AI Infrastructure

There is a network inside a GPU server that moves more data than your entire fabric, and you will never configure it. Where NVLink ends is where your job begins: the scale-up versus scale-out boundary, and how GB200 NVL72 moved it.

At a few thousand GPUs, the backend network becomes a probability problem: something is always broken and the hash is always unlucky. The fix is two identical fabrics with zero cross-links. A deep dive into dual-plane design, sized for B300 nodes with 800G per GPU.

Your AI cluster is two networks wearing one badge. One is judged by whether users notice it. The other is judged by whether GPUs wait on it. Frontend vs backend: the traffic, the rules, and how to design the GPU fabric so compute never waits.

5,412 connections on my Kubernetes cluster versus one RDMA flow moving 59.5 GB a minute. Then three identical transfers, one degraded path, and 32 percent of the workers' time gone. AI traffic measured, not described.

Longhorn on the workers (the storage leaves finally earn their name), 60Gi of language models surviving node death, a dozen apps behind one login, and the five-object checklist that publishes anything.

Every chip in the modern data center exists because one bottleneck got expensive enough to deserve its own silicon. The whole zoo in one sitting: real photos, one table, and a memory trick that makes the differences stick.