Marvell Expands AI Memory Infrastructure With CXL, Storage, and Optical Sharing

Marvell introduced technologies at FMS 2026 that address memory pressure in large AI inference deployments. The portfolio covers SSD-side infrastructure, CXL-based expansion at the rack level, and optical links for shared memory across larger systems. The focus is the growth of key-value cache requirements as models use longer context windows and produce more memory traffic. For the memory market, Marvell's announcement is less about selling DRAM or NAND directly and more about enabling architectures that can consume those resources more efficiently. If hyperscalers adopt more disaggregated designs, demand could shift toward higher-capacity SSDs, CXL memory modules, and networking components that keep accelerators supplied with data.