Hugging Face adds Xet-backed Storage Buckets

This is Hugging Face filling a real infrastructure gap in its platform stack, not just publishing another docs page. If Buckets work as advertised, they make the Hub more credible as an end-to-end home for active ML pipelines instead of just final model and dataset artifacts.

• –The core pitch is practical: mutable storage for checkpoints, processed shards, traces, and logs that change too often for versioned repos
• –Xet-backed chunk deduplication is a strong fit for ML workloads where successive checkpoints and dataset variants share lots of bytes
• –Native support in the CLI, Python, JavaScript, and `fsspec` means teams can drop this into existing training and data workflows without much glue code
• –Pre-warming data near AWS and GCP compute hints that Hugging Face wants to compete on performance and workflow convenience, not just storage semantics
• –The bigger strategic move is platform expansion: Hugging Face is turning the Hub into operational infrastructure for AI development, not only a publishing surface