Produced by | Huxiu Tech Group
Authors | Chen Yifan, Liu Xuanqi
Editor | Miao Zhengqing
Header Image | AI Generated
This is the 21st installment of Huxiu's WAIC "Tracking Token Business New Paradigms" series.
In tech investing, age is never a sufficient condition for judging competence. But during periods of rapid technological paradigm shifts, young investors warrant separate observation.
The reason is not complicated. AI, robotics, commercial space, and quantum computing are advancing simultaneously. The speed at which new technologies move from papers to products, and from labs to industry, has noticeably accelerated. Traditional investment logic from the past no longer applies. To enter these projects, investors can no longer rely solely on financial models and experience in mature industries. They must also understand technical language, accept unconverged development paths, and make judgments before consensus forms.
In the mobile internet era, a company could quickly prove itself with user growth, retention, revenue, and network effects. Today's AI and hard-tech projects, however, may span algorithms, chips, sensors, supply chains, industrial scenarios, and regulatory systems. Investors must both understand a single technological leap and judge whether it can become a stable product, form a data and training feedback loop, and accompany the company through long cycles of engineering and commercialization.
At this year's WAIC, the "Alpha Youth Investment Leaders" list was released, naming 10 new-generation investors. But beyond the list, more importantly, when these investment directions and judgments are viewed together, they resemble a tech investment trend map in the process of being drawn.
Details of the WAIC FUTURE TECH Young Investor List
This map reveals a common change: capital is shifting from chasing technical buzzwords towards identifying the real pathways through which technology enters industry. The next wave of opportunities may not belong to the loudest concepts, but rather appear in segments where model dividends have yet to be released, in capability gaps still unfilled in the physical world, and in infrastructure requiring long-term capital to enter early.
Trend One: AI is Leaving the Screen, Entering the Physical World
Over the past few years, the market's primary entry point for understanding AI has been large models, chatbots, and generative applications. Now, investment attention is moving towards edge devices, robots, vehicles, industrial systems, and space infrastructure. AI is no longer just about generating text or an image; it must sense the environment, understand human intent, and complete actions in the real world.
This is also why embodied intelligence, humanoid robots, AI hardware, and edge intelligence are receiving concentrated focus. Zhou Xin, Investment Executive Director at Jinqiu Fund, began focusing on robotics in 2018.
She cautioned during the roundtable that the bubble in embodied intelligence doesn't stem from a lack of technological progress, but from the market applying the iteration speed of large language models directly to robots deeply coupled with the physical world. Improvements in model capability do not equate to synchronous completion of real-world deployment. An impressive demo can spark imagination but cannot answer whether a robot can perform a task one hundred times consecutively, whether the error rate is controllable, whether deployment and operational costs can decrease, and ultimately, whether the ROI makes sense.
This line of judgment pulls the competition in embodied intelligence back from "demonstration effects" to "delivery capability." There is no prompt to retry in the real world. Factories, warehouses, homes, and public spaces all require systems to handle noise, wear and tear, unexpected situations, and unpredictable human behavior. The model is only one layer; touch, sensing, actuators, control systems, data collection, and scenario feedback collectively determine whether a product can work.
Capital enthusiasm can be quantified.
According to the China Academy of Information and Communications Technology's "Embodied Intelligence Development Report (2025)", as of December 2025, there were 744 investment events in China's embodied intelligence and robotics field, with total funding reaching 73.543 billion RMB. IDC predicts global humanoid robot shipments will exceed 50,000 units in 2026, a year-on-year increase of 178%.
But cold data on deployment also exists: In the first three quarters of 2025, revenue from humanoid robots for leading company Unitree came overwhelmingly from the scientific research and education sector, accounting for 73.6%. A large number of devices were sold to universities and research institutions, indicating there is still a way to go before creating practical value in factories. This precisely confirms Zhou Xin's reminder: model progress cannot be directly equated with deployment progress.
Edge intelligence thus becomes another important thread. Zhou Xin believes that if AI is a cognitive revolution approaching the scale of an industrial revolution, intelligence cannot remain solely in the cloud for long. The future cognitive network will consist of both cloud and edge components: the bottom layer is chips and hardware carriers, the middle is the runtime environment formed around Agents, and the upper layer consists of applications that can understand individual needs. Agents may gradually become the interaction interface for hardware, understanding user preferences, history, and state upwards, while calling upon phones, glasses, cars, robots, and home devices downwards.
This will change the business model for AI hardware. Hardware will no longer rely solely on one-time sales for revenue but may create value through ongoing services. The real barrier is also not just "adding AI to a device," but whether the device can lower usage barriers and form stable data and service relationships around the user.
