# Are AI Neoclouds Rewiring Data Center Traffic Patterns? — retained record

> Retained news record from news:datacenterknowledge.

Canonical: https://highsignal.app/data/records/976cda37e3909971

[← sources](https://highsignal.app/data)


source: [news](https://highsignal.app/data/news) (news:datacenterknowledge)  published 2026-05-08 17:20 ingested 2026-05-09 07:45

[original source ↗](https://www.datacenterknowledge.com/infrastructure/are-ai-neoclouds-rewiring-data-center-traffic-patterns-)

## extract

Are AI Neoclouds Rewiring Data Center Traffic Patterns?
Neocloud-driven workloads are shifting data movement toward sustained, high-bandwidth transfers between storage and AI compute, according to a report.
New measurements suggest AI workloads are reshaping data center traffic patterns, consolidating network activity into fewer, larger, and more synchronized flows that sustain extremely high throughput.
Backblaze reports that neocloud traffic is driving sustained transfers between storage systems and compute clusters with observed rates ranging from 100 Gbps to 1 Tbps.
Sameh Boujelbene, vice president at Dell’Oro Group, said the shift is best understood by how these flows behave rather than how many exist.“It is better described as larger, more synchronized, lower entropy elephant traffic than simply fewer flows,” she said.
As a result, operators are rethinking how data center networks handle traffic. Instead of optimizing for large volumes of short-lived, distributed connections, AI clusters now depend on sustained, coordinated data movement among storage systems, GPUs, and backend fabrics. This places new pressure on switching, congestion management, and east-west network capacity in AI-focused facilities.
A growing class of AI-focused infrastructure providers, including CoreWeave and Lambda Labs, is building GPU-dense environments optimized for high-throughput data movement between storage and compute.
Concentrated, High Throughput Transfers
Within these environments, traffic is consolidating around fewer, more persistent endpoints tied to data movement between storage platforms and GPU infrastructure, according to Backblaze. Those connections carry larger volumes of data per flow, with a relatively small set of endpoints accounting for a rising share of total bytes transferred.
The activity also follows distinct phases: large datasets are first ingested into storage systems, then moved in bulk to compute clusters for training, followed by additional transfers to support inference and model updates.
Training and Inference Drive Different Patterns
The drivers behind training and inference differ in ways that transport behavior. Javier Antich, principal AI engineer at Cisco Provider Connectivity Group, said that training preparation involves large transfers between enterprise data sources and the data centers where models run. Inside training clusters, multi-GPU workloads create “elephant flows” as GPUs exchange model state.
“Training jobs may last hours, days, or even weeks or months, depending on the size of the model,” Antich told Data Center Knowledge.
By contrast, inference traffic is more request-driven and bursty, even as aggregate demand rises, Boujelbene said.
Transport behavior is shifting as well. Antich said he is seeing increased use of QUIC, a transport protocol, for inference traffic. “QUIC aggregates multiple streams over a single connection, potentially aggregating more traffic than TCP,” Antich said. In Cisco’s measurements, QUIC accounts for 53% of total flows.
While individual inference requests are shorter, batching can still sustain high bandwidth utilization inside the data center fabric.
Expanding Beyond the Data Center Fabric
Most training activity remains concentrated within the data center fabric, particularly in GPU-to-GPU communication, Antich said.
At the same time, data transfers tied to training preparation can extend across WAN links, and inference traffic increasingly travels beyond the data center as users and applications issue requests.
Agent-based AI workloads could further increase network demand by accelerating and multiplying interactions across systems, he added.
What began as a backend GPU fabric challenge is now spreading into broader Ethernet-based data center networks. As AI clusters adopt technologies such as RoCE and embrace more disaggregated architectures, the same traffic management issues are emerging across a broader portion of the infrastructure, Boujelbene said.
Consequently, network design priorities are shifting from raw throughput toward managing synchronized traffic, avoiding congestion, and maintaining performance under load.
Backblaze said neocloud activity remains concentrated in the US East regions, including Northern Virginia, while expanding into additional regions such as Finland and Brazil.
“The data tells a clear story: AI is reshaping global infrastructure investment,” said Gleb Budman, CEO of Backblaze. “GPU clusters are concentrating not only in traditional high-intensity regions such as Northern Virginia and California, but also in Finland, Brazil, and beyond.”
Backblaze noted that the dataset is early and may not reflect long-term trends.

## retained text

No document body was retained for this record — only its metadata was stored.