There were plenty of models the size of gpt3 in industry.

The core insight necessary for chatgpt was not scaling (that was already widely accepted): the insight was that instead of finetuning for each individual task, you can finetune once for the meta-task of instruction following, which brings a problem specification directly into the data stream.