Run AI workloads on any cloud, store on Hugging Face: zero-egress storage with SkyPilot

Hugging Face has partnered with SkyPilot to introduce zero-egress storage, allowing users to run AI workloads on any cloud provider while storing datasets and models on Hugging Face. This integration eliminates data transfer fees when moving data between Hugging Face storage and compute instances on AWS, GCP, Azure, or other clouds. For developers building AI workflows, this means they can choose the most cost-effective or performant cloud for training or inference without being locked into a single vendor's storage ecosystem. The setup leverages SkyPilot's orchestration capabilities to launch jobs across clouds, automatically fetching data from Hugging Face repositories. This is particularly useful for teams running large-scale training or batch inference that require flexible compute. According to Hugging Face Blog, the collaboration aims to reduce cloud egress costs—often a hidden expense—and simplify multi-cloud deployments. The practical angle for AI workflow builders is straightforward: they can now design pipelines where data resides on Hugging Face, and compute runs where it makes sense financially or technically, without worrying about transfer penalties. This is a shift toward more modular and cost-aware AI infrastructure.