Morgan Stanley, in its latest research report, stated that the AI memory shortage is the most persistent structural constraint in the current AI build cycle, a point also emphasized by NVIDIA CEO Jensen Huang, who stressed the need for the industry to adopt new thinking. The report proposes three main solutions: lowering memory specifications (e.g., NVIDIA's Rubin platform HBM reduced from 288GB to 192GB), disaggregating inference workloads (configuring dedicated hardware for different inference stages), and enabling memory pooling and sharing through CXL technology. The firm raised its CXL total addressable market forecast from $4 billion to approximately $6 billion by 2030, primarily driven by new deployment scenarios on the AI server side. Morgan Stanley identified Astera Labs (ALAB) and Marvell (MRVL) as potential beneficiaries in the CXL and scaling directions, and Cerebras (CBRS) as a core beneficiary of the inference workload separation architecture. Concurrently, it maintained an Overweight rating on Micron (MU) and SanDisk (SNDK), believing the memory shortage cycle is far from over.