Is it in model design (now published), data set, engineering needed to train it, cost of hardware needed to train it, cost of HLRF on top of GPT3 to make ChatGPT?
Stuff like GPT3 is trying to be recreated, but even the eleuther AI guys only collected like 800GB training data, which is much less than what OpenAI has (iirc around 45TB). And apparently their data is very high quality. EleutherAI is pretty much one of the few big model open source competitors with GPT-Neox etc.
Plus openai has great branding.
I wonder how LaMDA compares performance wise to ChatGPT. I definitely understand why training on Github is an advantage, but I'd expect Google to also be great at getting a good dataset, across the range of things they'd be interested in.
Their valuation comes from expectations of the value of future iterations of those.