Sure there is scarcity in accessing Nvidia supply. Consumer GPU pricing is based on them not making any. There is abundance of GPU time available for training that explains the fast pace of new model version releases. Fast model releases also juice inference revenue as shinny new models get investigated, until next week something better comes along. The bottom line is that burning money on GPUs has plenty of GPUs available to burn it on. Labs would be releasing slower if there wasn’t.
Sure there is scarcity in accessing Nvidia supply. Consumer GPU pricing is based on them not making any. There is abundance of GPU time available for training that explains the fast pace of new model version releases. Fast model releases also juice inference revenue as shinny new models get investigated, until next week something better comes along. The bottom line is that burning money on GPUs has plenty of GPUs available to burn it on. Labs would be releasing slower if there wasn’t.