The global semiconductor industry is witnessing a fundamental shift in its architectural priorities. According to Alper Ilkbahar, Chief Technology Officer at SanDisk, the artificial intelligence competition is rapidly evolving into a 'memory-centric' battle. This transition is not merely a technical adjustment but a structural transformation that is forcing the world’s largest tech companies to abandon traditional procurement cycles in favor of long-term supply security.
The demand is driven by the increasing complexity of Large Language Models (LLMs) and the rising importance of 'short-term memory' mechanisms like key-value (KV) caching. As AI systems like ChatGPT and Gemini expand their context windows to remember longer conversations, the memory capacity required to store these intermediate states grows exponentially. Furthermore, the industry-wide adoption of 'Mixture of Experts' (MoE) models—which activate only specific sub-networks to save compute—ironically places a much heavier burden on memory systems to keep all potential 'experts' ready for immediate recall.
This technical reality is rewriting the economics of the memory market, which has historically been notoriously cyclical and prone to 'boom and bust' periods of oversupply. For the first time in his thirty-year career, Ilkbahar notes that customers are moving to lock in multi-year supply agreements to prevent shortages. SanDisk recently secured five such agreements, some lasting up to five years, which are projected to generate at least $42 billion in revenue. This shift suggests that memory chips are being repositioned from volatile commodities to essential, long-term strategic assets.
Innovation is also moving beyond standard DRAM. While High Bandwidth Memory (HBM) currently dominates the high-end training market, SanDisk is positioning High Bandwidth Flash (HBF) as the next major breakthrough for AI inference. Developed in collaboration with SK Hynix, HBF utilizes stacked NAND flash to provide massive capacity at a fraction of the cost of HBM. Samples of HBF wafers are expected by the end of 2024, with full products hitting the market in 2025, potentially democratizing the hardware required for complex AI inference at scale.