The emergence of large language models has redefined productivity in the digital realm. Physical AI is now poised to redefine productivity in the physical world. QUANTGROUP (02685.HK) does not sell robot hardware or single-point solutions; it sells the general capability layer that enables machines to comprehend the physical world—the Physical World Foundation Model. This model has already been validated in global capital markets by companies like Physical Intelligence (valued at $2.4 billion) and Skild AI (valued at $14 billion). As the first listed "Intelligent Species" entity on the Hong Kong Stock Exchange, Quantgroup is building a similar technological moat within the Chinese market.
Embodied Intelligence Industry Evolution
The embodied intelligence sector is undergoing a strategic differentiation. Some companies sell robot hardware, others offer complete solutions, while Quantgroup focuses on the Physical World Foundation Model. From a standard valuation perspective, hardware companies are valued based on sales volume multiplied by gross margin, solution providers on project count multiplied by per-unit price, and physical AI model companies on API call volume multiplied by usage duration. Comparatively, the latter holds greater potential for growth, as its ceiling is not limited by hardware production capacity or project delivery capability but by the model's own generalization ability.
The Core Arena: Physical AI
The transition of AI from the digital to the physical world represents a leap from "knowing" to "doing." Large language models grant AI the ability to understand language and reason logically, but Physical AI requires comprehension of gravity, friction, material deformation, and spatial relationships—fundamental laws of the physical world not covered by text-based training. The entity that masters the ability for robots to rapidly adapt to new scenarios holds the key to scalability.
Quantgroup's core strategy revolves around a cross-scenario, cross-hardware open base model, representing a pivotal path forward in this field. It is not tied to specific robot hardware brands nor confined to a single vertical scenario. Instead, it offers general physical capabilities to the entire industry, achieving scale by empowering partners across the value chain. This mirrors the core logic of Anthropic in the digital AI domain—becoming the foundational infrastructure for the entire industry through a general capability layer, with industry enablement as the primary means for the base model's deployment.
This open base model inherently includes a world model. The currently debated World Action Model (WAM) is a core functional module within the base model system. The base model uses the world model for reasoning and predicting physical laws, and relies on WAM for action generation and execution, together enabling a robot's autonomous operation capabilities in the real world.
Three Key Catalysts
A transformation is underway, pushing robots from exhibition halls into real-world applications. In June 2026, a joint notice from the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission launched a special action for humanoid robots and embodied intelligence in practical scenarios across 10 provinces and municipalities, covering industrial, service, and special application fields.
The policy mandates "minimal intervention and reuse of existing resources," forcing robots to prove their utility under current conditions. This requirement pushes the industry decisively into real scenarios. Each region must select no fewer than 20 key scenarios, covering at least two of the three major fields, with central state-owned enterprises selecting at least 10. A summary of results is due by November 30th. This means that in the coming months, a cohort of companies will use operational data from real scenarios to gain policy recognition and secure follow-on orders.
Technologically, Physical World Foundation Models are transitioning from concept to usability. Leading players' models can already perform unstructured operations, adapt to open environments, and execute long-chain autonomous tasks in real settings. On the capital front, top-tier global investment institutions are placing significant bets on this track, signaling their assessment of its value.
Business Model Differentiation
The Physical AI track features three primary business models: selling hardware, selling solutions, and selling capabilities. Selling hardware involves a one-time delivery; the customer purchases the robot and is responsible for subsequent maintenance and new-scenario adaptation, making it difficult to reduce marginal costs at scale. Selling solutions provides bundled services charged per project, allowing for depth but limiting breadth.
The third model is selling capabilities, operating on the logic of "train once, reuse across multiple scenarios." The Physical World Foundation Model provides the general ability for robots to understand the physical world and make real-time decisions. Once this capability is proven, it can be called like an API by different hardware platforms, with marginal costs approaching zero for each new scenario added. It solves the problem of rapid cross-scenario transferability, enabling robots to be operational out-of-the-box in unfamiliar environments. This aligns with the policy goal of "validate one, deploy a batch, drive widespread adoption," offering greater scalability. Concurrently, the Robot-as-a-Service (RaaS) model fits this logic: payment is based on utility, selling capability rather than hardware, where customers pay for the outcome of the robot's work, not the robot itself.
