Zerohedge Says Bernstein and Goldman See High Bandwidth Flash as a Way Around AI's Memory Wall, if Nvidia Adopts It
Flash Forward: Bernstein, Goldman Say "High-Bandwidth Flash" Could Crumble AI's Memory Wall... If Nvidia Ever Shows Up
Context
Memory is a growing limit on AI inference costs, so a cheaper, larger memory tier next to GPUs could change chip economics and which memory makers win.
High bandwidth flash, or HBF, is an emerging memory tier built from stacked NAND flash. It is meant to sit next to AI chips alongside the much more expensive high bandwidth memory, or HBM, giving models far more capacity for inference at lower cost. No major AI chip maker has yet announced a product that uses it.
Deep dive
Zerohedge (@zerohedge) posted on Nov 1 a link to its article titled Flash Forward: Bernstein, Goldman Say High Bandwidth Flash Could Crumble AI's Memory Wall... If Nvidia Ever Shows Up. The post gives only the headline, so the details come from the article and the bank research it cites.
The article opens with a warning for the memory trade. It says Kioxia guided its December quarter operating profit to roughly flat after NAND price gains slowed from 70% to the mid 20s in the September quarter, and that its fab partner Sandisk fell more than 11% on Friday despite strong results.
The core argument is about a new kind of memory. AI models need huge amounts of fast memory to serve long contexts and large batches, and HBM is scarce and costly. HBF stacks NAND flash to offer much more capacity per chip at lower cost, with less speed. In an Oct 12 note reported by Investing.com, Goldman Sachs analyst James Schneider put the potential HBF market at about $40 billion to $150 billion by 2030, depending on whether 15% or 50% of AI chips adopt it. He estimated it could cut the cost of inference throughput by about 30%, and named Sandisk, SK Hynix and Samsung as likely winners. Bernstein's view is cited by zerohedge and could not be checked against the original note.
The catch is in the headline. HBF only matters at scale if Nvidia or another major accelerator maker designs it into its systems, and so far the leading platforms rely on HBM. The technology is still early at the system level.
What to expect next
Watch for Sandisk and SK Hynix sample timelines and for any sign that Nvidia, AMD or a hyperscaler chip team plans to support HBF in a future platform. Memory makers' next earnings calls will show whether NAND pricing keeps cooling or AI demand lifts it again. If a major chip design adopts HBF, flash makers could gain a much bigger share of AI spending.