$Amazon(AMZN.US) NVIDIA's GB300 houses 72 GPUs in a single cabinet with full all-to-all connectivity.

Any two chips are reachable in one hop, with traffic traversing its in-house NVSwitch only once.

By contrast, AWS Trn3 is constrained by the limited port count on its off-the-shelf switch chips.

A single NVSwitch provides 72 ports, enabling direct links to all 72 GPUs.

In the Trn3 switch tray, the Scorpio X exposes only 18 usable ports, so it can mate with just one Trn3 per compute tray.

This means only one Trn3 in each compute tray can be directly connected.

Each compute tray integrates four Trn3s, which must be split into four isolated 'vertical ladder' planes.

Within a plane, chips communicate in a single hop.

Cross-plane traffic must traverse eight Scorpio P switch chips inside the compute tray.

This is the key difference: GB300's compute tray carries no switches, and Trn3 today does not achieve full intra-domain all-to-all.