SK hynix and SanDisk Unveil First High Bandwidth Flash Standard Targeting AI Memory Bottleneck

SK hynix and SanDisk have published the world's first open specification for High Bandwidth Flash (HBF), a new category of AI memory that sits between high-speed HBM and low-cost SSDs. Unveiled at the Flash Memory Summit (FMS) 2026 in Santa Clara, the standard targets one of the most stubborn bottlenecks in large-scale AI inference: feeding enough data to accelerators without the prohibitive cost of stacking every chip with HBM. The release, backed by Google and Tenstorrent, promptly sent memory stocks higher as analysts revised price targets upward. For the latest AI industry coverage, the move signals how memory design is being re-engineered specifically for inference workloads.

A new memory tier for AI inference

For years, the AI memory conversation has been dominated by High Bandwidth Memory (HBM), the stacked DRAM that sits close to GPUs and delivers enormous throughput. HBM works for training, but it is expensive and capacity-constrained. At the other end, NAND flash-based SSDs are cheap and capacious but too slow to keep inference accelerators fed. The HBF specification, published as an open technical document through the Open Compute Project (OCP), is designed to fill that gap.

According to reporting by Tom's Hardware, The Korea Herald, and HPCwire, the specification defines stacked NAND packages of up to 16 layers (16-Hi) and targets aggregate bandwidth of up to 3 terabytes per second. The interface uses UCIe, the Universal Chiplet Interconnect Express standard, allowing HBF modules to be integrated alongside logic dies in advanced packaging. TrendForce described the goal as resolving AI inference bottlenecks by delivering large-capacity, high-bandwidth storage that is far cheaper than equivalent HBM capacity.

Google and Tenstorrent back the ecosystem

A standard only matters if the buyers adopt it, and SK hynix and SanDisk lined up heavyweight supporters for the launch. Google and AI chip startup Tenstorrent were named as early ecosystem participants, lending credibility to the proposition that hyperscalers and accelerator designers see HBF as a viable tier rather than a niche experiment. By publishing the specification openly through OCP, the companies are explicitly inviting rival memory makers and chip designers to build compatible products, a strategy intended to grow the addressable market the way open HBM standards did in previous generations.

The positioning is deliberate. SK hynix remains the dominant HBM supplier, but the company has argued that the future of memory lies somewhere between HBM and SSDs. HBF gives it a product that defends its memory leadership into the inference era while opening a path for customers who need capacity that HBM economics cannot support.

SK hynix also debuts 375-layer NAND

The HBF announcement was not the only headline from SK hynix at FMS 2026. The company also unveiled 375-layer NAND flash technology, which it said delivers roughly 2.5 times the performance per watt compared to its previous generation. The higher layer count increases storage density per die, which directly supports the capacity argument behind HBF: more bits per package at lower energy cost makes large stacked modules economically viable.

Together, the two announcements sketch a roadmap in which SK hynix is pushing on both the density and the bandwidth axes simultaneously. That matters because the AI inference market, projected by most analysts to grow faster than training as models are deployed in production, will reward whoever can supply vast amounts of fast, affordable memory.

Memory stocks jump on the news

Markets responded forcefully. According to 24/7 Wall St. and Yahoo Finance, Sandisk shares jumped around 8 percent, Micron gained roughly 6 percent, and SK hynix climbed about 4 to 5 percent as Wall Street analysts hiked their price targets on the back of the AI memory boom. The rally reflected an investor read that an open, multi-vendor HBF standard expands the total memory market rather than simply redistributing existing share.

The reaction also underscored how sensitive the memory sector has become to AI-specific product announcements. Where NAND and DRAM once traded on broad PC and smartphone demand cycles, the marginal buyer is now the data center building inference infrastructure. Any credible plan to lower the cost per terabyte of feeding accelerators is treated as a structural positive.

Why inference memory is the next battleground

Training a frontier model is a one-time, intensive compute task that justifies the cost of HBM. Running that model millions of times a day for users is a recurring cost, and it is dominated by memory. If every query requires shipping gigabytes of model weights and key-value cache through the system, the economics of inference hinge almost entirely on memory bandwidth per dollar.

That is why so many companies are now competing on memory rather than pure compute. Samsung has pushed advanced HBM packaging, Micron is investing heavily in HBM3E, and now SK hynix and SanDisk are carving out HBF as a distinct inference tier. The bet is that inference workloads will tolerate slightly higher latency than HBM in exchange for dramatically more capacity at a fraction of the cost.

Not everyone is convinced the tier will displace HBM. Some architects argue that the latency gap between flash and DRAM remains large enough that HBF will serve a narrow role, augmenting rather than replacing HBM for the hottest parameters. SK hynix and SanDisk counter that the open standard and the 375-layer density improvements make the economics compelling for cold and warm data in long-context models.

Looking ahead

The publication of the first HBF specification is a milestone, but it is a starting point, not a finished market. Products conforming to the standard are expected to sample over the coming quarters, and adoption will depend on whether hyperscalers design inference platforms around the new tier. With Google already signaling support, the odds of at least limited production deployment look stronger than for most new memory categories.

For an industry that has spent the last two years obsessed with GPU supply, the focus is shifting to the memory that keeps those GPUs busy. SK hynix and SanDisk have just made the case that flash, reinvented and stacked, can be part of that answer.

Stay Ahead of AI

The memory layer of the AI stack is moving fast, and the companies that win the inference era may not be the ones with the fastest chips but the ones with the smartest memory. Read more AI news to keep up with every shift in hardware, models, and policy.