Artificial intelligence is no longer just reshaping compute infrastructure, it is now fundamentally altering how data moves across networks. That’s the central conclusion of Backblaze’s Q4 2025 Network Stats report, which draws on real-world production traffic to reveal how AI workloads are driving new trends in network traffic.
The cloud storage provider’s analysis points to a decisive break from traditional internet traffic patterns. Instead of many-to-many connections spread broadly across regions, AI-driven workloads are producing fewer but far larger data flows, sustained over persistent links between storage and specialized compute platforms known as neoclouds. (Examples of neoclouds include CoreWeave and Nebius.) These patterns reflect the growing influence of large-scale model training and inference on network design.
Backblaze’s data shows that neocloud-related traffic rose steadily from July through November, peaking in October. At the same time, migration traffic, based on large data transfers over private fiber links used to onboard massive datasets, surged between August and October before tapering off. While those peaks have passed, the company says the overall baseline level of network activity has now set at a higher level.
The East-West Split
AI-driven data transfers were heavily concentrated in Backblaze’s US-East region, particularly near dense clusters of AI compute infrastructure in Northern Virginia, New York, and Atlanta. These locations align with where many neocloud operators have built GPU-heavy facilities optimized for training and experimentation. As network admins know, keeping storage and compute in close proximity reduces latency, a necessity for today’s AI pipelines.
The US-West region continues to carry the largest share of total traffic volume, largely driven by consumer-facing workloads. Backblaze interacts with a much larger number of unique IP addresses in the western region, reflecting the mixed nature of content delivery and general internet services. The split between East and West highlights how AI traffic differs structurally from traditional cloud usage.
AI Data Patterns
To better capture these differences, Backblaze analyzed traffic using a metric it calls magnitude, which measures the number of bits transferred per unique IP address. High magnitude values indicate large, sustained transfers involving relatively few endpoints.
In Q4, these high-magnitude flows became increasingly common in regions tied to AI workloads, reinforcing the idea that networks are evolving to support long-running, high-bandwidth connections rather than short, bursty exchanges.
This behavior mirrors AI data patterns. Large collections of images, video, and metadata are ingested and consolidated, then exported to compute platforms for training and testing. As models are updated or retrained, those assets move again, creating recurring waves of heavy traffic.
A More Interconnected Cloud Ecosystem
Early quarter-over-quarter comparisons add another layer to the picture. While based on a limited dataset, Backblaze observed a jump in cloud-to-cloud traffic and a notable increase in transfers to traditional hyperscalers, alongside continued strength in neocloud connections. The company cautions against drawing broad conclusions from a single quarter, but says the numbers point to a more interconnected cloud ecosystem driven by AI.
Taken together, the Q4 data offers an early look at what an AI-native network looks like in practice. Traffic is heavier, more concentrated, and more tightly coupled to where compute lives. As AI workloads continue to scale, Backblaze expects these patterns to persist, pulling storage and networking into closer alignment and accelerating the transition away from the open, many-to-many internet model that defined earlier eras of cloud computing.
