At the opening of the Future of Memory and Storage (FMS) 2026 conference in California, held on August 4, SK Hynix, together with SanDisk, presented the first specifications of the High Bandwidth Flash (HBF) technology. This development claims to be a key intermediate link between ultra-fast HBM memory and traditional SSD drives, which could radically change the architecture of modern AI servers.

Technical specifications of HBF

The new specification impresses with its parameters:

  • Capacity: up to 512 GB in configurations with 8 and 16 NAND dies per stack;
  • Bandwidth: from ~0.4 TB/s to an impressive 3.0 TB/s;
  • Interface: the open UCIe (Universal Chiplet Interconnect Express) standard, ensuring compatibility with GPUs and CPUs from various manufacturers.

It is important to note that the specifications have been submitted to the Open Compute Project — a community focused on data center technologies. This positions HBF as an industry-wide standard rather than a proprietary solution, which will significantly accelerate its adoption.

Four-tier memory hierarchy

With the advent of HBF, the memory architecture of AI servers becomes more complex and efficient:

  1. HBM — handles "hot" data directly in the compute core, providing maximum speed but with limited capacity and high cost.
  2. CXL — accelerates classic RAM with latencies in the tens to hundreds of nanoseconds range, maintaining full compatibility with processors.
  3. HBF — fills the critical gap between RAM and SSDs. Thanks to integration via the UCIe bus directly with the processor, these devices provide speed close to HBM but with large capacity based on inexpensive NAND architecture.
  4. NVMe SSD — remain for storing "cold" data.

Solving the key problem of LLMs

In the infrastructure of large language models, the main problem remains storing the vast context and model weights during inference. Here, HBF takes on the storage of gigabyte-scale context caches and AI model weights that do not fit into HBM and CXL due to their high cost and capacity limitations. This allows for significantly expanding data processing capabilities without the need to scale the most expensive types of memory.

The partnership between SK Hynix and SanDisk on HBF standardization began back in August 2025, and in February 2026, the consortium, which also included Google and Tenstorrent, officially launched the global standardization process. The exhibition also announced the development of 375-layer 4D NAND (V10) with performance increased by 2.5 times per watt of energy compared to the previous generation. Mass production of eSSDs based on this technology is planned for early next year.

It is worth noting that SK Hynix shares have risen by more than 330% since the beginning of 2026, although they have lost about half of their value since the June peak. This reflects market volatility, but the company's long-term potential associated with the development of HBF and new types of memory remains significant.

My analysis: HBF is not just another memory format, but a strategic step toward optimizing the cost of AI computing. The ability to use inexpensive NAND architecture at speeds close to HBM could significantly lower the entry barrier for large-scale AI projects and accelerate their commercialization. The open standard through OCP is a smart move that will avoid ecosystem fragmentation and ensure broad compatibility.