High-bandwidth flash (HBF) technology developed by SanDisk in the United States is drawing renewed attention in the market. SanDisk recently secured a patent for a technology that directly attaches NAND flash beneath a graphics processing unit (GPU) or artificial intelligence (AI) accelerator. As the technology gains attention as an alternative to reduce data bottlenecks in the AI Semiconductor market, analysts say it has also helped lift the share price.
On the 22nd (local time), SanDisk shares ended the session at $2,273.73 (about 3.5 million won), up 4.07% from the prior transaction day. Intraday, they rose to $2,352.99, hitting a 52-week high. With AI spreading, demand has surged for SanDisk's core products—data center-focused NAND and enterprise solid-state drives (eSSD)—and expectations for commercialization of next-generation technologies like HBF are cited as a driver pushing the stock higher.
◇ Stock rises on patent for "NAND under the GPU"
According to the U.S. Patent and Trademark Office (USPTO), SanDisk registered a patent (US 12,430,274 B2) in September last year for a "processing core integrating high-capacity, high-bandwidth storage memory." The core of the patent is a structure that places a processor such as a GPU or AI accelerator directly on top of high-capacity nonvolatile memory, with high-bandwidth memory (HBM) stacks arranged around it.
Market research firm TrendForce on the 22nd said, "SanDisk is accelerating development of HBF, a next-generation structure that vertically stacks NAND, while also advancing an additional memory concept to address structural capacity limits."
An industry source said, "NAND has slower data access than DRAM, but placing it directly under a compute chip shortens data travel distance and can ease bottlenecks in AI Semiconductor," adding, "SanDisk's patented technology shows a future hardware architecture that can break through storage bottlenecks."
◇ AI bottleneck shifts from securing GPUs to memory capacity
Until now, competition in AI Semiconductor was close to a race to secure as many GPUs and as much HBM as possible. But the market now judges that this alone is not enough.
The AI market has shifted its center of gravity from "training" large models to "inference," which delivers answers to actual users. In inference, countless users ask questions at once, AI remembers prior conversations and context, and processes images, video, and audio together. In this process, memory capacity and data movement are emerging as problems. The key-value (KV) cache that large language models (LLMs) use to quickly reference prior context is also growing. As AI becomes personalized and evolves into agent AI, this limitation becomes more pronounced.
HBM is memory that widens data pathways by vertically stacking volatile DRAM. It has become a key component in the AI era by quickly supplying the data needed when GPUs perform large-scale parallel computation. But while HBM is fast, it is expensive, and because it is based on volatile memory that loses data when power is cut, its capacity cannot be increased indefinitely.
SSDs are based on NAND, offering large capacity at a relatively lower price. They are also nonvolatile, meaning data does not disappear when power is off. However, they are physically far from AI chips, and their data transfer speeds are lower than HBM.
HBF is a technology that aims to fill the gap between the two. It is slower than HBM but faster than SSDs, and seeks to create a new memory tier that provides far larger capacity than HBM. As it becomes important to store more data closer and fetch it faster, analysis that SanDisk's patented technology could be an alternative has drawn market attention.
◇ "A technology that will change AI chip package architecture"
So far, AI chips have been designed by attaching HBM next to the GPU and leaving larger data to SSDs farther away. In contrast, SanDisk's patent pulls high-capacity NAND right under the compute chip. If HBM is a "reference book" spread out on the desk, NAND-based HBF is the "bookshelf" placed near the GPU.
Industry observers interpret the patent as showing that HBF can go beyond being viewed merely as a "new memory" placed between HBM and SSD, and could change how AI chips are built. In the patent, SanDisk proposes placing a compute chip such as a GPU or AI accelerator on top of a structure that integrates NAND storage and its control circuitry. This combined structure is attached to an interposer, the substrate that multiple semiconductor chips are mounted on, with HBM arranged around it.
Simply put, the idea is to attach high-capacity NAND directly under the GPU to improve efficiency. Placing NAND storage close to the compute chip can shorten data travel distance. HBM handles data that must be processed immediately, while the NAND attached under the GPU stores larger data and takes on repetitive reads and writes.
