Within each AI server, GPUs are linked by ultra-high-bandwidth, short-reach interconnects such as NVLink. Instead of sending data over the standard connection that normally links a GPU to the rest of the server, NVLink creates a direct ‘fast lane’ between GPUs so that they can exchange model data without going through the CPU, as it is the case for the standard connection. 61 This high-speed fabric effectively pools the memory and compute of multiple GPUs inside a rack but also raises local power density, adding extra heat that needs to be removed at rack level