Global Capital Validation
The commercial value of Physical World Foundation Models has been validated by the world's most prestigious investment institutions.
Physical Intelligence saw its valuation soar sixfold from $400 million to $2.4 billion within eight months, following a $400 million funding round in November 2024 backed by investors including Jeff Bezos, OpenAI, Sequoia Capital, and Khosla Ventures. The company does not manufacture robot hardware, focusing solely on the general AI model that enables robots to understand the physical world.
Skild AI's trajectory is even steeper. Founded by former core researchers from Meta's robotics lab, this company also avoids hardware, concentrating on a general-purpose brain for robots. Its Series A valuation in July 2024 was $1.5 billion. In less than a year, its Series B reached $4.7 billion, followed by a Series C led by SoftBank and Nvidia at a $14 billion valuation—all while the company's annual revenue was just $30 million. This 3000x revenue-to-valuation multiple demonstrates the market's belief in the innovation potential and vast application prospects of a general robot brain, signaling that the valuation ceiling for Physical World Foundation Models far exceeds that of hardware or solution companies.
Quantgroup's positioning in the Physical AI field is analogous to Anthropic's role in large language models—acting as a Model-as-a-Service platform for the physical world, enabling different hardware platforms to call upon the same AI capability layer. Physical Intelligence and Skild AI have demonstrated the potential height of this path's valuation ceiling. Quantgroup's mission is to establish a similar technological moat within the Chinese market.
Quantgroup's Strategic Positioning
Quantgroup's commercial positioning is clear: a provider of an open, cross-scenario, cross-hardware world model for living scenarios. It does not bind itself to specific hardware or lock into particular scenarios; it develops a general AI capability layer that can be called upon by robots from various manufacturers.
Core Technological Leap
Quantgroup's technical roadmap aims to advance robots from "motion automation" to "task-level autonomous operation." The former involves executing actions based on fixed programs, while the latter entails understanding a task objective and autonomously completing the full cycle of perception, decision-making, and execution. Quantgroup's approach involves a layered software-hardware architecture where the Physical World Foundation Model is not tied to any specific hardware; the same base model capabilities can run on different hardware platforms. This is the prerequisite for scaling the RaaS model: the Physical World Foundation Model, as a technical asset, can be continuously called upon, generating recurring revenue.
First Mover Advantage
Quantgroup possesses long-term technical accumulation. Its prospectus indicates mature technological reserves in areas like automated machine learning and NLP. Its Physical AI business represents an extension and upgrade of its digital decision-making capabilities, not a development from scratch. During the current policy-driven window for real-world industrial deployment, the entity that first accumulates data and validation records from real scenarios will gain a first-mover advantage.
The RaaS business model has already received policy endorsement. As a listed entity on the Hong Kong Stock Exchange, Quantgroup benefits from the listed company platform, gaining an additional layer of credibility in financing capability, compliance transparency, and commitment to long-term investment—advantages not typically held by private startups.
Real-World Validation in Kitchen Scenarios
Quantgroup has conducted four rounds of technical validation, all deployed in the dynamic, real-world conditions of commercial kitchens, not laboratory environments. At a time when much of the industry is still showcasing demo videos as results, this pace places Quantgroup ahead.
Validation 1: Sandwich Assembly – testing real-time compliant control with unstructured, soft ingredients like bread, lettuce, and sauces.
Validation 2: Shopping Bag Sorting – testing scriptless autonomous decision-making for identifying, grasping, and categorizing unknown items in a bag.
Validation 3: Steak Seasoning – testing a complete reasoning chain where the robot must search multiple drawers to locate salt and perform precise seasoning without predefined steps.
Validation 4: Milk Tea Preparation – testing system-level coordination where the robot must seamlessly interact with other machines like drink dispensers, stirrers, and sealers, managing liquid dynamics and positional accuracy.
Following these four validations, the next step is cross-scenario reuse. The capabilities honed in the dynamic kitchen environment—unstructured operation, open-environment adaptation, and long-chain autonomous task execution—are theoretically the foundational abilities required in other variable-rich sectors like warehousing and logistics, inspection and analysis, and healthcare.
