Hugging Face and SkyPilot launch zero-egress cloud storage for AI workloads across any cloud

Hugging Face announced on July 7 an integration with SkyPilot, the open-source multi-cloud orchestration framework from UC Berkeley, launching a zero-egress storage solution for AI workloads. The setup lets developers run compute-intensive AI tasks on any cloud provider — AWS, GCP, Azure, or on-premises infrastructure — while storing datasets, models, and artifacts directly on Hugging Face without incurring cross-cloud data transfer fees.

SkyPilot handles the orchestration and scheduling of compute resources across clouds, and the new integration routes the storage layer back to Hugging Face. The key selling point is zero egress: data moving between Hugging Face storage and SkyPilot-managed compute nodes does not trigger the expensive outbound transfer charges that have long plagued multi-cloud architectures.

This is particularly significant for AI training and inference pipelines, which frequently read and write large datasets and model weights. Egress fees have been one of the hidden cost drivers in cloud AI, often catching teams off guard when they try to distribute workloads across providers or move data between regions.

The partnership extends Hugging Face's storage ambitions. The platform recently launched its Buckets storage service, and this SkyPilot integration positions Hugging Face as more than a model and dataset hub — it is evolving into a full-stack AI infrastructure platform that decouples compute from storage.

For AI development teams, the practical benefit is freedom from single-cloud lock-in. Compute decisions can be made based on GPU availability, spot pricing, or regional latency, while the storage layer remains unified under Hugging Face's interface and data governance policies.

SkyPilot confirmed the integration is available immediately through the latest release. Hugging Face highlighted that the zero-egress design applies across training, fine-tuning, inference, and data preprocessing workloads.

The timing is notable. As LLM training and inference demand continues to drive GPU consumption, the ability to flexibly schedule compute across clouds while controlling storage costs has become a critical infrastructure requirement. This integration could accelerate the adoption of multi-cloud strategies in AI development and push the broader ecosystem toward vendor-neutral toolchains.