It's crazy anyone ever thought there could be a moat. Almost all the research was/is published in the open. There's no secret trick, no hidden method. LLMs are a commodity technology. Anyone can read a paper, write software, and train models. How do people think llama.cpp works? It's not magic... it's software.

Hardware is the real differentiator. Not everyone has billions to make more advanced chips, and only a few companies can make them anyway. Both OpenAI and Anthropic would already be dead in the water if we had cheaper GPUs, because we'd all be running open models on local machines with 8 graphics cards. They're gonna have to force a hardware shortage to prevent a collapse in 2-3 years. My guess is it'll be tariffs or import restrictions or licenses to buy newer hardware.