High-bandwidth flash could become a new way to ease the memory bottleneck in AI inference, with Goldman Sachs estimating it might cut inference-throughput costs by about 30%. The technology is still early, but the forecast puts SK Hynix, SanDisk and Samsung in the frame as potential beneficiaries.
SK Hynix Watch analysis
What happened
A Goldman Sachs analyst says High-Bandwidth Flash, or HBF, could help AI chips make better use of their computing capacity by supplying more memory bandwidth. Yahoo Finance, carrying an Investing.com report, says the bank estimates a potential HBF market of roughly $40 billion to $150 billion by 2030, depending on adoption. Its estimate of around 30% lower inference-throughput costs is a forecast, not a measured result.
Goldman names SanDisk and SK Hynix as technology leaders in HBF, and also rates Samsung a Buy on the strength of its NAND portfolio. The analyst warns that these technologies remain at an early stage of system-level development, with a wide range of possible outcomes. The report also points to Kioxia’s XL-Flash and Samsung’s zNAND-O as alternative approaches.
Why it matters
Inference is where AI models do their work for users, and moving data to and from memory can constrain how efficiently chips are used. If HBF can ease that bottleneck, it could alter the economics of running AI services and create a significant new market for memory makers. Goldman’s range is wide for a reason: adoption and technical execution remain uncertain.
That makes this more than a stock-picking note, but not yet a shopping list for the data-centre aisle. A large projected market is not proof that the technology will arrive on schedule, deliver the estimated savings or be adopted across the industry.
Our read
The useful signal is that memory makers are looking beyond the familiar race for faster, higher-bandwidth memory. HBF is an intriguing possible addition to the AI infrastructure toolkit, and SK Hynix is one of the companies Goldman thinks could benefit. For now, treat the cost reduction and market-size figures as analyst estimates, not settled engineering results. The next interesting evidence will be working systems and customer adoption, not another grand total with several zeroes.
What to watch
- Whether memory makers publish clearer HBF development or deployment plans.
- Whether system-level tests support the claimed inference-cost savings.
- How adoption compares with Goldman’s wide range of possible market outcomes.
- Whether HBF gains ground against other flash-memory approaches.
Discussion spark: If HBF really can lower inference costs, should AI companies prioritise it over adding more conventional memory bandwidth, or is the technology too early to bet on?
Sources and evidence
- Goldman names 3 names to benefit as high-performance NAND tackles AI’s memory wall – Yahoo Finance (12 October 2026, 11:24 UTC)
not affiliated with or endorsed by SK Hynix