The real AI infrastructure play isn't chips or models—it's the unglamorous middleware. While everyone watches Nvidia's stock and chases frontier LLMs, Indian manufacturers are quietly capturing value in power delivery, thermal management, and deployment systems that actually make AI workloads run. llama.cpp's relentless optimization releases (5 builds in your list alone) matter more than most funding announcements because inference efficiency is where 80% of real-world margin lives. The noise around Anthropic's testing access and South Korean stock volatility is just noise—watch instead for who's winning the standardization wars (OSReward's cross-platform reward model framework, the emerging llm-as-infrastructure stack) because that's where lock-in and defensibility actually compound.