Hugging Face Storage is now a first-class backend in SkyPilot, letting teams mount Hub repos or Hugging Face Buckets into any SkyPilot job via hf:// URLs. The key benefit: no egress fees when reading data onto GPUs, regardless of which cloud the job lands on. This removes the cost penalty that has tied AI workloads to the same cloud as their data.

SkyPilot schedules compute across 20+ clouds, Kubernetes, Slurm, and on-prem. With store: hf, a model or dataset is mounted into a task using MOUNT (lazy, FUSE-based) or COPY (download upfront). The MOUNT mode uses Hugging Face‘s hf-mount FUSE backend, which fetches only the bytes your code actually reads. An on-disk cache keeps repeat reads local. Authentication uses the same HF_TOKEN across all clouds, so no per-cloud bucket keys are needed.

Hugging Face Storage charges no egress or CDN fees, at $12-18/TB/month versus S3’s roughly $23/TB plus egress. Writing back still incurs the compute cloud’s usual egress, but for most AI workloads reads dominate. A benchmark fine-tuning Qwen3.5-4B on H100s across AWS, GCP, and Lambda showed model loading in ~30 seconds at up to 500 MB/s with zero egress cost. Checkpoints wrote to the bucket at 112-168 MB/s depending on the cloud and GPU.

Buckets are built on Xet, which uses content-defined chunking to deduplicate at ~64 KB granularity. This saves on incremental checkpoints, model variants, and appended datasets. For example, re-uploading an 8.43 GB blob already in the bucket took ~8 seconds instead of 24, because only chunk hashes moved. The hf-mount FUSE fixes were upstreamed by the Hugging Face team to work in unprivileged containers, making the solution viable on Kubernetes clusters. The entire stack is open source.