I don't feel like compute for pretraining the model was a huge constraint? The supply bottlenecks have been around commercializing the ChatGPT product at scale

I don't feel like compute for pretraining the model was a huge constraint? The supply bottlenecks have been around commercializing the ChatGPT product at scale. But pretraining the underlying model I don't think was on the same order of magnitude, right?