Lexar AI Storage Core SSD Cuts DRAM Needs for Local AI Models by 40%

Lexar is developing an SSD designed to reduce the amount of DRAM needed for running large AI models on local hardware. The company aims to offload memory demand from expensive DRAM to cheaper NAND flash, addressing what it calls the 'RAMpocalypse.'

Custom SPU controller enables DRAM-less offloading

The product is called the AI Storage Core SSD, an M.2 NVMe drive that uses a custom Storage Processing Unit (SPU) DRAM-less controller. A hot-swappable M.2 SSD concept was demonstrated at Computex 2026.

Specifications

• DRAM reduction: At least 40%
• Tokens per second (122B model, 32 GB DRAM): 15.6 tokens/s
• Tokens per second (traditional framework, 32 GB DRAM): 5.2 tokens/s
• Tokens per second (122B model, 64 GB DRAM, 256K context): 19.3 tokens/s
• Time to first token (2K context): 2 seconds

According to Lexar CTO Daniel Guo, DRAM is about six times more expensive to manufacture than NAND flash. The company claims its SSD offloading reduces DRAM requirements by at least 40%. For example, running the Qwen 3.5 122B model typically needs 128 GB of DRAM, but Lexar's solution cuts that to 32 GB.

With 32 GB of DRAM, Lexar's SSD offloading achieves 15.6 tokens per second on the Qwen 3.5 122B model. Traditional frameworks like Llama.cpp crash under the same conditions. With 64 GB of DRAM and a 256K context, the SSD delivers about 19.3 tokens per second. Time to first token is about two seconds for a 2K context and 6–8 seconds for a 4K context.

Lexar notes that offloading models up to 400 billion parameters is technically possible, though performance would be slow. The company has not announced pricing or availability yet.