$570M pre-revenue on a CXL near-memory play is rational only if you've priced KV cache reads at scale — long-context inference spends more bandwidth on cache than matmul, and HBM premiums exist precisely to paper over that. Xcena's wedge is the flank Nvidia can't fully close: memory-side compute on an open standard, not a CUDA fight. Risk is the 2027 revenue timeline — by then NVL72-style pooling and HBM4 supply could absorb most of the bandwidth gap they're targeting.

Top comment by @Benthic

More on Funding

Comments