In this structure, the compute chip can be a large GPU or an AI processor, with HBM stacks arranged around it. What SanDisk is aiming for is to elevate NAND from a simple storage device to a high-capacity memory that operates alongside HBM within an AI Semiconductor.
The structure SanDisk proposed is closer to dividing roles with HBM than replacing it. HBM handles immediate computation and ultra-low-latency tasks, while NAND-based HBF handles large read-write operations and data that must be retained for a long time.
If implemented, this architecture would allow AI data centers to process longer context, more user records, and larger multimodal data without indefinitely scaling expensive HBM. It can lower inference expense and power consumption, making it attractive technology for AI service corporations.
However, HBF is not suitable for every AI task. NAND has longer latency than DRAM and also faces write endurance issues. Therefore, HBF is more likely to be used first for inference workloads that repeatedly read large-scale data, maintain long context, and reference massive KV caches, rather than computations where latency is extremely critical.
◇ SanDisk standardizes with SK hynix… signs that Samsung Electronics may join
Industry watchers interpret SanDisk's focus on HBF technology as an attempt to pull its strength—NAND—inside the AI Semiconductor package to find a new breakthrough in the HBM-centric race. Compared with DRAM, NAND has drawn relatively less attention in the AI market.
An industry source said, "HBM is a market led by Samsung Electronics, SK hynix, and Micron, and is enjoying a boom as AI demand rises," adding, "Rather than directly chasing this competition, SanDisk chose a strategy of bringing its strength—ultra-high-capacity NAND—inside the AI Semiconductor."
SanDisk has said its goal is to develop HBF that provides bandwidth close to HBM. The company aims to reshape the market with products that, at a similar expense, offer 8 to 16 times the capacity of HBM while delivering higher bandwidth than existing SSDs.
SanDisk has already formalized HBF standardization work with SK hynix. In February, the two companies held a kickoff event for an HBF specification standardization consortium at SanDisk's headquarters in Milpitas, California.
SK hynix is contributing its stacking, packaging, and interposer expertise accumulated in HBM, while SanDisk brings NAND and flash design capabilities in a collaborative setup. SK hynix has established itself as a powerhouse in AI memory by securing Nvidia's supply chain in the HBM market. If it builds HBF standards together with SanDisk, it can shape the AI memory market around both HBM and HBF.
Samsung Electronics has not officially outlined concrete plans for an HBF business. However, industry and securities circles expect the company is more likely to enter the HBF market aggressively, given its experience of losing leadership once in HBM.
A source familiar with Samsung Electronics' semiconductor business said, "Although it has not been announced externally, Samsung Electronics has continued HBF-related research since the early 2020s," adding, "It has secured multiple technologies close to commercialization, and some have already been filed for or registered as patents."
◇ Heat, yield, and software hurdles before commercialization
Although SanDisk's patented technology has been spotlighted in the market, it is not yet commercialized. Industry observers say that to realize the structure SanDisk proposed in actual products, issues such as power consumption, manufacturing expense, packaging Production yield, and thermal management must be addressed.
A researcher at a semiconductor corporations said, "If you place a processor, NAND structure, and HBM stack in one package, thermal control will be extremely difficult," adding, "GPUs or AI accelerators already consume high power, and attaching NAND and HBM stacks on top of that raises internal package heat density and complicates power delivery network design, making implementation challenging."
Production yield is also cited as a concern. HBF is a technology that seeks to use NAND in a high-bandwidth structure similar to HBM. Three-dimensional (D) NAND with hundreds of stacked layers is already widely used. When you add hybrid bonding, interposer wiring, and controller architecture, process complexity increases.
Controllers and software also matter. HBF is not a technology that works simply by placing a large NAND close to the GPU. The system must precisely decide what data to place in HBM, what to place in HBF, when to prefetch, and in what order to feed the GPU. If AI models and system software cannot fully leverage HBF, merely attaching physically close NAND will make it hard to achieve the expected performance.
Shinyoung Securities projected the HBF market could form in 2027 and grow to $12 billion (about 18.4632 trillion won) by 2030. The analysis says HBF is more likely to grow gradually in line with the pace of AI inference expansion and standardization, rather than replacing the HBM market in the short term.