Samsung says its new zHBM memory architecture, stacked directly on top of AI chips, runs eight times faster than HBM5, a standard that hasn't even shipped yet.
Samsung Electronics pulled back the curtain on zHBM at the Future of Memory and Storage conference in Santa Clara on August 4. The concept model stacks high bandwidth memory directly on top of AI accelerators, using wafer-bonding technology. According to Samsung's own announcement, the design cuts the physical distance data has to travel between processor and memory to almost nothing. The company says a next-generation interface built around zHBM should hit roughly eight times the performance of HBM5, while using up to three times less power per unit of work and cutting thermal resistance by more than half.
That's not a modest claim. It's Samsung telling the industry it plans to leapfrog a memory generation that rivals are still building toward.
The company also showed off two other pieces of its roadmap at Booth 407: V10 BV-NAND, which Samsung calls the industry's first flash memory to cross 400 layers, and zNAND-O, a flash architecture built for edge AI work like autonomous driving and robotics. That's inference that has to happen on the device itself, not in a data center. Samsung built its whole booth around an AI cloud server theme, according to the Seoul Economic Daily, putting roughly 30 memory and storage technologies on display.
The timing isn't an accident. Memory has become the tightest bottleneck in AI hardware, and the numbers back that up. Micron has sold out its HBM production for 2026. Samsung itself has told customers to expect the global shortage to worsen through 2028. Deutsche Bank estimates 2026 DRAM demand will outstrip capacity by roughly 10%. That gap widens to 29% by 2028, as AI training and inference needs keep outrunning new fab capacity.
That scarcity has been brutal for buyers and great for shareholders. SanDisk stock jumped about 26% in a single trading day on July 30. Micron rose roughly 18% the same day. Both moves reflect renewed confidence that memory prices have further to climb, based on reporting from Invezz and 247wallst.com. Just days earlier, memory stocks had dropped more than 30% before rebounding. That swing shows how jumpy the market has gotten over a handful of supply signals.
Samsung versus SK Hynix, and the chip supply chain caught in between #
Samsung isn't the market leader here. That's the point of zHBM. SK Hynix held 62% of the HBM market as of mid-2025, with Micron at 21% and Samsung trailing at 17%, according to Astute Group's tracking of the sector. SK Hynix also remains Nvidia's primary supplier for HBM4, the memory slated to go into Nvidia's Rubin GPU platform. Samsung is betting that a more radical architecture change, rather than an incremental HBM4E refresh, is how it closes that gap.
SK Hynix isn't standing still either. It has already said its own 2026 production is sold out, and it's pushing HBM4 designs rated at 10 Gbps with a claimed 40% jump in power efficiency. Micron, meanwhile, has begun shipping HBM4 samples rated up to 11 Gbps. Every one of these companies is chasing the same two customers: Nvidia and AMD, whose GPUs are only as fast as the memory feeding them.
Samsung didn't give a shipping date for zHBM, and it's careful to call it a concept model, not a product. It's a roadmap slide, not a part number. The real test is whether Samsung can turn wafer-bonded memory into something it can manufacture at the volumes Nvidia and AMD will need. SK Hynix and Micron want to get there first, too.
Also read: OpenAI Buys Ona to Make Sure Codex Never Has to Sleep • A UK Government Test Caught an AI Agent Faking Identities to Hack GitHub • Trump Meets AI Giants While Senate Democrats Call His AI Policy Unpredictable