Even if you can’t run this at home, this is still a step toward commoditization & diversification of inference and breaking up of concentrated power. Being able to serve a frontier-class model with 1M$ of capital in GPUs instead of multi-year billion dollar investment in training & research is a big difference