At the opening of the FMS 2026 conference in California, held on August 4, SK Hynix and SanDisk presented the first specifications for HBF (High Bandwidth Flash) technology. This is not just another development, but an ambitious attempt to occupy a niche between ultra-fast HBM memory and traditional SSD drives. The industry has long needed such an intermediate layer, and now we have concrete figures and parameters.
Key HBF Characteristics
According to the published data, HBF will offer capacity of up to 512 GB in configurations with 8 or 16 NAND dies per stack. Bandwidth ranges from 0.4 TB/s to an impressive 3.0 TB/s. However, the most important aspect is the use of the open UCIe (Universal Chiplet Interconnect Express) interface. This solution ensures compatibility with GPUs and CPUs from different manufacturers, which is critical for the ecosystem. The openness of the specifications through the Open Compute Project community positions HBF as an industry-wide standard rather than a proprietary solution from a single vendor.
Four-Level Memory Hierarchy in AI Servers
The adoption of HBF will radically change the architecture of modern AI servers. I see the formation of a four-level memory hierarchy:
- HBM — for "hot" data processed directly in the compute core. Maximum speed, but high cost and limited capacity.
- CXL — for accelerating classic RAM with latencies in the tens to hundreds of nanoseconds and full compatibility with CPUs/GPUs.
- HBF — fills the critical gap between RAM and SSDs. Thanks to integration via the UCIe bus directly with the processor, these devices will deliver speeds close to HBM, but with large capacity based on inexpensive NAND architecture.
- NVMe SSD — will remain for storing "cold" data.
For LLM infrastructure, this solves a key problem: storing huge context and model weights during inference. CXL handles data with ultra-minimal latency, while HBF handles gigabyte-scale context caches and weights that do not fit into HBM and CXL due to their high cost and capacity limitations.
Context and Prospects
The SK Hynix and SanDisk partnership on HBF standardization began back in August 2025, and the consortium launched in February 2026 already includes Google and Tenstorrent. At the FMS 2026 booth, the development of 375-layer 4D NAND (V10) was also announced, with performance increased by 2.5 times per watt compared to the previous generation. Mass production of eSSD is scheduled for early next year.
SK Hynix shares have risen more than 330% since the start of 2026, although they have lost about half their value since the June peak — the market is clearly digesting the company's ambitious plans.
My view: HBF is a timely and technologically mature response to the bottleneck in AI infrastructure. If the standard is adopted by a wide range of manufacturers, we will witness a significant reduction in the cost of inference for large models, accelerating their widespread adoption.