SK Hynix Bets on High Bandwidth Flash, Building AI Memory Tiering Architecture
nashnova research
SK Hynix is repositioning NAND flash from traditional storage to AI inference infrastructure, building a new memory tier around High-Bandwidth Flash (HBF) — a signal that AI chips' memory systems are shifting from HBM-only to a multi-layer division of labor.
Why can't AI chips rely on HBM alone anymore?
AI agents handle increasingly complex conversations and tasks. The data that needs caching is ballooning — but HBM is small and expensive, making it impractical to store everything there.
This means → chip memory needs to work in tiers, like the human brain: the most-accessed data stays in the fastest layer (HBM), while bulkier, less-accessed data moves to a cheaper tier.
SK Hynix's answer is HBF — high-bandwidth flash. In plain terms = it is a new "middle floor" inserted between HBM and conventional SSDs.
What exactly is HBF, and how does it perform?
HBF combines NAND flash's large capacity and low cost with bandwidth higher than a conventional SSD, manufactured using the same through-silicon via (TSV) stacking — vertical signal channels punched through the chip — that HBM uses.
Chief Development Officer Hyun Ahn positioned it at SEDEX 2026 in Seoul: capacity over ten times that of comparable HBM; bandwidth roughly half of HBM's.
This means → HBF does not chase peak speed. It trades ten-times-more capacity for an acceptable speed step-down — for large volumes of "warm data," the cost-performance ratio far exceeds simply stacking more HBM.
How far along is the spec, and when will it ship?
In August 2026, SK Hynix and Sandisk jointly published the first HBF specification through the Open Compute Project (OCP): 8-layer and 16-layer stacking, up to 512 GB capacity, bandwidth in three tiers from roughly 0.4 TB/s to 3.0 TB/s, using the UCIe open interconnect standard.
Google and Tenstorrent have joined the standardization effort. The spec is targeted for finalization by January 2027. Sandisk has taped out the first HBF memory die and aims to deliver initial inference-product samples in 2027.
SK Hynix itself has not announced a mass-production date or disclosed commercial customer orders. In plain terms = the roadmap is drawn, but there is a clear gap before real shipments begin.
How does data move between tiers?
Hardware tiers alone are not enough. Software must decide which data sits where. SK Hynix is developing two technologies for this:
SALT-KV assigns KV-cache data — the context a large model stores temporarily during inference — to the right tier based on reuse value and storage cost. This means → it acts as an intelligent dispatcher, keeping "likely to be reused" data on the fast tier and moving the rest to cheaper storage.
CMM-Hybrid uses the CXL standard — a high-speed interconnect protocol linking CPUs with various memory devices — to store the full KV cache on SSD while predicting hot data and pre-loading it into DRAM. The company demonstrated this at the OCP Global Summit in October 2026.
What is making money right now? Where does NAND stand?
HBF remains a mid-to-long-term play. The near-term revenue pillar is enterprise SSDs, built on SK Hynix's 321-layer TLC NAND process.
Q3 NAND bit shipments dipped slightly quarter-on-quarter — but this was deliberate: the company shifted toward higher-value products, lifting average selling prices by roughly the mid-20% range QoQ.
Q4 NAND bit shipments are guided to grow by mid-single-digit percentages QoQ. The 321-layer process is targeted to exceed 50% of domestic NAND capacity. This reflects a "price-over-volume" pivot, concentrating capacity on the higher-margin enterprise market.
What else is SK Hynix doing around this strategy?
In September 2026 the company launched SK Hynix Ventures, focusing on AI compute, data centers, system software, and optical interconnects.
It had already invested in several frontier companies: Avicena (micro-LED optical interconnects), Eliyan (chiplet interconnects), and Etched (HBM-SRAM fusion architecture).
A multi-year agreement with Nvidia targets joint development of next-generation AI memory. This means → SK Hynix is not just selling memory chips — it is trying to embed itself into the core supply chain of AI compute infrastructure. Whether HBF delivers samples on schedule in 2027 will be the key proof point for this entire tiered-memory strategy.
市场有风险，内容仅供研究参考，不构成投资建议。