Hugging Face and SkyPilot Eliminate Cloud Data Egress Fees for AI

What Happened?

Hugging Face, the leading platform for hosting and collaborating on AI models, and SkyPilot, an open-source multi-cloud orchestration system, have announced an integration that eliminates egress costs when storing AI workload results on Hugging Face from any cloud provider. The solution, detailed on the official Hugging Face blog (February 2025), allows users to run training or inference on AWS, Google Cloud, Azure, or local clouds (such as private or on-premise clouds) and upload generated artifacts — models, datasets, logs, checkpoints — directly to Hugging Face repositories without incurring the usual outgoing data transfer fees.

SkyPilot, developed at UC Berkeley, is known for its ability to abstract the complexity of multiple clouds, allowing users to launch jobs with a unified configuration. In this integration, SkyPilot automatically sets up tunnels and network optimizations so that data transfer occurs at no additional cost. According to the documentation, the tool prioritizes internal routes when possible (e.g., if the compute cloud and Hugging Face repository are in the same region) and otherwise uses mechanisms like peering or interconnection agreements to avoid egress charges. This represents a significant shift in the cost dynamics of AI workloads.

Why Is This Important?

Egress fees have historically been a vendor lock-in mechanism in the cloud. Companies like AWS, Google Cloud, and Azure charge between $0.05 and $0.12 per GB transferred out of their network, making portability expensive and discouraging migration between providers. For AI workloads generating terabytes of data (e.g., training large models with hundreds of GPUs), these costs can be prohibitive, reaching tens of thousands of dollars. This integration breaks that barrier by offering a free egress path to Hugging Face, which acts as a neutral, centralized repository. According to an analysis by Andreessen Horowitz (2023), egress costs can account for up to 10-20% of total cloud spending for data-intensive companies. By eliminating this cost, the integration reduces the total cost of ownership (TCO) of AI projects and fosters competition among cloud providers.

Moreover, Hugging Face does not charge for data ingestion (upload), though users pay for storage in their repositories (with free plans limited to 50 GB for models, 100 GB for datasets, and paid options for additional capacity). This means that for many small teams, free storage may be sufficient, completely eliminating transfer and storage costs for prototypes and moderate-sized projects. The integration also simplifies the workflow: data scientists can train on the cheapest cloud (e.g., AWS spot instances or reserved GPUs on GCP) and store results on Hugging Face without worrying about egress charges.

Market Implications

This alliance pressures hyperscalers (AWS, Google Cloud, Azure) to reconsider their egress policies. Historically, these providers have used egress fees as a barrier to retain customers. However, initiatives like Hugging Face and SkyPilot, along with others like Google's Data Transfer Project (which allows free transfers to certain destinations), are eroding that model. According to a Gartner report (2024), 60% of companies already use multi-cloud strategies, and removing egress barriers could accelerate this trend. For hyperscalers, this could mean downward pressure on egress prices or offering free transfers to popular repositories like Hugging Face.

Additionally, the integration accelerates the adoption of multi-cloud architectures in AI, as teams can choose the cheapest cloud for compute (e.g., spot GPUs on AWS, TPUs on GCP, or Azure instances) and store results on Hugging Face without penalty. This is particularly relevant for startups and small teams, for whom egress costs can be a limiting factor. According to Hugging Face's own data, the platform hosts over 500,000 models and 250,000 datasets, and the integration could significantly increase upload volume, solidifying its position as the central repository of the AI ecosystem.

However, the solution depends on integration with SkyPilot, which adds a layer of technical complexity. SkyPilot requires configuring cloud credentials and familiarity with its command-line interface. For teams without multi-cloud orchestration experience, the entry barrier may be high. Nevertheless, SkyPilot is open-source and has an active community, facilitating support. Additionally, Hugging Face recently launched a native integration with SkyPilot on its Spaces platform, allowing jobs to be launched directly from the web interface, reducing friction.

What Readers Should Know

The functionality has been available since the blog launch (February 2025). It requires a Hugging Face account and configuring SkyPilot with cloud credentials (AWS, GCP, Azure, or local clouds). No additional plugins are needed; SkyPilot automatically handles authentication and network optimizations. Hugging Face does not charge for data ingestion, though users pay for storage in their repositories (with free plans limited to 50 GB for models, 100 GB for datasets, and 1 TB of download bandwidth per month). For intensive use, Pro plans cost $9/month (500 GB storage) and Enterprise has custom pricing.

It is important to note that the integration does not eliminate egress costs between clouds for other purposes (e.g., transferring data from Hugging Face to another cloud), only the upload from the compute cloud to Hugging Face. Additionally, transfer speed may depend on region and network; SkyPilot attempts to optimize but does not guarantee performance. Speculation: it is likely that other AI storage providers (such as Weights & Biases, DagsHub, or Comet ML) will follow this trend to attract users, offering similar integrations with multi-cloud orchestration tools. It is also possible that hyperscalers will respond with free transfer programs to specific destinations, as Google Cloud already does with certain partners.

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

In summary, this integration represents an important step toward democratizing AI, reducing costs and technical barriers. For industry professionals, it is a valuable tool that simplifies artifact management and fosters portability. However, its adoption will depend on ease of use and whether other ecosystem players replicate the functionality. Time will tell if this initiative marks the beginning of the end of cloud egress fees.