Amazon has begun deploying a new data center routing architecture called Resilient Network Graphs across most new AWS facilities, replacing the fat tree topology that has been the standard approach in large-scale data center networking for roughly two decades.
The company says the architecture delivers 33% more throughput from 69% fewer routers, and projects a 40% reduction in network infrastructure electricity consumption. AWS made it the default for most new data centers in April, and says three deployments are already carrying production traffic.
To understand why this matters, it helps to understand what fat tree routing is and why it has persisted so long. Fat tree is a hierarchical design, originally developed for supercomputing in the 1990s, that routes packets by moving them up and down layers of switches.
It scaled well enough in the early cloud era, but its fundamental problem is that maintaining high throughput as networks grow requires adding switch infrastructure in proportion to that growth. At today's data center scales, particularly those built around AI training workloads that generate enormous east-west traffic between servers, the hardware requirements become expensive and the opportunity for congestion increases.
The theoretical alternative has been known for years. Random graph topologies, in which switches connect to each other in a flat mesh rather than a hierarchy, are more efficient and more fault tolerant.
An early academic proposal called Jellyfish outlined this approach in 2012. The problem was always practical: implementing a random graph in a physical data center requires impossibly complex cabling between switches at varying distances, and demands that each switch hold a routing table large enough to describe every possible path in the network. Neither constraint was seen as manageable at scale.
Amazon's contribution is a build decision that makes the random graph approach operational rather than theoretical. Its researchers developed a routing algorithm called Spraypoint, which sprays traffic randomly across neighboring switches to take advantage of multiple available paths, then routes packets via designated waypoint switches using a conventional shortest-path algorithm as they approach their destination.
The more important hardware innovation is a device called ShuffleBox, which concentrates the complex inter-switch wiring that random graph topologies require into a single physical unit, eliminating the long cable runs that made previous implementations impractical.
The efficiency gains AWS claims are significant in the context of the industry's current constraints. Cloud providers are facing growing opposition to data center expansion on the grounds of power and water demand, and the cost of grid capacity has become a real limiting factor in siting new facilities.
A 40% reduction in network infrastructure electricity consumption, if it holds at scale, meaningfully changes the operating economics of a new data center and reduces the headline power draw that draws regulatory and community scrutiny. AWS has not had these figures independently verified, but the fact that the architecture is already running production workloads in multiple facilities provides more credibility than a lab claim alone.
The architecture is proprietary, which limits its near-term industry influence. AWS designs most of its own networking hardware, and the capital required to redesign and re-equip an existing data center with RNG is prohibitive. Amazon is applying it only to new builds for that reason, and outside observers have noted that most hyperscale customers and competing cloud providers are unlikely to absorb comparable redesign costs.
The result is that RNG validates the core idea that random graph networking is achievable at scale, but the approach will likely remain an AWS-specific advantage for the foreseeable future rather than a shift that propagates quickly through the broader data center industry.
