Amazon Web Services is pushing a major shift in how data centers are built, and the change could reshape the economics of cloud infrastructure. The company says its new quasi-random network design, called RNG, is built to replace the fat tree topologies that have long dominated large-scale data centers.
At AWS’s scale, even modest infrastructure gains carry outsized impact. Amazon holds 28% of the global cloud infrastructure market in the first quarter of 2026, so any improvement in network efficiency can affect a vast share of internet traffic and server operations.
A flatter network model with fewer bottlenecks
Fat tree designs have remained popular because they are predictable and stable as data center size grows. The tradeoff is that they can be less efficient and more vulnerable to congestion around critical upper-layer routers.
Amazon’s approach follows the idea of flat networking, which links routers and switches more directly instead of stacking them in rigid layers. The company says RNG reduces the chance that one router becomes disproportionately important, making the network more resilient to congestion.
How AWS makes randomness practical
Fully random networks are difficult to deploy in real data centers because of the physical complexity of fiber cabling and the routing overhead they create. To solve that, Amazon combines fixed network elements with randomized components rather than relying on a fully unconstrained design.
The two core parts are Spraypoint, a routing algorithm, and ShuffleBox, a passive optical device. Spraypoint sends packets to all neighboring nodes that could serve as the next hop, then uses Equal-cost multi-path routing, or ECMP, to select a path.
Amazon also uses waypoints to spread traffic more evenly so packets do not pile up at a single destination. This allows the network to capture some of the benefits of randomness without requiring routers to store or process every possible route in special CPU or memory resources.
ShuffleBox handles the physical connectivity of RNG. It is a sealed device that uses no power, adds no latency, and is designed to lower failure risk.
Inside the unit, fiber cables are rearranged in a specific pattern and then connected randomly to other devices. Amazon says the device costs about the same as a standard patch panel, which makes large-scale deployment more realistic.
| RNG Component | Function | Notable Benefit |
|---|---|---|
| Spraypoint | Routing algorithm that sends packets to multiple possible next hops | Helps distribute traffic without heavy router overhead |
| ShuffleBox | Passive optical device that randomizes physical connectivity | No power use, no added latency, lower failure risk |
| Waypoints | Traffic-spreading mechanism | Reduces packet buildup at a single destination |
The reported gains are substantial
In Amazon’s white paper, RNG is described as matching or exceeding fat tree performance across different traffic patterns and workloads. The company also says the design can move data up to one-third faster and can be up to 45% cheaper to deploy.
Those savings come partly from a reduced need for routers. Amazon says the quasi-random network cuts router count by 69%, which lowers hardware costs and simplifies the infrastructure footprint.
The power story is just as important. AWS estimates that data centers using RNG will consume 40% less electricity than hierarchical networks.
Why the change matters beyond bandwidth
Lower hardware counts and lower electricity demand also reduce cooling needs, which matters as data center water and carbon footprints face closer scrutiny. Amazon already uses recycled water to cool some data centers, but RNG is expected to lower the base load for both cooling and power use more broadly.
That direction fits Amazon’s wider environmental targets. The company aims to become water positive by 2030 and reach net-zero carbon emissions by 2040.
Amazon says the first quasi-random network went live in Dublin, Ireland, in 2024. The system then expanded to Spain and Germany throughout 2025, and by April 2026 it became the standard for new AWS data centers.