Hugging Face buys XetHub to ditch Git LFS

If you update a single row in a 10GB dataset file, you shouldn't have to upload all 10GB again. Yet that is exactly what AI developers deal with every single day on Hugging Face.

Hugging Face announced it acquired XetHub, a Seattle-based startup founded by former Apple ML infrastructure veterans. The plan is straightforward: rip out Git LFS from the Hub and replace it with something built for the modern AI scale.

Fixing the Git LFS bottleneck

Hugging Face CTO Julien Chaumond admitted that when they built the Hub in 2020, Git LFS was the obvious choice to get up and running quickly. But while LFS stands for Large File Storage, it was never designed for the massive weights and multi-gigabyte files standard in machine learning today.

Under the current setup, any change to a file requires re-uploading the entire object from scratch. With XetHub's chunking and deduplication technology, that changes completely.

For example, if you add one row to a 10GB Parquet file, you will only need to upload the specific chunks holding that new row. If a community creator wants to modify a single metadata field in the header of a Llama 3.1 405B GGUF file, they will only have to push a single chunk measuring a few kilobytes.

Massive traffic and the trillion-parameter future

Hugging Face shared eye-opening stats showing why this upgrade can't wait. The Hub now hosts 1.3 million models, 450,000 datasets, and 680,000 Spaces. It stores 12PB of data across 280 million files in Git LFS, compared to just 7.3TB in standard Git repositories.

The real pressure is on the network. The Hub currently processes 1 billion requests daily and burns through an astonishing 6PB of CloudFront bandwidth every single day.

XetHub's founding team of 12 includes Yucheng Low, Ajit Banerjee, and Rajat Arya. Low previously sold GraphLab to Apple, where he helped scale internal ML data infrastructure to more than 100PB across dozens of production teams. With trillion-parameter open-weights models just around the corner, Hugging Face needs that exact enterprise-grade architecture to survive the next five years of growth.

Why it matters

For AI developers and open-source contributors, this solves one of the most frustrating papercuts in model hosting. Pushing quick dataset updates or fixing quantization headers will no longer mean waiting around for multi-gigabyte uploads over congested pipes. It finally brings standard software engineering practices to massive model artifacts.

For the AI industry, this acquisition strengthens Hugging Face's moat where it matters most: developer ergonomics and platform margins. Serving 6PB of daily egress bandwidth is brutally expensive, and deduplicating chunks could save significant infrastructure dollars while keeping competitors from catching up on collaboration tooling.

We think buying XetHub rather than building an in-house storage engine from scratch is a very smart move, especially given the founders' track record at Apple scale. That said, migrating 12PB of live data and nearly 300 million files across an active community without breaking existing Git pipelines will be a massive technical hurdle. How smoothly they execute that backend transition remains the key detail to watch.