Scale-out fabric

Scale-out fabric is the network that connects many separate machines across a data center into one giant cluster. Where scale-up binds chips into a single tight domain, scale-out binds those domains into a whole facility, turning a room full of servers into one usable compute resource. It matters because at cluster scale the network effectively becomes the computer — and the model notes that optics, the light-based links carrying data between racks, governs how far that can stretch. With cluster sizes exploding, scale-out bandwidth and the optical transceivers behind it have become a primary constraint on how big a training run can get.

If the fabric runs short on bandwidth, distributed training and large-scale inference bottleneck: chips sit idle waiting on data moving across the hall. The exposed names are the switch and system builders and the optics suppliers that feed them — Arista Networks, Cisco-class switching, servers from Hewlett Packard Enterprise and Accton, hyperscale operators like Oracle, and optical-module makers such as Eoptolink alongside networking incumbents like Nokia.