extend this capability upstream into scientific discovery. Therefore, what Shenzhen should truly compete for is not which city hosts the most AI4S projects, or even which city possesses the most powerful scientific foundation model.
The more critical question is: when future global scientists design experiments, invoke AI, connect robots, and verify results, will they use the interfaces, tools, and standards defined by Shenzhen? If the answer is yes, Shenzhen will be exporting not just chips, robots, and intelligent hardware, but a higher-value public capability: the capacity for scientific discovery. This is the most valuable position Shenzhen can aspire to in AI4S development: not building yet another platform, but becoming one of the default interfaces for global scientific intelligence infrastructure.