Artificial intelligence isn’t just transforming how data is processed—it’s rewriting the rules of how it moves. Backblaze’s latest Q4 2025 Network Stats report offers a rare, real-world glimpse into this transformation, revealing how AI-driven workloads are creating unprecedented spikes in data transfer volumes, concentrated around specialized compute clusters.
The report, which tracks production traffic across Backblaze’s global infrastructure, paints a picture of a network increasingly dominated by AI pipelines. Unlike traditional workloads—where data trickles in from countless endpoints—AI operations generate bursts of high-volume traffic between a handful of key nodes, often at speeds reaching up to 1 terabit per second. This shift is being fueled by tools like B2 Overdrive, a direct-connect service launched in April 2025 that links Backblaze storage with neocloud compute environments. The result? A network where data gravity is pulling storage and compute closer together, reducing latency and enabling faster iteration on AI models.
The most striking trend is the concentration of AI-related traffic in specific regions. While Backblaze’s US-West region remains its largest hub—thanks to its proximity to major internet exchanges like Equinix-IX—the real action is happening in US-East. Cities like Chicago, Dallas, Denver, New York, and Northern Virginia (home to the Ashburn-Reston corridor) are emerging as hotspots for neocloud activity. This isn’t coincidental: these locations host much of today’s AI compute capacity, and shorter distances between storage and compute mean faster, more efficient data transfers.
Heatmaps in the report illustrate these patterns in striking detail. Traffic volume maps show US-West dominating overall data movement, but when it comes to AI-specific flows, US-East stands out. The heatmaps also highlight a shift in how data is transferred: traditional internet-style traffic, spread across thousands of endpoints, is being replaced by fewer but far heavier flows between specialized systems. This change is reflected in Backblaze’s magnitude metric—measuring bits transferred per IP address—which reveals that AI workloads often involve sustained, high-speed connections between a small number of endpoints, rather than the many-to-many exchanges typical of web or content delivery traffic.
Another key insight comes from Backblaze’s first quarter-over-quarter comparisons. Traffic destined for other cloud providers surged from 36.2% to 49.6% in Q4, while neocloud-specific traffic saw a slight dip (19.8% to 18.4%). Meanwhile, transfers to hyperscalers jumped dramatically, from just 3.5% to 18%. While these numbers are based on a single quarter and Backblaze’s customer base, they suggest a broader trend: AI workloads are driving more cloud-to-cloud movement, with data increasingly flowing between specialized compute environments rather than being shuttled through traditional internet pathways.
The report also underscores the challenges of managing these new traffic patterns. High-magnitude transfers—often exceeding 100Gbps—require careful infrastructure planning. Backblaze notes that while distributed traffic is easier to balance, concentrated flows between a few endpoints can strain network resources. The company is now tracking these shifts closely, with plans to dive deeper into IPv4 vs. IPv6 adoption, cross-cloud connectivity, and regional concentration trends in future updates.
What’s clear is that AI is no longer an abstract concept in network design—it’s a force reshaping how data moves. The question now is whether these concentrated flows will persist, or if they’re just the beginning of an even larger shift toward AI-optimized infrastructure. One thing is certain: the days of diffuse, internet-style traffic may be numbered.
Backblaze’s next report will offer the first full quarter-over-quarter comparison, providing a clearer picture of whether these trends are accelerating—or if they’re just the first wave of an AI-driven network revolution.
