Lexar pitches SSD offloading as a way to cut DRAM needs for local AI PCs

Lexar outlined an SSD-led approach for running large language models on local PCs with less DRAM. The company says its AI Storage Core SSD and software stack can lower memory requirements by at least 40%, and it presented internal tests showing a Qwen 3.5 122B model running on a machine with 32 GB of DRAM through SSD offloading. Lexar also claimed higher throughput for a 35B model versus conventional frameworks and said a traditional 32 GB DRAM setup failed to load the larger model at all. For RamTrend, the key point is not an immediate change in memory demand, but a possible architectural shift: if AI PCs can use NAND more aggressively to support model execution, some workloads that currently require large DRAM footprints could be partly redirected toward SSD capacity. That would matter most in cost-sensitive edge and consumer AI systems, though the claims are based on vendor-provided testing and still need broader validation.