High Bandwidth Flash could reshape the future of AI memory, but it may also keep SSD prices from falling
High Bandwidth Flash, or HBF, is emerging as one of the most interesting new memory technologies for AI hardware. If it reaches commercial production and performs as expected, it could become a lower-cost alternative to High Bandwidth Memory in certain workloads. However, that same success could create a new problem for consumers: cheaper SSDs may become less likely in the years ahead.
The main reason is cost. Analysts estimate that an 8-Hi HBF stack could cost around $0.64 per GB, compared with roughly $6.78 per GB for a similar 8-Hi HBM4 stack. That means HBF could be about one-tenth the cost of comparable HBM4 capacity, making it extremely attractive for AI systems that need massive amounts of memory-like storage near the processor.
To understand why this matters, it helps to look at how HBF works. High Bandwidth Memory stacks DRAM dies vertically and connects them using Through Silicon Vias, or TSVs, to deliver very high bandwidth in a compact package. HBF follows a similar stacking concept, but instead of DRAM, it stacks NAND flash dies. A controller logic die is bonded to the NAND stack to manage data access across the array.
The challenge is that NAND flash is far slower than DRAM. SRAM can deliver read speeds in the range of about one nanosecond, while DRAM is closer to around 100 nanoseconds. NAND flash, by comparison, can take around 100 microseconds for a read operation. In simple terms, NAND can be roughly 1,000 times slower than DRAM in read latency, which seems like a major disadvantage for any technology trying to compete with HBM.
HBF attempts to overcome this weakness through massive parallelism. Instead of relying on a single NAND read operation, the logic die schedules thousands of NAND reads at the same time. Each individual NAND cell is still slow, but when thousands of operations are performed in parallel, the total bandwidth can rise dramatically. Current expectations suggest HBF could deliver bandwidth ranging from about 0.3 TB/s to 3 TB/s.
That does not mean HBF is a direct replacement for HBM in every situation. NAND still has poor write speeds compared with DRAM, and write endurance remains a major limitation. HBF is therefore better suited for workloads where data is written less frequently but read repeatedly at high bandwidth.
This is especially important for artificial intelligence. One possible use case is storing AI model weights, where huge amounts of data need to be accessed quickly but are not constantly rewritten. Another promising application is shared prefix KV caching. In large language model systems, many users may send different prompts that rely on the same large system prompt, codebase, or long reference document. Instead of recalculating that shared context repeatedly, the system can compute the KV cache once and store it in HBF for repeated access.
If HBF can provide high bandwidth at a much lower cost than HBM4, AI companies and data center operators may have a strong incentive to adopt it. The technology could help reduce the memory cost of large AI infrastructure while still delivering enough performance for specific read-heavy tasks.
However, there is a major supply-side concern. An 8-Hi HBF stack is expected to use around 6.5 times more wafers than typical TLC NAND. Since NAND wafers are also used to make SSDs, a commercial boom in HBF could absorb a large portion of available NAND production capacity.
That could create a classic Jevons paradox scenario. Because HBF is much cheaper than HBM, demand for it could rise quickly. But as more NAND wafer supply gets redirected toward HBF production, the supply available for conventional SSDs may tighten. This could prevent SSD prices from falling as much as many consumers and businesses might expect.
In other words, HBF may make some AI memory systems more affordable while indirectly keeping consumer and enterprise SSD prices higher for longer. The technology could be a win for AI hardware economics, but not necessarily for anyone hoping to see a major drop in storage prices.
SanDisk is currently working with SK hynix on an HBF standard that is expected to offer 512GB of storage and bandwidth in the range of 0.4 TB/s to 3 TB/s. Commercial availability is reportedly targeted around the 2028 to 2029 timeframe.
If those timelines hold, HBF could become a key part of the next generation of AI data centers. Its combination of high capacity, strong bandwidth, and much lower cost than HBM4 makes it difficult to ignore. Still, the broader impact on the NAND market could be significant, especially if AI demand grows faster than manufacturers can expand wafer production.
For now, HBF remains a technology to watch closely. It may not replace HBM outright, and it will not solve every memory bottleneck in AI computing. But for read-heavy workloads that need large capacity and high bandwidth at a lower price, High Bandwidth Flash could become one of the most important memory innovations of the next several years.
