Amazon Replaces Fat-Tree With Flat Random-Graph Networks, Cuts Costs 45%
Amazon published a paper this week revealing that flat quasi-random graph networks have replaced fat-tree topology as the default for most datacenter workloads — at forty-five percent lower switch count. Plus: NetBox Replica Cache solves the automation read wall, Microsoft completes the agentic AI safety engineering stack, and Equal1 fits a working quantum computer in a server rack.
Welcome to Amaze Networks for Friday, May twenty-second.
Before you get into the rundown — I've been sitting on a paper since Monday and I think it deserves the lead today. Amazon quietly replaced fat-tree topology across most of their datacenter workloads, published the academic paper this week, and the headline number is a forty-five percent reduction in switch count.
Forty-five percent is a big number. And fat-tree has been the unquestioned answer to "how do you build a datacenter network" for so long that most engineers treat it as a law of physics rather than a design choice. What did Amazon actually build?
New episodes, every weekday.
Amaze Networks drops at 4 AM CT, Monday through Friday. Spotify and Apple Podcasts submissions in progress.