That's not true for the most well-known models. For example Meta's LLAMA training and architecture was predicated on the observation that training cost is a dro

That's not true for the most well-known models. For example Meta's LLAMA training and architecture was predicated on the observation that training cost is a drop in the well compared to the inference cost for a model's lifetime.