SK hynix and Sandisk to Standardize New AI Memory

- HBF is a non-volatile memory solution based on NAND flash technology, meaning it retains data without power and doesn't require the constant power refresh that DRAM-based HBM does. This characteristic is advantageous for AI inference workloads where large model weights are read frequently but written infrequently. - The technology fits into a new "context memory" or "warm tier" in the memory hierarchy, a concept highlighted by NVIDIA's CEO Jensen Huang. This tier is designed to hold large volumes of AI context—like retrieval databases, conversation histories, or embeddings for Retrieval-Augmented Generation (RAG) in claims processing—that are too large for expensive HBM but need faster access than SSDs can provide. - For agentic AI and multi-agent systems, this new memory layer directly supports more complex memory architectures. System designers can map an agent's "working memory" to HBM while using the larger capacity of HBF for "session" or "episodic memory," allowing agents to maintain context over longer interactions without overwhelming the fastest memory tier. - In insurtech applications, such as automated claims processing or underwriting, HBF can accelerate workflows by allowing AI models to access vast, unstructured datasets (e.g., images, reports, and policy documents) more quickly. This avoids the latency of retrieving data from slower NVMe SSDs, speeding up fraud detection and policy analysis. - From a backend system design perspective, the introduction of HBF necessitates designing APIs and orchestration layers that are aware of this heterogeneous memory system. A Staff-level engineer would need to architect data placement strategies, deciding which data structures (e.g., active KV cache vs. foundational model weights) reside in HBM versus HBF to optimize both latency and cost. - The creation of HBF reflects a broader venture capital trend of shifting investment back into "deep tech" and AI hardware, including memory and storage. For a technical founder, this signals a market opportunity but also highlights the challenges of longer development cycles and higher capital requirements inherent to semiconductor startups. - The collaboration is significant because SK hynix is the market leader in HBM, giving the HBF standard instant credibility and a direct path into the GPU ecosystem. The goal is to create a multi-supplier market to assure customers of stable supply and competitive development, a key lesson for founders in the hardware space. - The roadmap for this technology is aggressive, with Sandisk targeting first samples in the second half of 2026 and expecting AI inference devices equipped with HBF to be available for sampling in early 2027.