Allen Zhu, Partner at Jinsha River Ventures: After Technology Diffusion, Focus on Channels, Brands, Supply Chains, Costs, and User Entry Points
Allen Zhu, partner at Jinsha River Ventures, reiterated in a class on AI entrepreneurship and investment at Peking University: technological leadership is only a temporary advantage; after diffusion, the competition is about channels, brands, supply chains, costs, and user entry points. He screens projects using three criteria: a sufficiently large market size, the ability to replicate quickly, and long-term defensibility against being swallowed by large companies. He views highly customized businesses that redo processes for each client as more akin to service companies rather than scalable products.
He set retention thresholds for AI products: he hopes to see over 50% retention in the following month and over 20% retention after six months. The number of users does not equate to value; it depends on payment, retention, and the ability to upgrade to higher-value services. The best growth comes from users bringing in other users; social networks, communities, and bilateral platforms rely on network effects rather than long-term customer acquisition. He attributes Tencent's late catch-up to the long-standing presence of entry points like WeChat and QQ; the barriers for Didi and Meituan lie in simultaneously connecting a large supply and demand, not just simple matchmaking.
Regarding AI entrepreneurship, he identified the most dangerous form: creating features that large companies can integrate for free, lacking entry points, data, and network effects, which amounts to providing prototypes for large firms. User entry points are often more important than mere technological leadership. The more standardized the transaction, the better it scales; ride-hailing, which meets the demand from point A to point B, can become a large platform, while highly non-standardized home services are more challenging. High-frequency usage is preferred over low-frequency; the ideal is to use the service at least once a week. Do not educate the market; instead, improve existing validated demands that just have poor experiences by enhancing them tenfold with AI.
At the model level, he advocates shifting from "stacking tokens" to "stacking value": not every task needs the strongest model, but rather the most cost-effective model to complete the task. Mature products should hide models and prompts so that users do not need to choose between GPT, Claude, or DeepSeek. He likens open-source models to Linux, suggesting that once capabilities and costs converge, they may become infrastructure; Chinese software going overseas can also use open-source as a customer acquisition strategy rather than first building an overseas sales team. He specifically mentions role-playing, programming, and video as high-frequency areas with poor experiences that can be improved by models.
On robotics, he continues to oppose developing general humanoid robots first. In 2025, he publicly questioned the commercialization of humanoid robots and withdrew from some projects, shifting focus to tasks like underwater cleaning and industrial assembly, which are clear, have low generalization requirements, and high average transaction values. Jinsha River Ventures has previously invested in Didi and Xiaohongshu but has recently avoided large Chinese model companies, focusing only on applications that can be commercialized immediately.
The market mechanism is changing the selection criteria, not a new financing boom. Buyers are early funds in RMB and USD seeking scalable revenue and retention curves; sellers are model shell teams with only demonstrations and no entry points or payments. Beneficiaries are products with bilateral networks, standardized transactions, and hidden models; those under pressure are projects customized for clients that rely on buying traffic to boost daily active users. Funding is shifting from "which model is stronger" to "who can be used weekly, still around next month, and cannot be easily replicated by large firms."
Source: Public Information
ABAB AI Insight
Allen Zhu's criteria have not changed in ten years: first calculate the numbers, then see if it can be replicated. Didi validated standardized demand amplified by weather; Ele.me validated that campus density can be expanded city by city. He applies the same criteria to AI, concluding not that "the model is not strong enough," but that "features will be integrated for free by WeChat and operating systems." Avoiding large models, withdrawing from humanoids, and focusing on underwater cleaning reflects a consistent rejection: hardware without visible ongoing orders and models without visible entry points are excluded from the portfolio.
The capital path is retreating from token consumption narratives back to unit economics. Charging for software in China is difficult, so he now looks at vertical agents that charge based on results; consumer projects need to show profit levels, and B2B needs to demonstrate revenue steps, otherwise, five-year RMB funds cannot sustain them. Open-source is seen as customer acquisition cost, not a belief: developer communities pay for global distribution. Retention numbers serve as filters to discard daily active users inflated by initial week subsidies.
Analogies should reflect Bezos's "customer experience is irreversible" and Peter Thiel's "competition is what losers do." Tencent's late entry uses existing entry points, not a better search engine; Meituan uses high frequency to drive low frequency. OpenAI treats models as products, while Allen Zhu views models as interchangeable parts. The industry position shifts from comparing parameters to comparing distribution; the phase is application liquidation: those that can hide models, achieve weekly usage, and prevent integration remain, while others become feature lists for large firms.
Structural judgments belong to the transfer of pricing power. Pricing power shifts from training clusters to user entry points and standardized transactions. The mechanism is that as model capabilities converge, differences can only be made in acquisition costs, retention, and network density; whoever controls the entry that opens weekly can replace any open-source model as a cost item. The half-life of technological leadership is shortening, while the half-life of commercial structures is lengthening.
ABAB News · Cognitive Laws
After technology diffusion, entry points are more valuable than models.
Features that large firms can integrate for free do not constitute a company.
Users retained in the following month are more valuable than those who rush in during the first week.