PanSwitch™ is the industry's first switch that connects system devices across multiple racks in any topology, without a network in between. As AI outgrows single accelerators, performance depends not only on each chip but on how efficiently devices work together. PanSwitch™ sits between memory, accelerators and CPUs and removes the overhead between them: every data path is accelerated in hardware, minimizing software involvement including the network stack, and every device joins one cache-coherence domain sharing a unified memory space. Devices in different racks exchange data as memory accesses, not network transfers. It also lets users compose a cluster to the workload: resources are disaggregated by type and combined selectively, lowering CAPEX. One silicon adjusts lanes per port, works as a PCIe or CXL interconnect switch, and links switches in topologies beyond the tree structures of conventional scale-up interconnects. With PanSwitch™, customers accelerate large-scale AI on clusters they compose themselves.