The growth of artificial intelligence is changing a less visible part of digital infrastructure: network traffic. Until recently, many cloud operators, CDNs, and storage providers were accustomed to more evenly distributed flows, with many users, destinations, and relatively predictable patterns. AI is introducing another logic: large volumes of data moving between a few points, in intense bursts, with bandwidth needs that are much harder to forecast.
The Q1 2026 network statistics report from Backblaze provides figures and context for this shift. The company, known for its cloud storage services, observes how flows related to neoclouds, hyperscalers, and AI workloads behave differently from traditional CDN, hosting, or regional ISP traffic. The key point is not just how much traffic is moving, but how it moves.
From distributed traffic to large flows between few points
AI workflows do not only require storing data. They need to repeatedly move data throughout the model lifecycle. Datasets are ingested, transformed, exported for training, retrieved for evaluation, re-stored, and updated as models evolve. When dealing with multi-petabyte datasets, each of these phases can become a significant network operation.
Backblaze summarizes this as a shift from “diffuse” patterns typical of the internet towards large, high-bandwidth flows between fewer endpoints. For network engineering teams, this difference is transformative. CDN traffic is usually distributed across many users and locations, which aids load balancing and growth modeling. AI traffic, by contrast, may concentrate on a few GPU clusters or compute networks, with peaks occurring during training windows or dataset updates.
This type of traffic is often referred to as “elephant flows”: massive, concentrated streams that are difficult to absorb if the network isn’t prepared. They differ from video streaming, traditional web hosting, or residential traffic. They involve heavy transfers between storage and compute, often constrained by limited GPU cluster availability. When a company gains access to a block of GPUs for hours or days, it needs to move data quickly, run the workload, and extract results before the window closes.
This introduces new pressures on infrastructure. Scaling linearly is not enough. Capacity must be available to handle bursts, internal data center links must support these speeds, with 100G or 400G ports, private interconnection agreements, and architectures that can prioritize large flows without degrading other services.
A winter pause and a rebound in March
The first quarter of 2026 showed an interesting dynamic. Backblaze noted a slowdown during winter months in traffic associated with neoclouds and hyperscalers, followed by a rebound in March. At the same time, CDN traffic increased during winter, while regional hosting and ISP traffic stayed within more predictable levels.
The company offers two possible explanations. The first is seasonal: AI projects also depend on human teams, budgets, development schedules, and periods of lower activity. If teams pause or slow down over winter, traffic can decrease. The second relates to data lifecycle. If a large dataset is already stored, it may not move for weeks or months until a new update, training, or evaluation triggers another massive transfer.
Percentages help clarify this shift. According to Backblaze, combined neocloud and hyperscaler traffic dropped from 36.4% of total volume in Q4 2025 to 25.5% in Q1 2026. Meanwhile, CDN traffic rose from around 20% to 32%, and regional ISP traffic increased from 21.5% to 27.8%. The takeaway isn’t that AI is losing importance; rather, its flows can be less regular and more cycle-dependent.
This irregularity complicates planning. A network designed for stable growth may fall short when a client moves large volumes to a training cluster. Overbuilding for unpredictable peaks also has costs. That’s why AI network engineering leans more toward scenario planning than simple monthly traffic projections.
Where AI traffic is concentrated
Backblaze’s geographic analysis shows a strong concentration of neocloud, hyperscale, and CDN traffic in the United States. The company notes that the U.S. hosts around 40-45% of global data centers, aligning with observed concentration. Within the U.S., California stands out for neocloud traffic, while hyperscaler activity predominantly appears in California and Virginia, especially along the Ashburn and Reston corridors.
This concentration is not surprising. Ashburn and northern Virginia have long been major interconnection hubs and data center regions worldwide. California combines cloud providers, tech companies, IA demand, and proximity to parts of the software ecosystem. What’s interesting is that AI traffic isn’t spread evenly across all newly announced data center regions; existing infrastructure and established network routes still carry significant weight.
Outside the U.S., Backblaze detects neocloud activity in Finland, Brazil, France, and Canada. CDN traffic is prominent in the Netherlands, in part due to connectivity with AMS-IX, while hosting activity is notable in Germany. Europe also shows a different dynamic from the U.S., with more local exchange points and interconnection decisions driven by cost, network policies, and operational preferences.
For Spain and Southern Europe, the takeaway is clear: attracting AI data centers depends not only on energy, land, and tax incentives. Interconnection, routing, proximity to customers, network capacity, private agreements, and large bandwidth availability also matter. AI doesn’t forgive bottlenecks between storage and compute.
What this means for infrastructure teams
Backblaze’s report highlights a practical consequence: network teams must manage two worlds simultaneously. One is the steady traffic of CDN, hosting, and regional ISPs, with more predictable curves. The other is the more concentrated, dynamic traffic of neoclouds and hyperscalers, with peaks that require capacity jumps.
To handle neocloud and hyperscaler traffic, Backblaze recommends adding bandwidth in 100G increments, often with 400G ports; ensuring internal data center links support bursts; and establishing private interconnections with specific partners when appropriate. This approach resembles high-performance wholesale infrastructure more than traditional web service scaling.
The growing diversity of models—training not only with text but also with images, audio, video, and synthetic data—further amplifies this need. Larger datasets and increased movement between storage, training, evaluation, and inference mean operational traffic will rise.
For companies building AI products, this reality carries cost implications. Moving data isn’t free. Storage provider choice, region, GPU cluster, interconnection, and pipeline architecture can impact both performance and cost. In AI, computing power is critical, but network strategy is becoming equally important.
The takeaway isn’t to move everything to private networks or specialized data centers; rather, AI workloads behave differently from typical web traffic. Organizations training, tuning, or evaluating models with large datasets must design storage and network solutions for repeated movements, bursts, and proximity of data and compute.
AI is making traffic larger, noisier, and less predictable—necessitating a rethink of how we size cloud networks. The next competitive edge won’t just be access to GPUs; it will be ensuring data is delivered timely, congestion-free, and at a cost that doesn’t eat into project value.
Frequently Asked Questions
What is neocloud traffic?
It’s the traffic associated with AI-focused compute networks, typically linked to GPU clusters, training, inference, or model evaluation. It involves large transfers between a few points.
Why is AI traffic harder to predict?
Because it depends on training cycles, dataset updates, cluster availability, and work windows. It can be low for weeks and then spike suddenly when moving large data volumes.
What’s the difference between AI traffic and CDN traffic?
CDN traffic generally distributes content to many users and destinations. AI traffic tends to concentrate on large flows between storage and compute, with fewer endpoints and high bandwidth needs.
What should AI-focused companies consider?
They need to plan where they store data, where they run compute, how much bandwidth they require, transfer costs, and whether private interconnections or proximity to clusters are necessary.