Training clusters are all-to-all, not client-server. Port counts per rack multiply, optics become the dominant interconnect cost, and thermal headroom becomes the binding constraint.
お客様が運用するネットワークのために
Count ports per GPU tray, not per server — that drives volume
Confirm switch line-card thermal headroom before choosing 800G-class optics
AOC is usually the cheapest option below 5 m inside a rack
Coding matters: fabric interoperability depends on EEPROM profiles
