DeepSeek's latest published paper reveals an interesting technical direction. The core idea is to separate the long-term memory component of large language models from the model weights and store it in memory hardware, which can significantly reduce VRAM pressure. The n-gram technical solution mentioned in the paper is based on this idea—storing long-term dependency information in external memory rather than relying on internal model parameters.

From a hardware perspective, what does this architectural shift mean? The demand for memory will increase substantially. As large models evolve in this direction, the market demand for DDR5 memory could enter a new growth cycle. Micron, as a mainstream memory supplier, is a long-term beneficiary.

From an investment perspective, the price pressure on DDR5 memory may continue to rise. If this type of technical solution is truly implemented and promoted, doubling this year is not an exaggeration—depending on the actual progress of model training and deployment.