OP here. Imagery file downloads still dominate today because they’re currently the simplest way to get reliable GPU throughput from disk for geospatial images. While in truth for majority of affordable AWS VMs, network is faster than disk.

I also found that in practice there are much bigger bottlenecks before model training/inference: STAC catalog discovery/filtering, iterating S3 URIs, and Rasterio/GDAL I/O/decoding starving GPUs. Not to mention the difficulty in getting paid sat data.

I open sourced Rasteret last year, its 10x faster than any other tool to read geo imagery, but it wasn't enough.

I was inspired by Youtube and how they didn't focus on inventing or selling new cameras but standardized the 'player' and its metrics, which helped everyone to monetize. I believe EO needs such a solution, without needing a new file format we can improve the usage of its images,so after 6 months of tinkering i've built an MVP.

I make the case in the blog that this not only improves today's dev experience, but it also prepares the foundation for when Agentic Commerce tries to use EO images. Happy to share private beta access with anyone. Would love to hear feedback