The investments of several people on the list have already unfolded along this path: Bai Zerén, Vice President at Houxue Capital, focuses on unmanned delivery, tactile sensing, and highway logistics autonomous driving. Guo Jing, Investment Manager at YaoTu Capital, focuses on vertical specialized robots, drone swarms, and photovoltaic installation robots. Unmanned delivery is one of the fastest-moving sectors in "AI entering the physical world": in 2025, total industry financing approached 10 billion RMB. White Rhino, invested in by Linear Capital where Bai Zerén is located, completed three rounds of financing totaling over $100 million in the same year. Its active operational vehicles grew from about 100 at the end of 2023 to over 2,000 by December 2025, operating routinely in over 170 cities globally. This type of business doesn't have stunning demos but undergoes daily tests of efficiency, cost, and reliability. After AI moves from the digital to the physical world, technological imagination remains important, but the ability to scale and deliver will become the harder dividing line.
Trend Two: The Large Model Dividend Isn't Over; Competition Shifts Focus to the "Intelligence Flywheel"
While the market constantly seeks the next-generation architecture and new concepts, an easily overlooked fact is: the industrial dividend of this generation of large models may be far from exhausted.
Hu Qi, Executive Director at Qiming Venture Partners, suggests that one of today's biggest false consensuses is that people pay too much attention to the next leap-forward technology and discuss too little about how to truly enable existing large models to reconstruct productivity. After TCP/IP became a standard protocol, the internet dividend was released over a long period; Bayesian theory and CNN also continued to influence new technical systems for years after their proposal. The journey from the emergence of foundational technology to its transformation of industries is typically not a linear process of one or two years. Large models have only been in the public eye for a few years; commercial systems, organizational processes, and industrial infrastructure have not yet fully caught up with model capabilities.
This means investors don't have to choose between "continuing to invest in large models" and "searching for the next-generation technology." A more important question is: which companies can embed existing model capabilities into real workflows, creating value that cannot be easily erased by the next model upgrade.
Liu Yunjie, Investment Director at Jingya Capital, refers to this capability as the "intelligence flywheel." Users generate data by using the product; this data enters the training or optimization process; improved model capabilities enhance the experience; a better experience attracts more users, scenarios, and feedback. The data feedback loop, training loop, and user feedback loop push each other, giving the company a continuously self-reinforcing capability.
Applying this framework to specific companies, Zhipu and Kling serve as two illustrations. The host mentioned during the roundtable that Hu Qi heavily invested in Zhipu before its valuation surged. Zhipu's revenue grew from 57.4 million RMB in 2022 to 312.4 million RMB in 2024, with a three-year compound annual growth rate of about 130%. Its revenue for the first half of 2025 was 190.9 million RMB, a year-on-year increase of 325%. It listed on the Hong Kong Stock Exchange in January 2026, becoming the "first listed large model company globally." Kuaishou's Kling AI demonstrates the commercialization speed possible when model capabilities are embedded into a creation workflow: it opened for testing in June 2024, launched a收费体系 in about 45 days, generated approximately 10.4 billion RMB in revenue for 2025, exceeded $200 million in monthly revenue in December 2025, and by the end of 2025 had over 60 million global users, cumulatively generating over 600 million videos, serving over 30,000 enterprise clients and developers. User creation, data feedback, model iteration, experience improvement—each revolution of the flywheel directly translates into revenue.
Viewed through this framework, the core metrics for AI companies also need to change. ARR, GMV, and SaaS efficiency remain relevant but are insufficient to indicate whether a company possesses long-term moats. A feature might rapidly depreciate due to a base model upgrade; a system that truly integrates into business processes, accumulates proprietary data, and continuously learns could become stronger as models improve.
Liu Yunjie further divides the intelligence value chain into three layers: the generation of intelligence, the distribution of intelligence, and the manifestation of intelligence in the physical world. World models and drug discovery belong to the generation of intelligence; Agents and enterprise Agents handle the distribution of intelligence; embodied intelligence and autonomous driving manifest intelligence as physical actions. This framework breaks down the "AI赛道" into a complete value chain. Capital is no longer just searching for a hot application, but for companies at each layer that can occupy a platform position and form a feedback loop.
Therefore, the focus of large model investment is shifting. The past question was "whose model is stronger." The upcoming question is "who can turn model capability into a system that grows stronger with use and deeper integration into business."
Trend Three: Data Bottlenecks Spur New Foundational Technologies; Scientific Foundation Models Become a Non-Consensus Direction
As large models continue to expand, they will eventually face a more fundamental problem: high-quality human data is finite, and its collection and labeling are expensive. When internet corpora and human feedback gradually approach their limits, where can models obtain new experiences?
Hu Qi sees reinforcement learning, self-play, and continuous iteration as potential answers. He cites AlphaGo Zero as an example: the model only knew the rules of Go and generated experience through self-play, improving its ability without relying on human game records. Whether similar mechanisms can move from rule-clear games to more complex scientific research, engineering systems, and real-world tasks will determine whether AI can break its dependence on existing data.
