Lexar is pushing SSDs to shoulder AI workloads, letting consumers run large language models without buying pricey memory.

The company’s AI Storage Core SSD slashes DRAM needs by up to 40 %. Internal tests show the Qwen 3.5 122B model runs on 32 GB RAM, generating 15.6 tokens per second versus 5.2 with standard setups.

Running the 122B model cost a typical build around $4,500 for a 128‑GB RAM rig. Lexar’s solution cuts that to 32 GB, keeping performance high while cutting hardware expense.

Even models as massive as 400 billion parameters can be offloaded, though token rates drop and latency climbs to seconds per response. Still, for many users, the trade‑off between cost and speed favors Lexar’s SSD‑based approach.