Amazon has implemented a new data center network architecture that accelerates data transmission and reduces energy consumption, according to Wired.
The solution is based on a “quasi-random” topology. AWS claims to have been working on it since 2023, with implementation in infrastructure starting at the end of 2025. For the project, the company developed its own optical device, ShuffleBox, which automatically rearranges cable connections between routers.
Matt Rayner, Vice President of AWS Network Engineering, stated that the team managed to “flatten the network” and eliminate bottlenecks typical of traditional architectures. According to him, Amazon was the first to scale this approach to the level of real infrastructure.
In April, the company released a research paper titled RNG: Flat Datacenter Networks at Scale. RNG stands for resilient network graphs.
According to Amazon, compared to classic networks, the RNG architecture:
- requires 69% fewer routers and switches;
- increases throughput by 33%;
- reduces energy consumption by 40%;
- cuts operational costs by 27%.
The first launch took place in Dublin in 2024. The architecture was then deployed in Germany and Spain. Currently, most new AWS data centers are being built using RNG.
Since the mid-1980s, data centers have predominantly used the fat-tree topology—a multi-level hierarchy of switches. It is reliable but rigid and requires complex cabling infrastructure. Amazon’s network connects about 20 million kilometers of fiber optic cable.
The starting point for the project was the Jellyfish concept, proposed in 2012 by researchers at the University of Illinois. It suggested a random graph topology as an alternative to fat-tree but created issues with routing and cable layout.
According to one of the authors, Giacomo Bernardi, the team initially tested a more regular scheme inspired by Penrose tiling. Simulations showed weak resilience and modest efficiency gains. Ultimately, engineers moved to a quasi-random model.
AWS emphasized that the architecture was not specifically created for generative AI. The training patterns of models are too centralized for a random graph. The focus is on optimizing the basic network infrastructure.
Earlier, the largest regional electricity transmission operator in the US, PJM Interconnection, reported that the data center boom led to additional costs amounting to $23.1 billion in its area of responsibility alone.