Prioritize AI hardware infrastructure like NVIDIA (NVDA), as persistent compute bottlenecks and massive model-training demand make high-end accelerators the highest-conviction core holding.

Capture near-term cash flow through AI Application Companies, but actively manage platform risk as major model developers increasingly integrate specialized features natively.

Approach pure-play AI Foundation Model Companies with caution due to heavy capital burn and rapid price erosion driven by capable open-source alternatives like DeepSeek.

Begin building long-term growth allocations in Embodied AI and Humanoid Robotics supply chains—particularly sensor and actuator manufacturers—positioning for mainstream physical deployment on a 5- to 10-year investment horizon.

Artificial Intelligence Foundation Models & Frontier Labs

• Frontier foundation model developers (such as OpenAI, Anthropic, Moonshot AI / Kimi, and DeepSeek) are engaged in massive capital expenditure to train increasingly advanced models aiming for AGI (Artificial General Intelligence).
  • Frontier research is capital- and compute-heavy; training even small models independently costs hundreds of thousands to millions of RMB in GPU hours.
  • The performance gap between US frontier models and open-source or domestic alternatives (such as DeepSeek and Kimi) is closing rapidly.
  • Most foundation model developers are currently burning significant cash, with profitability remaining elusive across the sector.

Takeaways

• High Capital Barriers: Frontier foundation model development is largely restricted to well-funded hyperscalers and top-tier labs with immense compute access.
• Erosion of Exclusivity: Open-source and lower-cost alternatives are catching up faster than expected, potentially pressuring long-term pricing power and margins for proprietary foundation model providers.

AI Model Companies vs. AI Application Companies

• There is a clear structural divide between Model Companies (infrastructure/model providers) and Application Companies (software wrappers or end-user solutions built on top of third-party models).
  • In the current market cycle, application-layer software companies often generate revenue and cash flow faster than pure model companies due to lower training compute overhead.
  • Application companies face substantial long-term platform risk: model providers can natively integrate application-level features (such as specialized coding tools, co-workers, or native agent workflows), directly disrupting downstream wrappers.
  • If AGI or highly capable autonomous systems are achieved, advanced base models may independently build software or replace vertical application layers altogether.

Takeaways

• Near-Term vs. Long-Term Advantage: While application companies offer quicker cash-flow realization today, investors must monitor platform risk—specifically whether the underlying model providers (e.g., OpenAI, Anthropic) will cannibalize the application's core functionality.

Semiconductor Hardware & Infrastructure (NVIDIA)

• Advanced AI research and experimental scale-ups depend entirely on top-tier GPU hardware, specifically references to clusters of NVIDIA H100 GPUs.
  • Compute demand remains non-discretionary for scaling data training runs, residual architecture modifications, and competitive model benchmarks.
  • High model iteration frequency and enterprise token consumption continue to drive sustained hardware demand.

Takeaways

• Essential Infrastructure Play: Compute remains the core bottleneck and capital sink in the AI race, sustaining structural demand for high-end accelerator hardware like NVIDIA across enterprise and research workloads.

Embodied AI & Humanoid Robotics

• Embodied AI (Physical AI / Robotics) represents the next frontier after large language models reach logical maturity.
  • The expected timeline for mainstream breakthroughs in embodied robotics navigating public environments is approximately 5 years.
  • Future AI evolution involves multi-modal agency interacting with the physical world through specialized robotic hardware (e.g., multi-arm machines, customized robotic bodies) rather than software-only interfaces.

Takeaways

• Long-Term Secular Growth: Watch for investment opportunities in the convergence of AI reasoning models and physical automation (robotics supply chains, sensors, and actuators) on a 5- to 10-year investment horizon.