Highlights

• AI model competition is broadening from capability to efficiency, monetisation and global adoption, with cybersecurity emerging as a high-value AI use case. DeepSeek V4.1 Flash highlights lower compute requirements and more dynamic pricing, while MiniMax expands overseas through deeper model adoption. Funding intensity remains high as Z.AI moves upstream into compute infrastructure, while potential IPOs could weigh on sector liquidity. 
   
• Maintain MARKET WEIGHT. Top BUYs: Z.AI and Alibaba. 
  
   

Analysis

• Funding intensity remains high, although September demand is supported by long-term institutional investors. Z.AI has raised HK$70.6b since July, comprising HK$31.4b net proceeds from its July placement and HK$39.3b in September, including HK$15.7b from new shares and HK$23.6b from a zero-coupon convertible bond (CB). The September share placement represents 4.5% immediate dilution, while full conversion of the CB would take the combined potential dilution to 9.4%. The September deal was launched immediately after the July placement’s 60-day issuer lock up expired, while the latest financing introduces another 60-day restriction through mid-November, highlighting near-term funding needs. That said, the investor mix was supportive, with overseas institutions from Asia, Europe and the US forming the majority of demand. Nearly 30 long-only investors participated, including existing holders, alongside close to 20 technology-focused investors. The top 20 investors accounted for over 85% of the equity allocation and over 88% of the CB allocation, suggesting demand was concentrated among larger institutional investors rather than short-term capital.
   
• Cybersecurity is emerging as an increasingly important layer of AI infrastructure. Rising model autonomy and progress in AI-assisted recursive self-improvement could reduce direct human intervention, increasing the need for stronger safeguards. At the same time, cybersecurity could become an early high-ROI enterprise use case for agentic AI, particularly in vulnerability discovery, secure coding and security operations centre automation. Recent incidents highlight the dual-use nature of AI cyber capabilities: an OpenAI autonomous agent gained unauthorised access to Hugging Face systems, while Hugging Face subsequently used Z.AI’s GLM-5.2 to support its investigation. As model intelligence improves, we expect AI’s role in cybersecurity to expand, with GLM-5.3’s stronger cyber benchmarks positioning Z.AI to benefit. Independent evaluation, red-teaming, runtime monitoring and “trusted checkpoints” should also become increasingly important.