Via en.wikipedia.org
The new memory technology promises up to 512 GB per stack and could reshape the AI inference hardware landscape by offering a cheaper alternative to HBM
SanDisk just finished designing its first High Bandwidth Flash memory die, a milestone that puts the company on track to deliver product samples in 2027.
High Bandwidth Flash, or HBF, is SanDisk’s bet that you can take existing NAND flash technology, stack it in clever ways, and get performance that comes remarkably close to the pricier High Bandwidth Memory chips currently dominating AI hardware. Close as in within 2.2% of theoretical HBM limits, according to SanDisk’s projections.
What HBF actually does #
The first generation targets up to 512 GB per stack with read bandwidths around 1.6 TB/s. For context, that capacity figure represents roughly 8 to 16 times what existing HBM systems can offer in comparable configurations. Future iterations on the roadmap push performance to 2 TB/s and eventually 3.2 TB/s.
SanDisk plans to offer initial HBF configurations in 8-high and 16-high stacks, with bandwidth grades spanning from 0.4 TB/s to 3 TB/s depending on the configuration. The technology leans on SanDisk’s proprietary CBA and BiCS NAND processes, both of which the company has been refining for years in its conventional flash business.
The critical distinction here is the target workload. While HBM excels at training AI models, where raw speed matters most, HBF is designed specifically for inference. That’s the phase where a trained model actually does useful work: answering questions, generating images, recommending products.
The roadmap and key partnerships #
SanDisk, which now operates as an independent flash memory entity following its separation from Western Digital, formed a technical advisory board in July 2025 to guide specification development and attract ecosystem partners.
In August 2025, SanDisk signed a memorandum of understanding with SK hynix. The two companies agreed to collaboratively develop HBF specifications, a notable arrangement given that SK hynix is currently the dominant supplier of HBM chips.
The first open specifications for HBF are scheduled for release in August 2026. Those specs will detail memory stack configurations, achievable bandwidths, and power efficiency characteristics.
SanDisk’s timeline puts initial HBM samples arriving in late 2026, with the HBF die samples following in 2027. The company also picked up recognition at the Flash Memory Summit for the technology’s innovative approach.
Why this matters for AI infrastructure #
A technology that can deliver 512 GB per stack at near-HBM speeds could meaningfully change the economics of deploying large language models at scale. SanDisk’s ability to leverage existing NAND manufacturing infrastructure is another practical advantage, as its approach of using established BiCS and CBA processes means production could scale more predictably than a greenfield memory technology.
If the August 2026 specifications earn broad industry buy-in, system integrators and chip designers can start planning HBF into their 2028 and 2029 product roadmaps. Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our