There are also diminishing returns from additional data if it does not bring vastly new scenarios to train against.
Access to raw data alone is rarely a problem, so I would personally doubt that Google would have much value in Google Photos for training.
That's not true. There is a large and growing body of research on semi-supervised, self-supervised, and unsupervised learning that can take advantage of these unlabelled images.
Perhaps semi-supervised could be utilized, which reduces the required annotation by some factor k, but still leaves it as a function of the dataset.
Self-supervised basically replaces human annotation with machine annotation, making it only applicable to a small subset of tasks in which this is possible (e.g. you could train "guess time from picture" using EXIF timestamp).
Unsupervised is only applicable to very specific tasks.