https://www.statista.com/statistics/502208/tesla-quarterly-v...
And I would expect that those generate more data than they can crunch, so that the bottleneck is the number of GPUs they have.
They recently reported that they have the equivalent of 40k H100 GPUs:
https://www.autoevolution.com/news/elon-musk-explains-the-la...
they will be using 85K H100s (up from 35K now) for training by the end of 2024:
https://www.shacknews.com/article/139619/tesla-ai-training-n...
> This vastly expanded the FSD testing pool, with the Tesla fleet adding 1 billion FSD miles every 2-3 months
It doesn't mean they collect the video feed and all input from every second of those miles. Just that the FSD drives those miles.
There's also a matter of deduplication where they really don't need to use ~250 days a year of the same commute. They probably have more data than others, but I'm asking for a proper confirmation of this common knowledge.