Run AI Workloads on Any Cloud with Hugging Face Storage: Zero-Egress via SkyPilot

Hugging Face and SkyPilot have jointly announced a new integration that allows AI workloads to be run on any cloud while keeping data on Hugging Face Storage with zero egress costs. Users can mount a Hugging Face Bucket (read-write) or any model/dataset/Space repo (read-only) into a SkyPilot task using a single hf:// URL and their existing HF_TOKEN. The job can then be launched on any GPU cluster across 20+ clouds, Kubernetes, Slurm, or on-prem infrastructure. Hugging Face charges no egress or CDN fees for reading data, so reading models and datasets from the same bucket is free regardless of where the GPUs run. The mount uses Hugging Face's hf-mount FUSE backend for lazy reads, meaning only the bytes actually touched by the code cross the network, and an on-disk cache keeps repeat reads local. Checkpoint writes go directly to the bucket at speeds up to ~170 MB/s. Xet-backed storage provides content-defined deduplication, so incremental checkpoints, adapter weights, and model variants only store and transfer changed chunks, reducing upload time and storage costs. Authentication uses the same HF_TOKEN across all clouds, eliminating per-cloud bucket keys. A benchmark fine-tuning Qwen/Qwen3.5-4B on AWS, GCP, and Lambda showed the model loaded ready to train in about 30 seconds with zero egress cost. The integration is open source and available via pip install skypilot[huggingface].