At a global AI conference in Hong Kong, Fu Sheng, the entrepreneur behind Cheetah Mobile and service-robot maker OrionStar, set out a compact thesis for the next phase of China’s tech transition: AI will invert the old human–machine relationship so that machines “orbit” around people. A veteran of three waves of consumer internet products, Fu sketches a commercial roadmap that moves beyond flashy models toward AI-native applications, hardware integration and narrowly scoped robots that can actually run as businesses.
Fu argues the arrival of large generative models creates what he calls an “AI Native” moment — one that requires rebuilding many existing apps. He says AI-native software changes both production and experience: people who cannot code can now use AI to turn ideas into functioning programs, and interfaces will increasingly act on user intent instead of relying on clicks and menus. Cheetah Mobile is already testing a PC assistant spun out of an antivirus product and enhancing light-office applications with AI features.
Hardware matters in Fu’s view. After visiting CES, he points to AI glasses, toys and business devices such as smart voice recorders as signs that embodied AI is crossing an experience threshold. He sees robotics not as a distant sci‑fi prospect but as a partially commercialized field with concrete deployments, and he has put money where his mouth is — founding OrionStar in 2016 to focus on intelligent service robots and acquiring lightweight collaborative-robot maker UFactory in 2025.
The common thread in Fu’s strategy is what he calls “start with the end” (以终为始): find a strong use case first and then build the product and business around it. That pragmatic stance drives OrionStar’s approach to deployment in exhibition guides, restaurant delivery and hotel service robots, where current AI technologies already deliver measurable value and predictable returns.
Fu’s comments sit against a fast‑moving global landscape. He recalls Cheetah Mobile’s early commitment to AI and traces the industry’s jump from pattern recognition to large models that claim a deeper form of understanding. He also notes the accelerating competition and capital flows — from sky‑high private valuations for OpenAI and interest in Anthropic, to new features such as Google’s Gemini shopping tools and the rush by phone makers and apps to add agent capabilities and “personal assistant” functionality.
Despite the optimism, Fu issues a caution: foundational models still have meaningful limits. He says many agent failures and hard productivity problems are rooted not in user interfaces or integration but in the base models themselves, especially when systems are asked to make real decisions. That assessment shifts the battleground from mere product demos to model reliability, domain adaptation and safety.
For global readers the takeaway is twofold. First, China’s AI trajectory combines rapid application-layer innovation with a hardware and robotics push that emphasises monetisable, narrow deployments rather than general‑purpose autonomy. Second, gaps in base-model performance create commercial and technical opportunities for specialised industry models, better evaluation metrics, and system designs that combine models with deterministic engineering.