Network Stats for Q3 2025: The Magnitude of AI Workflows (www.backblaze.com)

Backblaze has expanded its Network Stats snapshots into a regular quarterly transparency report (starting with data captured from August 2025) that publishes anonymized TCP flow telemetry to show where data lives and how it moves in the AI era. The dataset is captured monthly by region and direction (ingress/egress) and records averages, 95th-percentile weighted values, and maximums for bits, packets, flows, and unique IPs. A new “magnitude” metric—bits per IP—combines transfer volume with number of actors as a proxy for how concentrated flows are. Metadata adds BGP ASN grouping and categories (neocloud, hyperscaler, CDN, ISP, etc.), and Backblaze plans quarterly raw data releases and QoQ tracking (including IPv4 vs IPv6 and AI workflow trends). The headline finding: traffic to “neocloud” providers (GPU/AI compute clouds) already represents nearly a quarter of Backblaze’s ingress and egress, and those flows show an order-of-magnitude higher bits-per-IP than traditional CDN/hyperscaler traffic. In practice that means “fewer talkers, bigger flows”: concentrated, bursty transfers tied to dataset replication, checkpointing, and inference pipelines as object storage pairs with external GPU clouds. For operators and ML teams this signals a shift in architecture (storage close to specialized compute), changes to cost and egress modeling, and new operational demands to handle high-peak, cross-cloud AI traffic.