Investing.com -- The rapid expansion of artificial intelligence workloads is creating a growing demand for memory capacity and bandwidth, opening the door for a new storage technology that could ease infrastructure constraints while providing a fresh growth opportunity for the NAND flash industry, Bernstein said.
In a research report, Bernstein highlighted the potential of high-bandwidth flash (HBF), a technology designed to combine read bandwidth comparable to high-bandwidth memory (HBM) with eight to 16 times greater capacity at a significantly lower cost per gigabyte. The technology could make AI inference more efficient by reducing the amount of expensive GPU memory required to run increasingly large models and allowing AI systems to retrieve previously processed information instead of repeatedly performing costly computations.
The shift from AI training toward advanced inference and agentic workloads is making memory capacity, bandwidth and data reuse increasingly important. Larger models, longer context windows and more complex interactions are pushing data beyond the limits of HBM, expanding demand for conventional DRAM, NAND flash and hard disk drives.
Bernstein said HBF could address this challenge through a tiered memory architecture in which HBM handles frequently accessed and latency-sensitive data, while HBF stores larger model parameters, less frequently used model components and reusable key-value (KV) cache data. This could allow AI infrastructure providers to run larger models with fewer GPUs and reduce computing, power and capital expenditure requirements.
The technology could also create a significant opportunity for NAND manufacturers. Bernstein estimates that HBF could consume approximately 6.5 times more NAND wafer capacity per bit than conventional triple-level cell (TLC) NAND, reflecting its use of higher-performance single-level cell (SLC) flash, additional silicon area and advanced packaging requirements.
Bernstein estimates that HBF manufacturing costs could be roughly 10 times higher than those of conventional enterprise SSDs, depending on the configuration. However, if manufacturers price HBF modules at levels comparable to HBM while offering substantially greater capacity, the technology could generate attractive gross margins. Under its base-case assumptions, Bernstein estimates gross margins of about 86.5% for an eight-high HBF configuration and 88.9% for a 16-high configuration.
Despite the opportunity, Bernstein cautioned that HBF faces significant adoption hurdles. The technology is not a direct replacement for HBM and has limitations involving latency, write bandwidth, endurance and data access. It would also require changes to processor designs and system software, alongside careful optimization of data placement and memory management.
A key concern is the lack of formal endorsement from Nvidia, which has instead promoted alternative technologies such as Storage Next and its CMX content memory storage platform to address the growing demand for KV-cache capacity. Alphabet Inc Class A (NASDAQ:GOOGL), Meta Platforms Inc (NASDAQ:META), SK hynix Inc ADR (NASDAQ:SKHY) and Tenstorrent are among the companies involved in the HBF ecosystem, but broader industry adoption remains uncertain.
SanDisk Corporation (NASDAQ:SNDK) has moved HBF beyond the conceptual stage toward standardization and initial silicon development, with the company targeting its first product samples for inference applications in 2027. Bernstein said the technology could provide additional upside to the company’s financial targets, although HBF is not currently included in those targets.
Bernstein maintained an Outperform rating on SanDisk with a $3,000 price target. It also rated Seagate Technology PLC (NASDAQ:STX), Western Digital Corporation (NASDAQ:WDC), Samsung Electronics Co Ltd (KS:005930), SK hynix and Micron Outperform, while assigning KIOXIA an Underperform rating.
The analysts said HBF could ultimately do for the NAND industry what HBM has done for DRAM by creating a premium AI-driven demand category and tightening supply through higher wafer consumption. However, its commercial success will depend on technological execution, customer adoption and competition from alternative memory architectures.
The fastest way to find out is looking at our Fair Value estimate. It runs 17 proven valuation models on GOOGL and gives you one number, plus the upside or downside against today’s price. Do the same for thousands of other stocks and you stop guessing which ones are already priced for perfection. 370,000+ investors rely on powerful InvestingPro tools from Investing.com to validate their trading ideas . Full access now 55% off for Black Friday early birds.