admin The shift marks a notable change in the way one of the world’s largest cloud operators builds the internal networks that connect servers, storage systems and routers across its data centres. The new design, called Resilient Network Graphs, moves away from the conventional “fat-tree” model that has dominated hyperscale cloud infrastructure for years and uses a more distributed layout intended to create multiple efficient paths for data.
AWS says the architecture can cut the number of routers by about 69 per cent, increase data throughput by up to 33 per cent and reduce network equipment electricity consumption by around 40 per cent. Internal cost comparisons indicate that the system could be as much as 45 per cent cheaper than traditional designs at scale, depending on workload and deployment model.
The technology has already been deployed in production facilities in Europe, beginning near Dublin at the end of 2024, with additional rollouts in Germany and Spain. By April 2026, quasi-random wiring had become the default design for most new AWS data centre builds globally, although older sites are expected to move gradually as hardware refresh cycles allow.
The network redesign comes as cloud providers face sharply rising demand from artificial intelligence, video, enterprise software, financial services and data-heavy consumer applications. Generative AI workloads, in particular, require high-bandwidth, low-latency communication between large clusters of processors, making the efficiency of internal data centre networks more commercially significant.
Traditional fat-tree networks use a layered structure, with traffic passing through set tiers of switches and routers. That model is predictable and relatively easy to operate, but it can create bottlenecks when traffic flows become uneven or when workloads require large numbers of servers to exchange data at high speed. AWS’s new model introduces a flatter arrangement in which routers are connected through a partly fixed and partly randomised pattern, giving data more possible paths across the network.
A central part of the system is a passive optical component known as ShuffleBox. It allows AWS to preserve the benefits of random graph-style connectivity without turning physical cabling into an unmanageable engineering problem. The device does not require electrical power and helps standardise connections between routers and server rooms, reducing the complexity that would otherwise come with quasi-random wiring across large sites.
AWS also developed a routing system designed to spread traffic across multiple available paths rather than relying on rigid, pre-determined routes. That approach is intended to improve resilience when links become congested or fail, while also raising the total amount of traffic the network can carry under demanding conditions.
The practical significance for customers is that the change should be invisible at the application level. Database queries, storage operations, cloud APIs and machine learning jobs should continue to run without code changes, while gaining from a more efficient underlying fabric. For Amazon, however, the gain is strategic: fewer routers mean lower capital spending, reduced power draw, less cooling demand and simpler operations at a time when data centre expansion is constrained by electricity supply in several markets.
Europe has become an important testing ground because cloud operators are expanding capacity while facing pressure over grid connections, land use, energy consumption and sustainability targets. Data centre developers in some European markets have faced long waits for power connections, making efficiency improvements commercially valuable beyond their engineering appeal.
AWS’s move also reflects a broader industry race to redesign infrastructure for AI-era computing. Cloud providers are investing in liquid cooling, higher-density server racks, custom chips, optical links and more flexible data centre layouts. The network layer is now part of that competitive push, as the cost and speed of moving data between processors can determine how efficiently expensive AI hardware is used.
The company’s work draws on long-standing academic research into random graph networks, an area that has promised strong performance and fault tolerance but was difficult to deploy in real-world facilities because of cabling and routing challenges. AWS’s claim is that its combination of quasi-random topology, passive optical hardware and new routing software has made the concept workable at hyperscale.
Competitors are pursuing their own approaches. Google has used optical switching in machine learning supercomputers, Microsoft continues to expand specialised AI infrastructure, and several chip and networking suppliers are pushing faster Ethernet, InfiniBand and optical interconnect technologies for large accelerator clusters. AWS’s decision to make Resilient Network Graphs a default design signals that data centre networking is becoming a core battleground in cloud economics rather than a background engineering function.
The financial implications could be considerable if the technology performs consistently across regions and workloads. Amazon’s cloud division remains a major profit engine for the group, and even modest efficiency gains can translate into large savings when applied across hundreds of facilities and millions of servers. Lower equipment counts may also help AWS respond to environmental scrutiny by reducing embodied hardware demand and ongoing electricity use.
