Hugging Face is ripping out its standard S3-and-CloudFront plumbing and building a custom content-addressed protocol from scratch. The numbers driving this rearchitecture are staggering: over a 24-hour snapshot, the Hub handled 8.2 million upload requests and 130.8 terabytes of data from developers in 88 countries. When you’re moving Meta’s Llama-3-70B weights—all 131GB split across 30 separate files—CloudFront’s 50GB file size cap stops being an edge case and starts being a daily bottleneck.
The core idea is deceptively simple. Insert a content-addressed store (CAS) as the first stop for every request. On the read path, a CAS server hands back reconstruction instructions while CloudFront still does the heavy lifting for downloads. The write path is where things get clever. Instead of treating files as opaque Git LFS blobs, the protocol analyzes data at the byte level, deduplicating chunks so clients only push genuinely new data. Early experiments with compressing Safetensors files suggest upload speed gains of 10 to 25 percent are on the table, and the team is also exploring similar tricks for Parquet files.
Enterprise users get a different kind of win. Inserting a control plane means uploads stop being a black box. Malicious or invalid data gets caught before it lands, and enhanced telemetry gives the infrastructure team audit trails and detailed logs—the kind of operational visibility that makes debugging storage issues go from a multi-day headache to a quick query.
Geography matters too. The team analyzed traffic patterns across AWS’s 34 regions and found a classic power-law distribution: seven countries account for 80 percent of uploaded bytes, with the United States alone acting as the dominant source. Rather than chasing global coverage with Lambda@Edge—which proved incompatible with the new upload path—they’re placing CAS nodes in a handful of core regions. It’s a pragmatic tradeoff. The top 20 countries cover 95 percent of traffic, so that’s where the infrastructure will live. For AI researchers in those long-tail regions, the experience won’t degrade; it just won’t benefit from the new protocol’s smart-chunking magic until broader expansion happens. The real question is whether this byte-level intelligence can keep pace as model sizes continue their relentless climb toward the trillion-parameter mark.
