← Back to feed

Amazon has deployed flat random graph networks across production AWS data centers and made them the default for most workloads — a shift away from hierarchical fat-tree topologies that have dominated data center design for three decades. The underlying paper, RNG: Flat Datacenter Networks at Scale, appeared on arxiv in April 2026 (arXiv:2604.15261) and has since been revised twice as the production deployment matured.

The core problem RNG solves is structural. Fat-tree networks arrange switches in fixed tiers — access, aggregation, core — which creates choke points at upper layers while lower-layer links sit idle. The hierarchy simplifies operations and routing but systematically concentrates traffic, especially the synchronized all-to-all communication patterns that dominate large-scale AI training jobs.

How RNG Works

RNG replaces the hierarchical topology with a quasi-random graph where routers connect without a fixed layer structure. Instead of a few privileged upward paths through aggregation tiers, the random topology creates many independent edge-disjoint paths between any two endpoints.

Three components make this practical at AWS scale:

Spraypoint routing. A distributed protocol that exploits random graph properties to find large numbers of edge-disjoint paths between endpoint pairs. Rather than chasing the shortest path, it fans traffic outward through distributed waypoints and guides it back — distributing load across many simultaneous paths instead of concentrating it on the hierarchical core.

ShuffleBox. A passive optical device that internally shuffles fiber connections, making the random-looking cabling topology physically manageable. The authors describe cabling complexity as “similar to that of fat trees” despite the underlying topology being fundamentally different — solving the primary deployment objection that has blocked flat networks in production since the theory was established.

The design was tested in two real Amazon production fabrics against fat-tree baselines across transport and storage workloads. RNG matched fat-tree application performance on all tested workload patterns while finding significantly more independent paths than standard routing approaches.

Why It Matters for AI Infrastructure

The cost and fault-tolerance advantages of random graph topologies have been known since the early 2010s. The deployment gap was always practical: random wiring is hard to build, maintain, and route without bespoke hardware and protocols. AWS solved the hardware problem with ShuffleBox and the routing problem with Spraypoint, both developed internally.

For AI training workloads specifically, the benefits compound. Large GPU clusters generate synchronized traffic bursts — collective operations, gradient exchanges, checkpointing — that are poorly matched to fat-tree topologies where simultaneous upper-layer demand creates hot links. A network that spreads load across many more independent paths keeps expensive GPUs from waiting on network congestion.

At the cost side: the paper reports RNG is up to 45% cheaper than fat trees. With hyperscaler capex running at $130B+ per quarter across the major cloud providers, topology-level savings at that magnitude are material — not marginal.

Scale of Deployment

The paper states RNG is now the default data center network for most AWS workloads, after quiet deployment across production fabrics beginning in 2025. Amazon disclosed this through the arxiv paper rather than a product announcement. It is the first production deployment of flat datacenter networks at hyperscaler scale.

The question of whether the other hyperscalers — Google, Microsoft, Meta — have been running comparable research will likely become clear as AI infrastructure capex disclosures force more transparency about where the efficiency gains are actually coming from.