At SC26, SK hynix moved KV cache into a shared CXL pool for multiple servers to reuse, and used in-memory DRAM computing to reduce data movement. So far, the only source for these demonstrations is the company itself.

What it wants to sell is no longer just HBM, but an entire rack redesigned around memory. If MARS goes a long time without customer deployments and reproducible real-world tests, this assessment no longer holds.