AWS's math-powered network cuts data center routers by 69%
AWS has deployed Resilient Network Graphs (RNG) in its data centers, cutting routers by 69%, boosting throughput 33%, and reducing network energy use by 40%.
Data
Amazon Web Services today published a detailed technical account of a data center networking architecture it has been quietly rolling out since late 2024 - one built on three decades of mathematical theory that was long considered too impractical to deploy at commercial scale. The architecture, called Resilient Network Graphs, or RNG, is now the default network design for most new AWS data center builds globally, according to Amazon Science.
The announcement on May 28, 2026, marks what AWS engineers describe as the first ever scalable flat-network data center deployment based on random graph theory in production at hyperscale.
What RNG replaces and why it matters
To understand RNG's significance, it helps to understand what came before it. For decades, cloud and enterprise data centers have relied on a design called the fat-tree topology. The structure resembles a corporate org chart: servers connect to edge switches, edge switches connect to an intermediate layer, and that layer connects to a top-of-tree core. When two servers communicate, their data travels up through the hierarchy until it finds the right branch, then descends to the destination.
This approach works, but it has persistent weaknesses. According to Amazon Science, the layered structure is inefficient - the extra routing hops add latency overhead - and the top-of-tree switches become congestion bottlenecks when traffic concentrates. Worse, the architecture is fragile: losing a single top-tier router can sever connectivity for large sections of the network beneath it.
Mathematicians identified a better theoretical alternative as far back as the early 1990s. Random network topologies, where each router simply connects to a few others at random, produce an abundance of distinct paths between any two points. That path diversity all but eliminates bottlenecks. And because no single router is structurally more important than another, failures degrade performance in a predictable, proportional way rather than causing cascading outages. According to Amazon Science, losing 1% of routers in a random-topology network causes roughly 1% capacity loss - not the catastrophic, concentrated failures that fat-tree designs are prone to.
The catch is implementation. Calculating efficient routes through a random graph is computationally intensive, demanding memory resources 20 to 80 times beyond what standard commodity routers carry. And physically cabling routers in a random pattern produces a data center threaded with wires running in every direction, making installation and maintenance a practical nightmare.
RNG is AWS's answer to both of those constraints.
The quasi-random topology
Rather than adopting a fully random wiring scheme, AWS engineers designed what they call a quasi-random topology - one that preserves most of the mathematical benefits of genuine randomness while remaining cableable and computationally manageable.
The central physical innovation is a passive optical device called a ShuffleBox. Each ShuffleBox has router-facing ports on one side and connects to other ShuffleBoxes on the other side. Its internal fiber runs follow a specific shuffled pattern, so that connecting ShuffleBoxes to each other produces an overall network topology that is quasi-random at the macro level without any individual cable run needing to span an entire data center floor.
The practical result is significant. According to Amazon Science, when a new server rack arrives, a technician plugs it into an available port on the local ShuffleBox. No rewiring is required elsewhere in the building. The physical installation complexity is comparable to that of a fat-tree network, even though the logical topology is fundamentally different.
Critically, the ShuffleBox is passive - it contains no electronics and consumes no electricity. This is not a minor detail. Traditional network switches are power-hungry devices. By replacing active switching hardware with passive optical components at key junctions in the network, AWS removes an entire category of electrical load from its infrastructure.
How Spraypoint routes traffic
The physical topology is only half the story. RNG also requires a new approach to routing - the protocol that decides how data packets travel through the network. Fat-tree routing is relatively straightforward because the hierarchical structure itself provides directional guidance. A quasi-random graph offers no such obvious structure to exploit.
AWS developed a custom distributed routing protocol called Spraypoint. It operates in two stages. First, a source router "sprays" its outgoing traffic randomly across all of its immediate neighbors. From there, each packet uses a shortest-path algorithm to reach a designated waypoint - a router that has been pre-assigned to feed traffic toward a specific destination. The waypoints funnel packets into a series of concentric rings around the destination, with each ring passing traffic inward to the next until it arrives.
This combination of random initial spray and structured convergence gives Spraypoint nearly twice as many independent routing paths between any two routers as standard shortest-path techniques, according to Amazon Science. The additional path diversity means congestion in one part of the network can be automatically circumvented without explicit rerouting decisions.
The memory requirements of Spraypoint fit within commodity router hardware - a requirement the team treated as non-negotiable for cost reasons.
