Samsung’s zHBM Stacks Memory on AI Chips, 8x Faster

As of August 5, 2026, Samsung Electronics has unveiled zHBM, a next-generation 3D memory that stacks vertically right on top of AI accelerator chips. Compared with the still-in-development HBM5, it promises up to 8x the GPU performance and up to 3x better performance-per-watt. At the same event, SK hynix countered with its own open-standard NAND-based memory tier, HBF.

8xzHBM's max per-GPU performance gain over HBM5

3xzHBM's performance-per-watt improvement over HBM5

1/6Memory cost for running a 120B-parameter model with zNAND-O, vs. without it

400+ layersLayer count of the newly unveiled 10th-gen V-NAND

Key points

Samsung showed off a physical zHBM mockup for the first time during a keynote at a major US memory industry event.
Unlike conventional HBM, which sits beside the accelerator, zHBM is stacked directly on top of the chip, shortening the distance data has to travel.
Samsung says its on-device AI NAND solution, zNAND-O, cuts the memory cost of running a 120-billion-parameter model to one-sixth.
SK hynix, together with SanDisk, unveiled the HBF standard at the same event, betting on an open ecosystem multiple companies can adopt.

As warnings mount that AI chip bottlenecks are shifting from processors to memory, Korea's two big memory makers shared the stage at the world's largest memory and storage industry event, held in California. Both companies agreed that explosive growth in tokens processed during AI inference has outpaced what conventional memory architectures can handle. The presentations took place the day before, local time.

Samsung's answer is to stack memory vertically instead of spreading it out horizontally. zHBM takes high-bandwidth memory — traditionally placed beside the accelerator — and stacks it directly on top of the chip, shrinking the distance signals need to travel. By combining full-wafer bonding with direct copper-to-copper connections, Samsung packed in far denser wiring, cutting thermal resistance to 10-25% of what's expected from the still-in-development HBM5. The Samsung executive who presented said the company plans to co-design zHBM with processor makers from the start, rather than selling it as an off-the-shelf product.

Samsung's 3D push extended to storage too. The company unveiled the industry's first 10th-generation V-NAND, stacked past 400 layers, alongside zNAND-O, an on-device AI solution that bundles multiple NAND layers together. The design hands fast, temporary data to DRAM while storing the bulky AI models themselves on NAND-based memory; mass-production samples are slated to ship starting in 2028.

SK hynix took a different route. Together with SanDisk, it finalized and unveiled the HBF standard, which stacks NAND in 8 or 16 layers to deliver up to 512GB of capacity and transfer speeds of up to 3TB per second. Built around the industry-wide UCIe interconnect standard, HBF is designed to plug into a range of different processors, with Google and Tenstorrent both taking part in the standardization effort. Where Samsung is betting on doing everything in-house, SK hynix is betting on a wider, shared ecosystem.

Market conditions form the backdrop for both announcements. Samsung held onto the No. 1 spot in global memory revenue in the first quarter of this year, and market researchers expect the overall memory market to expand sharply next year. One French research firm predicted Korea will keep its lead in memory for years to come, citing both companies' rising investment and government support — while also flagging a new wrinkle: as AI shifts from training to inference, product development cycles are speeding up.

In Korea right nowAt home, attention has focused on how the two companies' bets point in opposite directions. Samsung and SK hynix shares both ticked up slightly after the announcements, with analysts saying the real test is whether Samsung's vertically integrated approach wins over major customers, and how many companies rally behind SK hynix's open HBF standard. An overseas forecast that Korea will hold its memory-industry lead through 2031 is also drawing attention.

FAQ

How is zHBM different from conventional HBM?

Conventional HBM sits beside the AI accelerator chip, while zHBM stacks memory vertically directly on top of the accelerator, minimizing the distance data has to travel.

How does SK hynix's HBF differ from zHBM?

HBF is a NAND-flash-based memory tier built for much larger capacity. Rather than replacing HBM, it's meant to work alongside HBM to support large-scale AI inference.

When will these technologies actually reach products?

Mass-production samples of zNAND-O are expected to ship starting in 2028, and the industry broadly expects NAND-based memory tiers like HBF to see full-scale commercialization around 2030.