A deeper layer of potential lies in the concept of "Intelligent Species"—any physical terminal equipped with perception modules, AI decision-making capability, and the ability to autonomously perform physical interactions can integrate with the Physical World Foundation Model to become an intelligent carrier. This concept extends the reuse path from餐饮 into a much broader hardware ecosystem.
Competitive Barriers: Data Flywheel and Full-Stack R&D
The viability of Quantgroup's business logic hinges on two factors: whether the technical capability of its Physical World Foundation Model can support cross-scenario reuse, and whether the company can establish barriers in data accumulation and system integration.
Data is the one resource that cannot be simply purchased. In Physical AI, high-quality, real-world operational data is paramount and can only be accumulated over time through actual deployment. For Quantgroup, the data accumulation barrier stems from a multi-path collection system: B2B commercial scenario deployment through partnerships with robot hardware companies in sectors like餐饮; C2C smart hardware布局 through diverse hardware collecting behavioral and environmental data during daily use; supplemented by user置换式采集 and scenario co-creation data sharing. These four parallel paths transform data collection from a single point into a systematic approach.
Competition in this field is, at its core, a race against time. Successfully establishing a virtuous cycle of data accumulation is, to a significant extent, more critical than any single-point technological breakthrough.
Globally, a definitive leader has not yet emerged. Established players have validated scenarios that are relatively standardized, focusing on warehousing, logistics, and household desktop tasks. Quantgroup has chosen the commercial kitchen—an environment with a far higher density of variables, offering theoretically greater generalization value but also presenting higher validation difficulty. Domestically, other startups are targeting industrial and logistics scenarios, forming two differentiated paths alongside Quantgroup's餐饮 focus. It remains to be seen which path will more effectively lead to a truly general Physical World Foundation Model.
Growth Path: From RaaS to Platform Revenue
The commercial potential of Physical World Foundation Models must overcome two hurdles: whether the marginal cost of cross-scenario reuse can truly approach zero, and whether the gap between theoretical potential and practical performance can be narrowed to a commercially acceptable range.
Quantgroup's long-term path can be divided into three stages. Stage 1: Scenario-specific solution deployment and data monetization (present). Using high-frequency民生 scenarios like commercial kitchens as entry points to complete technical validation and商业闭环, accumulating real-world operational data. The RaaS model has official backing, and Quantgroup, as a listed entity, is positioned to capitalize quickly during this policy window.
Stage 2: Model calls and value-added services (medium-term). Providing model calls and增值服务 to smart hardware manufacturers. The Physical World Foundation Model serves as a general capability layer called upon by robots from different manufacturers across different scenarios, with收费 based on call volume and usage duration. The revenue model shifts from project-based to platform-based.
Stage 3: Base model plus compute (long-term). Establishing a revenue structure combining the base model and compute power, aligning with capital market valuation logic. The Physical World Foundation Model becomes the "operating system" for all robots, while the compute layer provides the infrastructure for training and inference.
Market signals are clear. The national special action is highly specific: involving 10 provinces/municipalities and relevant central state-owned enterprises, covering industrial, service, and special fields. The combination of policy endorsement for the RaaS model and this impending deadline transforms the notion that "selling capability" is more aligned with industrial direction than "selling hardware" from a mere judgment into an unfolding inflection point.
The Inflection Point is Here
The true industry scaling inflection point will arrive when third-party external customers independently purchase Quantgroup's Physical World Foundation Model and deploy it in entirely new industry scenarios for commercial delivery. Prior to this,赛道 valuation relies more on industrial expectations. Once third-party commercialization is successfully demonstrated, the underlying Physical AI model will transition from a conceptual赛道 into a core industrial platform generating sustainable revenue.
The "ChatGPT moment" for Physical AI is imminent, driven by policy acceleration for real-world deployment, maturing underlying technology, and continued global capital investment. QUANTGROUP (02685.HK), as the first listed "Intelligent Species" entity on the Hong Kong Stock Exchange, is already positioned on the Physical AI赛道, showcasing its distinctive business model.
Comments