Amazon has outlined a new data center networking architecture called Resilient Network Graphs, or RNG, and the headline numbers are hard to ignore. The company says the design can deliver up to 33% higher throughput, use 69% fewer networking devices than traditional architectures, reduce network power consumption by 40%, and lower infrastructure costs by as much as 45%.
Those figures come from Amazon’s own claims, and they have not been independently verified. Still, the announcement is worth attention because networking has become one of the less visible constraints behind cloud performance, AI infrastructure, and large-scale application reliability.
AWS says it has been quietly deploying the design since 2024 and that RNG is now part of the direction for future data center builds. The company has also said the architecture has been used in European facilities, including an initial deployment in Dublin followed by expansion into other sites. A firm public rollout timeline for all regions, workloads, or customer-visible services has not been confirmed.
What Amazon Says RNG Changes
Most cloud buyers tend to evaluate compute, storage, accelerator availability, region coverage, and price first. Network architecture usually sits deeper in the stack, even though it affects how quickly applications, databases, AI models, and distributed services can move data between servers.
Amazon’s pitch is that RNG replaces the more familiar hierarchical data center network model with a flatter design influenced by random graph theory. In a traditional fat-tree-style architecture, traffic moves through layers of switches and routers. That approach is widely used, but it can concentrate traffic in certain parts of the hierarchy while other capacity remains less used.
RNG is designed around a more distributed set of paths between endpoints. Instead of depending mainly on predefined layers, the network creates many possible routes across the fabric. In principle, that can improve resilience and make better use of available bandwidth, though the real-world benefit depends on workload type, congestion patterns, routing behavior, and deployment scale.
For AWS, the business case is straightforward: fewer networking devices, lower power draw, and higher throughput would reduce both capital expense and operating cost. For customers, the question is more specific: whether the architecture produces measurable improvements in application latency, AI training throughput, cross-service performance, or regional availability.
| Amazon claim | Why it matters | Buyer caveat |
|---|---|---|
| Up to 33% higher throughput | Could help data-heavy workloads move traffic more efficiently | Actual gains may vary by workload and service |
| 69% fewer networking devices | Could reduce infrastructure complexity and cost | This is an AWS-reported figure, not an independently verified benchmark |
| 40% lower network power use | Could support lower operating costs and sustainability targets | Cloud pricing effects, if any, are not guaranteed |
| Up to 45% lower infrastructure cost | Could improve AWS data center economics | Savings may remain internal unless reflected in pricing or service capability |
The Two Pieces AWS Says Made It Practical
Randomized or flatter network topologies are not a brand-new academic idea. The harder problem is making them work in very large facilities where routing, cabling, monitoring, repair, and expansion all have to be operationally manageable.
AWS says RNG depends on two main components. The first is a custom routing protocol called Spraypoint. Rather than relying primarily on the shortest path, Spraypoint is described as distributing traffic across many available paths. That matters because a flatter network only becomes useful if traffic can be spread intelligently without creating unpredictable congestion or operational confusion.
The second component is a passive optical device called ShuffleBox. AWS describes it as a way to organize and standardize the large volume of fiber cabling needed for this architecture. That detail is easy to overlook, but physical cabling is often one of the hardest parts of changing data center network design at scale. A topology that works on paper can become expensive or fragile if technicians cannot deploy, trace, replace, and expand it consistently.
AWS Certified Advanced Networking Study Guide
For readers who want deeper grounding in AWS networking concepts, this study guide covers VPC design, hybrid connectivity, routing, monitoring, and network automation. It is most useful for engineers who need structured AWS-specific context rather than a general overview.
As an Amazon Associate I earn from qualifying purchases.
For infrastructure teams watching this from the outside, the important point is not just that AWS has a new topology. It is that Amazon is claiming it has combined routing software and optical cabling hardware into a deployable system. That is the difference between an interesting network design and something a hyperscaler can repeat across facilities.
Why This Matters for AI and Large Cloud Workloads
AI infrastructure discussions often focus on GPUs, custom accelerators, memory, and power availability. Those pieces matter, but large AI clusters also depend heavily on data movement. Training jobs, inference systems, storage layers, and distributed databases all rely on the network fabric to move traffic quickly and predictably.
As AI models and cloud applications spread across larger pools of hardware, network efficiency becomes more important. Faster chips do not help as much if data cannot reach them efficiently, and adding more hardware can expose new bottlenecks elsewhere in the stack.
That is why Amazon’s claimed reductions in hardware and power are commercially relevant. If AWS can build high-throughput networks with fewer switches and routers, it may be able to expand capacity with less pressure on energy use, space, and hardware supply chains. It may also improve internal economics at a time when cloud providers are spending heavily on AI-ready data centers.
For buyers, however, the benefits should be evaluated through service-level outcomes rather than architecture claims alone. Procurement teams and engineering leaders should look for signs such as improved performance in specific AWS services, better availability characteristics in new regions, clearer sustainability reporting, or pricing changes tied to infrastructure efficiency.
What Cloud Buyers Should Watch Next
RNG is not something most AWS customers will directly configure. It sits beneath the managed services, compute instances, storage systems, and AI platforms that customers actually buy. That makes the announcement strategically important but difficult to evaluate from the outside.
The practical questions are narrower than the headline numbers:
- Will AWS identify which regions or services benefit from RNG-backed infrastructure?
- Will customers see measurable gains in throughput-sensitive workloads?
- Will the lower hardware and power claims affect pricing, availability, or capacity planning?
- Will AWS expose any service-level metrics that make the impact visible to buyers?
- Will competing cloud providers respond with similar network architecture disclosures?
Systems Performance, 2nd Edition
Teams evaluating cloud infrastructure claims need a disciplined way to measure latency, throughput, bottlenecks, and system behavior. This book is a practical reference for engineers doing performance investigation across Linux, cloud, and large-scale systems.
As an Amazon Associate I earn from qualifying purchases.
For now, the safest read is that RNG is a significant infrastructure claim from AWS, not a direct buying feature by itself. It may help Amazon build data centers more efficiently and support larger AI and cloud workloads, but customers should wait for workload-level evidence before treating the architecture as a reason to move applications or change vendors.
The most useful near-term takeaway is simple: networking is becoming a bigger part of cloud differentiation. Compute capacity still gets the attention, but the companies that can move data faster, with less power and less hardware, may have an advantage as AI and distributed workloads keep growing.