Hugging Face has opened Storage Buckets, an object storage service for models, datasets and training artifacts priced between $8 and $12 per terabyte. The company positions that rate against AWS S3 at $23 per terabyte and Backblaze’s Overdrive tier at $15, figures Hugging Face publishes directly on the product page. Egress is included, not metered separately, up to eight times a customer’s total stored volume.
That detail matters more than the headline price. Object storage vendors often advertise low per-terabyte rates while treating egress, the cost of moving data out, as a separate line item that ends up dominating the bill once a team trains across multiple clouds. Hugging Face bundles CDN delivery and egress into the base rate, a structure closer to Cloudflare R2’s egress-free model than to S3’s default terms.
That deduplication layer, called Xet, splits files into byte-level chunks using content-defined chunking and skips re-uploading anything that has not changed. Hugging Face says retraining a model that alters only 5 percent of its weights uploads only that 5 percent, cutting data volume by roughly four times in the company’s own workload tests. Storing a raw dataset alongside its processed version without paying to store both copies is limited to the Enterprise and Enterprise Plus tiers.
Hugging Face frames the product as compute-agnostic: data sits in one bucket while training or inference jobs run on AWS, Google Cloud, Nebius or a private cluster, with no need to re-upload data to switch providers. That framing doubles as a retention pitch. Once petabytes of training data live inside a Hugging Face bucket, moving off the platform carries a switching cost similar to the one egress fees were built to enforce, even though Hugging Face waives that particular fee.
The service also folds in AES-256 encryption, audit logging, SSO with role-based access control and a choice of US or EU regions, positioning it for enterprise buyers rather than solo developers experimenting with a side project. Hugging Face is also pitching Storage Buckets at coding agents that run in short-lived environments and need somewhere to persist checkpoints between sessions, syncing with a single command from its CLI.
Teams currently paying S3 or Backblaze egress charges to shuttle datasets between training clusters now have a concrete comparison point. Run a month of actual transfer volume against Hugging Face’s per-terabyte rate and check it against the included 8
egress allowance before signing a migration contract, since that allowance, not the headline price, determines whether the switch saves money.Hugging Face detailed Storage Buckets pricing, egress terms and Xet deduplication on its own storage product page, which lists no publication date.
