The analogy holds on capital and compute but breaks on the input that actually made LLMs work. Anthropic and its peers inherited a decade of scraped internet text. Humanoid players have no equivalent corpus of real-world manipulation data, and you can't scrape someone folding laundry. Whoever solves the data-collection loop for physical tasks ends up mattering more than which lab has NVDA and Google backing.(Sahara AI 🔆)