i was _very_ surprised by that number as well. 10s for inference on 300 chars? using an entire A100 GB GPU? that does not make any sense. OpenAI and other companies would NEVER be able to scale to 100s of millions of users if that was the real number, right?