The NVLink and NVSwitch scale-up fabric that links accelerators inside an AI server rack so they behave as one system. It sits between packaged accelerator modules and full rack integration, and depends on leading-edge logic wafers for the switch silicon.
A sustained tight read means the switch fabric that binds accelerators into a single training system, not the accelerators themselves, limits how large a cluster can act as one machine. Chips may be available, but rack-scale bandwidth cannot expand until more switch silicon is fabricated and integrated. Because the switch silicon uses leading-edge logic wafers, relief arrives only as added fabrication and rack integration qualify and ramp, over successive quarters rather than a single order cycle.