Amazon, a big investor in AI, has pioneered technology that it says will make its data centers both more resilient and more energy efficient. A 2026 report from Wired states that Amazon claims the new architecture will allow the company to use 69% fewer routers and switches and 40% less energy in its massive Amazon Web Services (AWS) data centers, while delivering 33% higher throughput.
The heart of this advancement is called resilient network graphs, or RNGs. It’s a way to use random cable connections to make a network work more efficiently. It contains two key innovations: On the hardware side, there is a device called ShuffleBox, which randomizes the physical cable connections between network components to make the network structure more efficient. It pairs with software called Spraypoint, a custom traffic routing algorithm specifically designed to work within RNG design.
RNG builds on a theory pioneered by Hungarian mathematicians in 1959 called random network graphs, specifically the Erdős-Rényi model. To understand this, it helps to imagine a graph with several points. The points are connected randomly but, above all, according to fixed rules of probability. This means that you get different final graphs each time, but also graphs whose statistical properties can be predicted. Although RNG is quasi-random rather than truly random, the random elements are governed by strict precepts. The resulting improvement in efficiency means less power-hungry hardware, which in turn means less power consumption on the network, a vital consideration as data centers struggle to find sources to meet their voracious appetite for power without overloading electrical networks and infrastructure.
How the RNG system compares
A typical data center uses a “big tree” network structure. Data flows up and down a stack, with large nodes at the top. These large nodes are powerful server clusters capable of breaking through bottlenecks created by the linear nature of the “tree,” which is a stack primarily composed of switches (which move data within a network) and routers (which move data between networks). Near the “roots”, the paths become thinner. Fat tree structures require masses of wiring and tons of hardware, and while they are generally reliable, they are not particularly efficient.
Instead of the big tree model, in which traffic must take a small number of fixed paths and may be blocked, a random network graph could allow them to move on a larger number of random routes, creating many alternative paths between any two points. The challenge is implementing this idea with physical cabling. Enter the ShuffleBox, which can shuffle physical cables randomly, following the concept of a random network graph. An important part of the design is optical circuit switching (OCS). This method transmits data from a starting point to an end point only in the form of light inside fiber optic cables. It is more efficient than traditional data transmission because, unlike the conventional method, it does not require data to stop at different points, be converted into electrical energy, and then translated back into light at the next node or destination.
Spraypoint helps by doing what it says on the tin: rather than sending data packets along a single best path, it randomly “sprays” them onto multiple neighboring routers. The packets are then picked up by a number of waiting intermediate routers to send them to the appropriate destination. It distributes traffic to avoid traffic jams and reduces crowded hotspots.
Why this advancement is so important
Reducing the power consumption of data centers is essential to ensure that their construction does not decimate the environment and the power grid. For example, Syracuse said there are currently 30 data centers seeking building permits in New York state. If all of these networks are built, they will consume more energy than the nation of Ireland, putting additional pressure on an aging network, already badly deteriorated after years of deferred maintenance and underinvestment.
This is a common problem nationwide. An analysis by Bloomberg showed that customers just an hour’s drive from new data centers could expect to pay up to 267% more on their utility bills than five years ago. This means almost tripling utility costs, which are already at record levels. While not a panacea, Amazon’s RNG technology could have a significant impact, given that the company uses a staggering amount of energy and was revealed in 2025 to have twice as many active data centers as expected (according to SourceMaterial with Bloomberg).
The same report highlights that Amazon’s data centers are spurring the construction of new gas-fired power plants and extending the life of existing coal-fired power plants that might otherwise have been retired. The size of the data centers is also a concern for nearby residents, with the Meta data center in Louisiana measuring nearly 70 football fields long. The environmental impact is also significant. A separate SourceMaterial report from 2025 illustrates how Amazon, alongside Google and Microsoft, operates (and continues to build) data centers, which rely on enormous amounts of water for cooling, in some of the driest regions in the world.