This also provides a set of criteria for judging foundational technologies. First, it must bring universal improvement in a sufficiently large industry, not just benefit a single company's localized环节. Second, it must expand AI's capability boundaries, such as solving data, reasoning, or continuous learning problems. Finally, it must withstand the test of time, not just hold true within one cycle of technological narrative.
Following this logic, scientific foundation models become a non-consensus direction worth attention. This is not entirely the same as the common AI for Science. The latter is often understood as lab automation, protein structure prediction, material discovery, or single-point research tools. Scientific foundation models attempt to establish more foundational, more general model capabilities in fields like life sciences, materials, and simulation, serving multiple tasks within a discipline.
The challenges of this direction are also apparent. Scientific data is characterized by high specialization, inconsistent standards, and high acquisition costs; research results themselves may contain errors or even fabrication. Models must not only learn from literature and experimental data but also judge evidence quality, understand disciplinary规律, and form a validation cycle with real experiments. It remains uncertain whether future development will lead to separate foundation models for each major discipline or a unified跨学科 architecture. But the problem it points to is clear enough: as the marginal value of general internet data declines, high-quality data from specialized domains, scientific laws, and experimental feedback may become a crucial source for the next wave of model capability growth.
Trend Four: The Closer to the Deep Waters of Hard Tech, the More Patient Capital is Needed
Compared to AI applications, commercial space, quantum computing, space-based computing power, and advanced energy have longer validation cycles. They can rarely prove themselves with data from a few months after a product launch. Technology roadmaps, engineering capabilities, supply chains, qualifications, industry standards, and industrial demand must mature together over a longer timeframe.
Wang Shuhe, Vice President at Houxue Capital, expressed such restraint in his assessment of quantum computing. Current approaches like photonics, neutral atoms, and ions each have their own advantages and bottlenecks. When general-purpose quantum computing will materialize remains highly uncertain. Beyond engineering issues like device design and supply chains, the industry also faces fundamental challenges in materials, algorithms, physical mechanisms, talent, and evaluation standards. A more realistic path might be for quantum computing to form modular collaboration with classical supercomputing, GPUs, and CPUs, rather than expecting general-purpose quantum computing to independently replace existing systems in the short term.
Uncertainty does not equate to a lack of investment value. Moderate heat in frontier industries attracts talent, capital, and industrial resources. Even if not all companies succeed, the process still accumulates engineering experience, supply chain capabilities, and transferable technical成果. The key is that capital cannot directly convert distant possibilities into short-term certainties, nor can it completely abandon early-stage judgment just because commercialization is far off.
The same applies to commercial space. This industry involves launches, approvals, licensing, testing, quality management, and supply chain coordination, with a chain far longer than typical software products. But its industrialization is accelerating: China completed 87 space launches in 2025, of which private commercial rocket companies executed 23, deploying 324 orbital spacecraft. Total industry financing reached 18.6 billion RMB, a year-on-year increase of 32%, and commercial launch service orders grew by about 40% year-on-year. Venture capital cannot wait until companies have completely crossed the "valley of death" to enter, as the window for early-stage investment often closes by then. Investors need to find the constants amidst change: will market demand ultimately emerge, does the team possess underlying R&D and engineering organization capabilities, can it handle supply chain and commercialization issues, and does it have sufficient resilience to wait for the industry's door to truly open.
Wang Shuhe's focus on space-based embodied robotics is a concrete example of this judgment. It sounds like an amalgamation of hot concepts like commercial space, robotics, and AI, but clear demand is already emerging. As the number of satellites increases, end-of-life satellite disposal, space debris清理, on-orbit maintenance, and on-orbit construction will become part of the space ecosystem. As commercial space shifts from "how to get equipment up there" to "how to use and maintain space resources long-term," new infrastructure and service systems will appear.
The list's attention to space-based computing power, aerospace energy materials, and quantum computing shows that young investors are not limiting their gaze to quickly monetizable AI applications. They are also entering areas where technology roadmaps have not fully converged but may reshape the foundational elements of future industries. Patience in hard tech doesn't mean forsaking returns, but placing those returns on a longer time scale. Such investment requires capital to be both early enough and patient enough.
As AI moves from the cloud to devices, from digital content to physical actions, and as tech investing expands from software to chips, energy, robotics, and space infrastructure, the role of capital will also change. It needs to understand technology earlier, engage with industry more deeply, and make long-term choices when outcomes are still unclear. What makes this list worth observing is precisely how this generation of investors will learn to make such choices.
This article is from Huxiu, Author: Chen Yifan_YF Original link: https://www.huxiu.com/article/4877273.html?type=text
