InfiniBand vs Ethernet: Wiring the AI Supercomputer
The network fabric of a training cluster isn't plumbing — it's the critical path. Choose wrong, and thousands of GPUs spend more time waiting than working.

Ten thousand GPUs are only as fast as the wires between them. InfiniBand, Ethernet and the fabric of the AI supercomputer.
This section is depth 03 — hall of the descent, expanded · 3 articles
The network fabric of a training cluster isn't plumbing — it's the critical path. Choose wrong, and thousands of GPUs spend more time waiting than working.

Selecting the right switching fabric is one of the most consequential calls in any data-center build. Get it right and the network keeps pace with your GPU clusters. Get it wrong and a billion-dollar training run waits on a bottleneck measured in microseconds.

Training gets the campuses. Inference needs the meet-me room — what carrier neutrality actually buys when models graduate to production.
