Amazon's RNG Architecture Flattens Data Center Networks Without Sacrificing Scale

The Quasi-Random Solution to Network Bottlenecks

Amazon Web Services has been quietly deploying a fundamentally different approach to data center networking since late 2025, according to Wired AI. The company detailed its breakthrough in a May 2026 research paper titled “RNG: Flat Datacenter Networks at Scale,” which describes a hybrid architecture that combines structured and random elements to eliminate the hierarchical bottlenecks inherent in traditional designs. AWS VP of Network Engineering Matt Rehder told Wired that this approach “essentially flattens the network,” allowing Amazon to claim a competitive advantage it believes no competitor has yet matched at operational scale.

From Fat-Tree to Resilient Graphs

Data center networks have relied on fat-tree topology since the mid-1980s—a hierarchical design with concentrated, high-bandwidth switches at the top and progressively thinner branches toward endpoints. This architecture works, but it creates a chokepoint: all traffic flows through the top-tier routers, constraining throughput when simultaneous communications exceed available bandwidth at that layer.

Random network graphs have long been theorized as a solution, but according to Brighten Godfrey, a networking expert at the University of Illinois at Urbana-Champaign who was not involved in Amazon’s work, they represent a “mind-bending problem to solve, in general.” The technical challenge lies in randomness at scale: while random topologies can theoretically distribute traffic more evenly, deploying them reliably across thousands of servers introduces complexity in routing, resilience, and operations that previous attempts could not overcome. Wired reports that Amazon’s team, assembled in 2023 and including researchers recruited from academia, designed a novel piece of equipment called the ShuffleBox to automatically manage the cable routing required for this quasi-random design.

Why This Matters Beyond AI

Notably, AWS is not positioning RNG as an AI-training accelerator. Rehder emphasized that AI workloads “are far more coordinated and centrally orchestrated, so they don’t approximate a random graph.” Instead, RNG targets the baseline efficiency of everyday cloud operations—a strategic choice that suggests Amazon sees its advantage not in specialized performance for cutting-edge workloads, but in foundational infrastructure efficiency that compounds across all use cases. If the deployment holds, this could tighten Amazon’s margin on margin-sensitive services like compute and storage, making its cloud platform structurally harder for competitors to undercut.

The fact that AWS has been operating this in production for six months before publishing signals confidence in the design’s stability. The claim of being first to scale quasi-random networking, if validated independently, represents a rare instance of fundamental infrastructure innovation in cloud computing—a domain where incremental improvements usually dominate published breakthroughs.