I recall hearing (maybe in a Level1Techs video?) that something akin to this can already be done with SAS. Each CPU in a dual-socket system gets its own HBA and each drive has two SAS connectors, so there's no performance hit for accessing a drive from the "wrong" CPU.
It's common that a external SAS chassis had multiple connections that can access all drives. The one I use I believe has 12 connections for 60 drives. CPUs these days generally have a PCIe controller inside.
Socket 0 can use the direct connected HBA 0 to access all drives while socket 1 can use the direct connected HBA 1 to access all drives. On top of that each SAS drive has TWO connections to the SAS fabric.
My favorite storage building block is two 60 disk chassis (120 drives) cross connected to two servers with 4 SAS connections each. So with 12gbit SAS that's 48Gbit per cable x 4 = 192 Gbit/sec or 24GB/sec per chassis. For extra performance I use NVMe inside the server for caching, with hit rates usually in the 75-90% range.
High-end enterprise NVMe SSDs also commonly come in dual-port versions, where instead of a PCIe x4 link the drive has two PCIe x2 links over the same connector.
A Google search for this term returns this discussion and this article. What does it mean? The network controller is connected to multiple PCIe peers?