The numbers
The performance and cost figures AWS has published for RNG are substantial. According to Amazon Science, compared with fat-tree designs, RNG uses 69% fewer routers and delivers up to 33% better throughput. AWS projects a 40% reduction in network equipment electricity consumption. Infrastructure cost savings are estimated at between 9% and 45%, depending on configuration, according to a research paper co-authored by AWS engineers and posted to arXiv in April 2026.
These percentages translate into concrete sums at AWS's scale. The company plans to spend roughly $200 billion on data center and AI infrastructure in 2026 alone. A 9% to 45% infrastructure cost reduction across that envelope represents tens of billions of dollars in potential savings over the investment cycle.
The router reduction is perhaps the most striking individual figure. A 69% cut in networking devices means fewer units to procure, fewer to power, fewer to cool, and fewer to replace when they fail. For a company operating data centers across dozens of countries, the operational compounding of that reduction is significant.
AWS has also separately noted that its modular data center components are designed to deliver up to 46% lower mechanical energy usage for cooling, without increasing water consumption per megawatt. The company reported a global Power Usage Effectiveness of 1.15 across its 2024 data centers. PUE is the industry's efficiency benchmark: a score of 1.0 would mean every watt goes to computing. AWS's figure compares favorably with industry averages but remains a target for further reduction.
Validating before building
Operators committing to a new network architecture need confidence in performance forecasts before breaking ground. Fat-tree networks come with well-established mathematical models for predicting capacity and behavior. No equivalent existed for quasi-random topologies.
AWS developed new mathematical models for RNG covering path lengths, route counts, and link-level traffic distribution. According to Amazon Science, the team validated these models using 530 processor-years of simulation - the equivalent of running a single processor continuously for five centuries - executed on Amazon EC2. The models give network operators precise formulas to compute the cheapest compliant topology for a given server count and performance target.
That capability matters for planning at scale. Before committing to a capital-intensive build, AWS can now specify performance requirements, calculate what an RNG-based design would cost, and have mathematical confidence in the outcome.
From Dublin to global default
The first RNG network went live near Dublin, Ireland, at the end of 2024, carrying live production traffic. The deployment served as validation: AWS engineers measured real-world performance against the mathematical predictions, identified operational refinements, and applied them in two subsequent deployments. According to SiliconAngle, the technology is already operating across multiple data centers in Ireland, Germany, and Spain.
By April 2026, RNG had become the default architecture for most new AWS data center builds worldwide. In end-to-end benchmarks across production fabrics, AWS says the flat topology matched fat-tree performance for both multipath transport workloads and latency-sensitive storage operations, with no changes required from customers or their applications.
The European deployments are not coincidental. As PPC Land reported in March 2026, Amazon raised its total planned investment in Spain to 33.7 billion euros - the largest infrastructure commitment the company has ever made in that country - covering activity through 2035 and centered on the AWS Europe (Spain) Region in Aragon. Germany and Ireland have been the other primary anchors of Amazon's European cloud build-out. RNG is now the foundation on which that European expansion is being constructed.
Relevance to the advertising and marketing industry
For the marketing community, this story plays out several layers below the campaign dashboard but carries direct consequences. The performance and economics of cloud infrastructure set the boundary conditions for what advertising technology platforms can build and at what cost.
As PPC Land has tracked extensively, AWS has become the predominant infrastructure layer for the advertising technology sector. Supply-side platforms, demand-side platforms, measurement vendors, and clean room providers increasingly run on AWS compute and networking. A 33% throughput improvement in the underlying network fabric means those workloads run faster and handle higher traffic volumes at the same hardware footprint.
The RTB Fabric product that AWS launched in October 2025 - a managed service purpose-built for real-time bidding with single-digit millisecond latency requirements - and the subsequent custom domain support added in May 2026 both run on the same underlying AWS network infrastructure that RNG is now standardizing. Programmatic advertising places some of the most demanding latency requirements of any commercial workload on network infrastructure: bid requests must be evaluated, routed, and responded to within strict time windows. A network that delivers more parallel routing paths and eliminates fat-tree congestion points is directly relevant to those workloads.
AWS's Q4 2025 results, covered by PPC Land, showed advertising services revenue of 21.3 billion dollars - a 23% year-over-year increase - running on the same cloud infrastructure now being upgraded to RNG. The advertising revenue depends on AWS segment capacity: AWS hit 35.6 billion dollars in Q4 2025 revenue, growing 24% year-over-year.
