>Everything was parallelized on Burla, on a single dynamic cluster that scaled to ~1.7K CPU workers for photo download and CLIP, with 20 A100 GPUs running embedding clusters in parallel on the same cluster.
That's a lot of budget - would have been nice if they'd made an actual donation to the project, instead of pounding the project's servers and bandwidth when there are much better ways to interact with the data.