AI GPUs are bottlenecked mostly by high-bandwidth memory (HBM) chips and CoWoS (packaging tech used to integrate HBM with the GPU die), which are in short supply and aren't found in consumer cards at all