Hugging Face launches cloud storage with no egress fees

Hugging Face and SkyPilot launched a cloud storage integration on July 7, 2026. Developers can now run AI workloads on any of 20+ cloud services, reading model weights and datasets directly from Hugging Face with no egress fees.

Why egress became a problem

Teams that train models often hold GPUs at several cloud providers at once, wherever free capacity turns up. Storing data in every cloud is expensive and awkward, and moving data between providers costs ~$0.09/GB at AWS. With datasets of hundreds of gigabytes, traffic costs quickly overtake the cost of compute.

Key facts about the integration:

• Announced July 7, 2026, jointly developed by the SkyPilot and Hugging Face teams
• Hugging Face Storage: $12-18/TB/month, egress is free
• AWS S3: ~$23/TB plus $0.09/GB for egress
• Model loading speed when mounted: up to 500 MB/s, cold start ~30 seconds
• One-command install: `pip install "skypilot[huggingface]"`

How mounting via hf:// works

The new `store: hf` backend in the SkyPilot configuration accepts `hf://` addresses: model repositories, datasets or buckets (Buckets). Authentication goes through the `HF_TOKEN` environment variable.

Two access modes are supported. MOUNT streams data through FUSE: only the requested bytes are fetched, and the file is not copied in full. COPY preloads data through `huggingface_hub` with a disk cache. Hugging Face Buckets are readable and writable (for checkpoints), while model and dataset repositories are read-only, with optional revision pinning.

Hugging Face also added FUSE patches that allow mounting without superuser rights, a standard requirement in Kubernetes, where root is unavailable.

How fast checkpoints are written

The teams tested fine-tuning of the Qwen3.5-4B model on the Multilingual-Thinking dataset on three clouds at once: AWS with an L40S GPU gave ~168 MB/s, GCP with an L4 ~123 MB/s, and Lambda with an H100 ~112 MB/s. The spread between providers is small and falls within the range of typical workloads.

Deduplication through Xet technology adds a further gain: files are split into ~64 KB blocks by content, and re-uploads skip blocks already known. In practice, appending 10,000 rows to a 100,000-row Parquet table transferred ~10 MB instead of ~106 MB, and re-uploading an 8.43 GB blob took 8 seconds instead of 24.

What this means

Egress fees were a hidden barrier that forced teams to keep data and compute in the same cloud. The SkyPilot and Hugging Face Storage integration removes that barrier: teams can choose GPUs by availability and price at any provider without losing money on traffic.

Frequently asked questions

Can checkpoints be written back to Hugging Face?

Yes. Hugging Face Buckets support both reading and writing. Model and dataset repositories are read-only. When writing from a cloud to Hugging Face, the cloud provider's standard egress rates apply; only traffic coming out of Hugging Face is free.

Which infrastructures does SkyPilot support?

SkyPilot works with 20+ cloud providers, Kubernetes clusters, Slurm and on-premises infrastructure. The base image must have glibc 2.34+ and access to `/dev/fuse`.

How many clouds does the

Hugging Face and SkyPilot integration support?

The integration supports more than 20 cloud providers for running AI workloads.

Why is egress expensive when training models?

Moving data between cloud providers costs ~$0.09 per gigabyte at AWS. With datasets of hundreds of gigabytes, traffic costs climb quickly.