The real hardware story isn't about chips anymore—it's about the software stack eating everything. vLLM, ollama, and llama.cpp hitting constant release cycles while AMD quietly demolishes CUDA's moat with ROCm means inference is becoming commoditized infrastructure, not a competitive advantage. Nvidia's letter defending open-source AI is damage control from a company that built its empire on proprietary lock-in; they're getting undercut by projects that cost nothing and run anywhere. Watch the Roku price hike and SpaceX valuation as harbingers—when hardware makers can't justify their margins, they're dead money, and the winners are the platforms that abstract away which GPU you're using.