Scale-up fabric

Scale-up fabric is the set of very fast links that join many chips together into one big, tightly-coupled machine — the wiring inside a pod that lets a rack of accelerators behave as a single coherent compute domain. It matters because it defines both the size and the efficiency of the tightest compute unit: the more chips you can bind at full bandwidth, the larger the effective machine you can train on. The model frames this as tight coupling setting the unit of compute, where bandwidth is capability. Frontier training keeps demanding ever-larger coherent domains, pushing scale-up fabric right up against the physical limits of copper and advanced packaging.

If that bandwidth or reach runs short, pod size is capped — training jobs fragment across smaller islands, and efficiency drops as the machine spends more time waiting than computing. Exposure runs through the firms building these links and the operators that depend on them: accelerator and interconnect designers like NVIDIA and Huawei, signal-conditioning and connector specialists such as Astera Labs, Amphenol and BizLink, and the neoclouds like CoreWeave and Nebius whose economics hinge on how large a coherent domain they can stand up.