Efficiency gains at the infrastructure layer can reduce costs for the ad tech companies that build on AWS, though whether those savings pass through to advertisers or accrue to vendors depends on competitive dynamics within each product category.
Authors and the research paper
The Amazon Science post was authored by Giacomo Bernardi, principal applied scientist with AWS's Core Networking organization; Ratul Mahajan, a professor at the University of Washington and Amazon Scholar; and Seshadhri Comandur, a professor of computer science at the University of California, Santa Cruz, and an Amazon Scholar. The underlying research paper was posted to arXiv, catalog number 2604.15261, in April 2026.
The three authors were photographed at the AWS Networking Lab in Cupertino, California, in front of a server rack incorporating ShuffleBoxes - a facility where much of the experimental validation work was conducted.
The arXiv paper describes the deployment as the first large-scale production deployment of expander-based network fabrics, a term from graph theory referring to networks with high connectivity relative to their size. Expander graphs have been studied in mathematics for decades for their robustness and path diversity properties, but applying them to physical data center infrastructure required solving the cabling and routing challenges that AWS has now addressed.
Timeline
- Early 1990s - Mathematicians demonstrate that random network topologies produce optimal routing properties, with high path diversity and proportional failure degradation.
- October 23, 2025 - AWS launches RTB Fabric, a dedicated managed service for real-time ad bidding workloads, built on AWS infrastructure.
- October 29, 2025 - PPC Land reports on AWS becoming central infrastructure for advertising technology, covering the scale of dependence across SSPs, DSPs, and measurement vendors.
- End of 2024 - The first RNG network goes live near Dublin, Ireland, carrying production traffic for the first time.
- January 2026 - AWS launches the European Sovereign Cloud, establishing a separate infrastructure instance in Germany.
- February 6, 2026 - Amazon reports Q4 2025 results: advertising services at 21.3 billion dollars, AWS segment at 35.6 billion dollars.
- March 2, 2026 - Amazon raises total planned investment in Spain to 33.7 billion euros, the largest infrastructure commitment the company has made in that country, covering the AWS Europe (Spain) Region in Aragon through 2035.
- April 2026 - The arXiv paper (2604.15261) co-authored by AWS engineers is published, describing the deployment as the first large-scale production instance of expander-based network fabrics. RNG becomes the default architecture for most new AWS data center builds globally.
- May 14, 2026 - AWS RTB Fabric adds custom domain support, allowing advertising technology companies to route bid traffic through Fabric's private network while keeping existing public endpoints.
- May 28, 2026 - Amazon Science publishes the full technical explanation of RNG, detailing ShuffleBoxes, the Spraypoint routing protocol, the 530 processor-year simulation validation, and production performance results.
Summary
Who: Amazon Web Services (AWS), authored by AWS principal applied scientist Giacomo Bernardi, University of Washington professor Ratul Mahajan, and UC Santa Cruz professor Seshadhri Comandur, all working within AWS's Core Networking organization.
What: AWS publicly detailed the deployment of Resilient Network Graphs (RNG), a quasi-random flat network topology that replaces the fat-tree architecture that has dominated cloud data centers for decades. RNG uses passive optical ShuffleBox devices and a custom routing protocol called Spraypoint to achieve 69% fewer routers, 33% higher throughput, and a projected 40% reduction in network equipment electricity consumption. Infrastructure cost savings are modeled at between 9% and 45% compared to fat-tree designs.
When: The first production RNG deployment went live near Dublin, Ireland, at the end of 2024. By April 2026, RNG had become the global default for most new AWS data center builds. The full technical account was published on May 28, 2026.
Where: Initial production deployments in Ireland, Germany, and Spain. The architecture is now the standard for most new AWS data center builds globally.
Why: Fat-tree networks, which have been the industry standard for decades, suffer from congestion at the top of the hierarchy, fragility when routers fail, and inefficiency from extra routing layers. Random graph theory offered a theoretical solution since the 1990s but was considered impractical due to computational and cabling constraints. AWS's ShuffleBox and Spraypoint innovations resolved both barriers, enabling the first commercial-scale deployment of a flat, quasi-random network topology, with direct consequences for the infrastructure costs and performance characteristics of cloud-dependent workloads - including the programmatic advertising systems that run on AWS.
Members Discussion