SK hynix, Sandisk Target AI Memory Wall with New HBF Standard

MILPITAS, CA – February 26, 2026 – In a strategic move to define the next generation of AI infrastructure, memory titans SK hynix and Sandisk have announced a joint effort to standardize a new class of memory called High Bandwidth Flash (HBF). The collaboration, formalized through a dedicated workstream under the influential Open Compute Project (OCP), aims to solve a critical bottleneck threatening to slow the mass adoption of artificial intelligence.

At a kick-off event held at Sandisk's headquarters, the companies detailed their plan to create an industry standard for HBF, a novel memory solution designed to bridge the widening chasm between ultra-fast, but limited-capacity High Bandwidth Memory (HBM) and high-capacity, but slower Solid State Drives (SSDs). This new memory layer is engineered specifically for the demands of AI inference—the process of running trained AI models to deliver real-time services to users.

The AI Inference Memory Dilemma

The AI industry is undergoing a fundamental shift. After years focused on the monumental task of training Large Language Models (LLMs), the focus is now pivoting to inference, where these models are deployed at scale. As millions of users interact with AI services simultaneously, the demand on data center hardware is exploding, creating what engineers call a "memory wall."

Currently, AI accelerators like GPUs rely on HBM, a type of vertically stacked DRAM that provides the immense bandwidth needed to feed data to powerful processing cores. However, HBM is expensive to produce and constrained in capacity. A single AI server may require terabytes of memory to hold the parameters of massive models, a scale that is economically and technically challenging to achieve with HBM alone. On the other end of the spectrum, NAND flash-based SSDs offer vast, affordable capacity but lack the speed to keep the AI processors fed, causing them to sit idle and waste expensive computing cycles.

This gap creates a significant performance and cost dilemma for data center operators. HBF technology is designed to resolve this by creating a new, intermediate tier in the memory hierarchy that offers a compelling balance of capacity, bandwidth, power efficiency, and cost, specifically optimized for the read-intensive workloads typical of AI inference.

Bridging the Gap with High Bandwidth Flash

High Bandwidth Flash is not simply a faster SSD. It represents a fundamental architectural innovation that combines the cost-effectiveness and density of NAND flash with the high-speed interconnection techniques pioneered by HBM. HBF utilizes 3D NAND chips, the same technology found in modern SSDs, but stacks them vertically and connects them using Through-Silicon Vias (TSVs), creating a dense, high-speed package.

This design allows HBF to achieve bandwidth an order of magnitude higher than traditional SSDs. Initial performance targets for first-generation HBF aim for a read bandwidth of 1.6 TB/s with a capacity of 512GB per stack—a capacity 8 to 16 times greater than current HBM solutions. Sandisk's internal simulations demonstrate the potential, showing that for a task like reading the pretrained weights of a 405-billion-parameter LLM, an HBF-based system performed within 2.2% of a hypothetical system with unlimited HBM capacity.

"The key to AI infrastructure is to go beyond the performance competition of individual technologies and to optimize the entire ecosystem," said Ahn Hyun, President and Chief Development Officer at SK hynix, in a statement. "Through HBF technology standardization the company will establish a cooperative system and present an AI-era optimized memory architecture to create new value for customers and partners."

A Strategic Push for an Open Standard

The decision to pursue standardization through the Open Compute Project is a critical component of the strategy. The OCP fosters open-source hardware designs for data centers, aiming to drive innovation, efficiency, and interoperability. By establishing HBF as an open standard, SK hynix and Sandisk hope to accelerate its adoption, build a broad ecosystem of support, and prevent market fragmentation.

This collaborative approach comes as the entire industry grapples with the future of AI memory. Competitors are also exploring solutions. Samsung Electronics has reportedly begun its own HBF conceptual design and is participating in the standardization efforts. Meanwhile, AI chip leader NVIDIA is pursuing heterogeneous memory strategies for its next-generation platforms, considering different memory types for different stages of the inference process, further validating the need for a more diverse memory hierarchy.

By leading the OCP workstream, the partners aim to position HBF as the go-to solution for this emerging memory tier. This move positions them not just as component suppliers, but as architects of the future 'total memory solution' for the AI era, where system-level optimization across the CPU, GPU, and memory stack determines overall competitiveness.

Unlocking Scalability and Reducing Costs

The ultimate promise of HBF extends beyond raw performance to the economics of AI. The technology is expected to significantly reduce the Total Cost of Ownership (TCO) for deploying AI inference at scale. This cost reduction stems from several factors. First, by using NAND flash, HBF has a much lower cost-per-gigabyte than DRAM-based HBM. Second, because NAND is non-volatile, HBF consumes no power to retain data and avoids the constant power draw required for DRAM refresh cycles, contributing to overall data center energy efficiency.

Research from SK hynix has shown that a hybrid memory system combining HBM and HBF alongside a GPU can improve performance-per-watt by as much as 2.69 times compared to an HBM-only configuration. This efficiency is crucial for managing the massive power demands of AI data centers.

With initial HBF samples expected in the second half of 2026 and hardware integration anticipated in 2027, the industry is watching closely. While market demand for complex memory solutions like HBF is forecast to ramp up significantly around 2030, this initiative marks a pivotal first step toward building a more scalable, efficient, and cost-effective foundation for the next wave of artificial intelligence.