Amazon has rewired how its data centers work internally, and the numbers are significant. The new system, called Resilient Network Graphs (RNG), uses 69% fewer networking devices than the old design, delivers 33% faster data movement, and cuts the energy consumed by the network itself by 40%. It is already the default for most new AWS facilities worldwide.
To understand why this matters, it helps to know what the old design looked like. Since the 1980s, data centers have used a structure called a fat-tree: data flows up through multiple layers of switches and routers, reaches the top, then comes back down to its destination. It is reliable, but rigid. Traffic can pile up at certain points even when the rest of the network is sitting idle. Amazon's global data centers currently hold 20 million kilometers of fiber-optic cables connected in this way.
Amazon's new design removes most of those layers. Instead of a hierarchy, servers connect in a flatter, partially randomized pattern with many more possible paths between any two points. If one path is busy, data simply takes another. The engineering challenge was making this work in practice: truly random cabling inside a massive building is a physical nightmare, and routing data efficiently through such a web is far harder to compute.
Amazon solved both problems. The cabling problem was addressed with a new device called a ShuffleBox, a passive optical unit that physically organizes the cable connections without consuming electricity itself. The routing problem was solved with custom software called Spraypoint, which spreads data across many available paths rather than always trying to find a single shortest route. Together, these two inventions made a design that academics had theorized since at least 2012 finally workable at real-world scale.
The deployment timeline is already done, not planned. The first facility went live in Dublin in 2024. Germany and Spain followed. By April 2026, this became the default for most newly built AWS data centers globally. A research paper documenting the work was published on arXiv, meaning independent engineers can examine and verify the claims.
For companies that use AWS to run their software, databases, or AI tools, this is already running under their workloads. Nothing changes on their end. The benefit shows up as better reliability and, over time, potentially lower costs as Amazon's infrastructure becomes cheaper to operate.
For everyone else, the competitive gap is real. Google Cloud, Microsoft Azure, and Oracle have no publicly documented equivalent. The AWS research paper sets a verifiable benchmark that rivals will need to address. Amazon plans to spend roughly $200 billion on data center and AI infrastructure in 2026 alone, so the savings from a 40% cut in network energy across that base are not trivial.
This is not an AI story, interestingly. Amazon's own engineers explicitly said RNG is designed for general cloud workloads, not AI training, which moves data in a more coordinated, predictable pattern that does not benefit the same way. This is about making the ordinary, everyday cloud faster and cheaper to run. That is the part of AWS that touches almost every business that has moved any operation online.