AWS Unveils 102.4-Tbps Cloud Network Stack With Single ASIC and NetOS
Updated
Updated · InfoWorld · May 12
AWS Unveils 102.4-Tbps Cloud Network Stack With Single ASIC and NetOS
1 articles · Updated · InfoWorld · May 12
AWS said its redesigned cloud backbone now runs on a unified in-house stack—one switching ASIC across aggregation, core and border networks, plus a Linux-based NetOS for faster patching and automation.
102.4 Tbps is the capacity of AWS’s next-generation switch, up from 51.2 Tbps today; the current hardware supports 64 ports at 800 Gbps, with future ports reaching 1.6 Tbps.
30% lower latency from hollow-core fiber and a new UltraCluster topology are aimed at AI training, distributed databases and trading workloads that need larger low-latency regions and fewer network hops.
Roughly 2 million devices, 50 million to 60 million optical links and more than 20 million kilometers of fiber underpin the network, letting AWS push fixes and performance upgrades across its infrastructure at scale.
For enterprises, the shift moves network operations away from vendor-specific hardware management toward workload placement, cost control and compliance planning, while raising fresh questions about premium pricing and cloud lock-in.
As AWS's network becomes a black box, are businesses sacrificing control for performance they don't need?
Can AWS's custom network truly dethrone InfiniBand as the gold standard for large-scale AI training?
What is the hidden environmental price of AWS's massive global network upgrade?
Breaking the Bottleneck: How 102.4 Tbps Networks Are Powering the Future of Cloud and AI
Overview
The deployment of 102.4 Tbps network stacks in 2025-2026, led by Broadcom's Tomahawk 6 ASIC, marks a major leap for cloud and AI infrastructure. This innovation directly addresses the long-standing network bottleneck that has limited large-scale AI workloads, especially as trillion-parameter models push traditional systems to their limits. By delivering unprecedented bandwidth and efficiency, Tomahawk 6 enables hyperscale cloud environments to meet the growing demands of generative AI and other compute-intensive applications, redefining what is possible for data processing and communication in the next generation of data centers.