Zhu Xiaohu on AI Valuations, Applications and Startup Investing

At the end of 2025, Zhang Xiaojun again interviews investor Zhu Xiaohu about AI applications, valuations, startups and investment cycles. The episode contains personal views and forecasts, not current market facts or investment advice.

Zhu argues that AI competition was shifting from demonstrating model capabilities toward daily active users and engagement. He sees those measures as clues to building a major consumer entry point.

He discusses ChatGPT group chats and speculates that groups could create new social relationships and use cases. He acknowledges that whether the product would become a social network was unclear.

On China, he sees potential in several large companies competing for consumers while considering the landscape early. His rankings and assessments reflect his personal view at the end of 2025.

He sees competition between Google and OpenAI for application entry points and treats group chat as an experiment that might change usage, not a settled competitive outcome.

On subsidy competition among platforms, Zhu questions the efficiency of heavy spending and thinks the contest for AI applications may offer a different long-term opportunity.

Asked about an AI bubble, he says he did not expect a near-term burst, citing rising compute demand and slower infrastructure construction. This is a contestable forecast, not a guarantee about markets.

He believes further model improvements may be expensive even as application demand grows. Zhang asks who ultimately bears the cost, raising questions about subsidies, pricing and sustainable monetization.

Zhu describes low model-access prices at the time and expects large companies may subsidize strategic applications. His cost and revenue figures are personal estimates, not a universal economic model.

They discuss concerns about compute and financing arrangements. Zhu says real demand for infrastructure matters alongside financial structure; this does not by itself prove that any transaction is free of risk.

The episode discusses investment choices involving Nvidia and OpenAI. Zhu explains his preferences in terms of price, competition and expected returns; these are personal views from a specific time.

He credits DeepSeek with influencing the open-model ecosystem and believes access to more models can support competition. His outlook is a forecast, and the openness of each project depends on its actual license and releases.

He thinks model APIs could increasingly resemble infrastructure services and that price competition may make monetization difficult for independent providers. This is his view of a future market.

Discussing Chinese model companies, he says foundation-model training still requires scarce expertise, but startups need sustainable revenue when large companies offer low-priced models.

Zhu compares the cycle with earlier technology waves, in which infrastructure attracted attention before applications expanded. He predicts application growth in 2026 would further test compute and energy supply.

In robotics investments, he favors projects with clear tasks and customers, citing hull cleaning and massage as examples. His descriptions of individual projects are not independently verified performance results.

He discusses products combining software and hardware and startup ecosystems in Shenzhen, Hangzhou, Beijing and Shanghai. Proximity to suppliers, talent and customers can speed iteration, although competition remains intense.

In vertical applications, he values customer relationships, data and product experience built around a specific workflow. A simple wrapper around a general model may be easier for competitors to challenge.

He compares agent startups with early mobile apps: many grew quickly at first but later struggled to sustain growth. Founders need to consider retention, distribution and differentiation.

Zhu suggests looking for specific problems large companies are less willing to tackle and doing difficult, detailed work. He acknowledges that finding both a large market and a defensible position is hard.

He expects major technology companies may continue to grow but acknowledges that startup opportunities are hard to predict. His figures for future valuations and timing are speculative forecasts, not realized outcomes.

On evaluating startups, he says investors should meet founders and learn the industry. He looks for teams that understand a real application rather than merely following a trend.

On fundraising, he says a small stable business may not suit venture capital. Founders should understand return expectations, time horizons and possible terms attached to capital.

He describes post-investment work as staying in contact with teams, watching direction and important details, and deciding whether to participate in later financing.

Using AI toys as an example, he says user experience matters even in a crowded market. Executing a seemingly simple product well is not easy.

They discuss concentrated venture consensus and club deals. Zhu worries that when many firms chase the same startups for small stakes, valuations and eventual returns can diverge.

He compares competition among some foundation-model companies with earlier AI firms, focusing on technical convergence, pricing pressure and commercialization. An analogy can raise questions but cannot prove that two cycles will end alike.

On OpenAI’s long-term position, he distinguishes creating a new product from defending a market over time, which may demand different operating capabilities. He poses this as a question, not a settled conclusion.

He recommends watching active users, retention and real usage rather than model rankings alone. Rankings can change quickly; sustained value to users is more likely to support a durable business.

His advice to young investors is to develop independent views and test them through research and conversations with founders instead of mechanically following fashionable sectors.

He closes with an outlook on China’s open-model ecosystem, compute and energy, including long-term predictions about international competition. These were his 2025 judgments and depend on many uncertain factors.

The conversation ends with personal routines and leisure. It records one investor’s thinking at a particular time, not investment conclusions to apply automatically.