By Rebecca Stone

Goldman Sachs has flagged three stock names it expects to benefit as high-performance NAND flash steps into a larger role solving AI's so-called memory wall, according to a research note covered by StreetInsider.

The note singles out vendors positioned to supply AI data centers with the faster, denser flash that training and inference workloads increasingly demand. Goldman's three tickers weren't detailed in the headline surfaced by StreetInsider, but the thesis points squarely at NAND's most concentrated growth pocket.

What's the memory wall?

In accelerator-heavy systems, the gap between how fast GPUs and ASICs crunch numbers and how quickly memory can feed them has widened for years. HBM stacks and on-package SRAM address the problem inside the chip.

Outside the chip, AI clusters need storage that can ingest checkpoints, embeddings, and terabyte-scale datasets without becoming the bottleneck. That is where high-performance NAND — typically 3D NAND in TLC or QLC form riding PCIe Gen5 NVMe controllers — steps in. Enterprise SSDs built on that NAND now push past 30 TB per drive, and PCIe Gen5 controllers have roughly doubled sequential throughput versus Gen4 parts from 2022.

Why now?

Three forces are converging. Frontier model parameter counts keep climbing, with GPT-class systems jumping from hundreds of billions to trillions of parameters in roughly three years. Training datasets themselves have grown from tens of terabytes into the petabyte range. Inference at scale, including retrieval-augmented generation pipelines, reads from vast vector stores that must sit on fast flash. And QLC NAND's unit economics have finally made it cost-effective to push AI datasets onto flash rather than spinning media or tape.

The NAND market itself has consolidated hard after years of price erosion. Samsung, SK hynix (including Solidigm), Kioxia, Micron, and SanDisk together hold roughly 90% of bit shipments. On the controller side, Marvell, Phison, and Silicon Motion compete for the silicon that turns raw NAND into a usable PCIe SSD, and each new generation — from 176-layer to 232-layer and the upcoming 300-layer parts — has demanded fresh controller designs to keep pace with die-level throughput.

What does Goldman expect?

Goldman's three names sit somewhere in that supply chain, the StreetInsider headline indicates, though the underlying report behind a Google News paywall lays out the specific tickers. What the bank is signaling matters as much as the names: institutional capital is now treating high-performance NAND as an AI infrastructure play, not a commodity memory trade.

For NAND vendors, the timing is telling. NAND wafer prices began firming in late 2024 after roughly two years of downturn that pushed smaller players out of merchant 3D NAND. Samsung and SK hynix have both signaled tighter supply as they rebalance capacity toward HBM and AI-grade flash through 2025.

What's the catch?

NAND cycles are notoriously cyclical. The last boom, in 2017 and 2018, ended in a brutal 18-month price collapse that wiped out margins across the supply chain. Any new fab capacity from the major suppliers, or a pullback in AI capex from the hyperscalers, could reverse the current rally just as quickly. Goldman itself has hedged past NAND calls in earlier cycles, and its bullishness on AI infrastructure spending assumes the capex plans from Microsoft, Google, Amazon, and Meta hold.

The forward view is straightforward: as long as AI model training runs keep demanding more checkpoint storage and inference keeps leaning on fast vector databases, the high-performance NAND tier will grow faster than the rest of the flash market. Whether Goldman's three picks capture that delta will depend on which corner of the supply chain each one occupies — controller IP, branded SSD, or raw NAND wafer — and how exposed each is to the AI hyperscaler customer base driving most of the buying.