AWS replaced fat-tree datacenter networks with a quasi-random graph — and deployed it

Since the mid-1980s, datacenter networks have been built as 'fat-tree' hierarchies: data moves up and down stacked switch layers, and cabling is one of the largest costs — Amazon's data centers alone are connected by 20 million kilometres of fiber. Random-graph topologies like the 2012 'Jellyfish' proposal promised a flexible pool of capacity but failed to scale because routing is harder and cable endpoints are chosen randomly.

From 2023 an AWS team including Amazon Scholars Ratul Mahajan and Seshadhri Comandur hunted for the 'holy grail': flat, efficient, failure-resilient and cable-light. Lead author Giacomo Bernardi first tried Penrose tilings — a repeating aperiodic pattern — but the simulated network was unreliable; outcomes improved only when structured parts were replaced with randomness. 'We embraced the chaos, and adopted a quasi-random approach,' he told WIRED.

The blocker — who cables a random graph? — was answered with the ShuffleBox, an optical device Amazon built that mixes connections between routers internally, turning the chaotic fat-tree cable nest into tidy waves. AWS first deployed RNG in a Dublin data center in 2024, then Germany and Spain, and says most newly built data centers now get it. Vice president Matt Rehder claims 69 percent fewer routers and switches, 33 percent higher throughput, 40 percent lower network power and 27 percent lower operating costs.

Jellyfish co-author and UIUC professor Brighten Godfrey, uninvolved in the work, calls real-world use of random network graphs 'remarkable'. Notably, AWS is not pitching RNG for AI training clusters — coordinated training traffic does not approximate a random graph; this is about everyday cloud architecture. The design is detailed in the April 2026 paper 'RNG: Flat Datacenter Networks at Scale'.

Replace the fat-tree hierarchy with a quasi-random graph — a flat pool of paths — and build a ShuffleBox device to tame the random cabling that sank earlier attempts.

Performance figures are AWS claims, but deployment is verifiable: first Dublin 2024, then Germany and Spain, now default in new builds.

An independent expert from the original random-graph research calls production use 'remarkable'.

ShuffleBox reframes the failure point as a hardware problem, which is why 2012's idea finally scaled.

The numbers target the industry's real cost centre — switches, fiber and the power they burn.

The idea had waited 14 years (Jellyfish, 2012) for hardware that absorbs the operational mess. When a clever architecture fails in practice, automate the part humans cannot maintain.

As of May 2026, RNG runs in production across AWS data centers in Ireland, Germany and Spain and is being fitted into most newly built data centers; the team published the design in April 2026 and says no one else has done this at scale. AWS does not yet position RNG for AI training clusters.