Samsung has revealed a new idea for building computer memory called zHBM, which would place memory chips directly on top of the artificial intelligence processors that use them. The company showed early concept models at the Future of Memory and Storage (FMS) 2026 show in Santa Clara. The announcement, reported on August 4, 2026 by both Samsung's newsroom and Reuters, describes what Samsung calls a fundamentally new approach to 3D memory stacking for AI data centers.

To picture why this matters, think of a computer chip and its memory as two workers who need to constantly hand documents back and forth. Today, they sit in adjacent offices, separated by a hallway. Moving that hallway takes time and energy. zHBM would move them into the same room — stacking the memory directly on top of the processor so the data barely has to travel.

Samsung claims zHBM boosts performance by up to eight times compared to today's memory arrangements. The design uses a new manufacturing technique called wafer bonding — essentially fusing thin layers of silicon together — to pack over 10 times more memory into the same space than current technology allows. Samsung also reports a threefold increase in energy efficiency. The company presented these as concept models, not finished products you can buy. Engadget, publishing August 5, 2026, provided the detailed specifications.

Alongside zHBM, Samsung introduced V10 BV-NAND, a type of flash storage memory for SSDs, memory cards, and other devices. Flash memory is the technology inside your phone and USB drives that stores data even when powered off. Samsung builds its flash by stacking memory cells in vertical layers, like floors in a building. V10 BV-NAND stacks more than 400 layers, increasing memory density by about 58 percent over the previous generation, called V9. Samsung says it also improves read, write, and data transfer speeds.

Samsung also unveiled zNAND-O, a new flash memory designed for "edge AI" — artificial intelligence running on everyday devices like phones and cars instead of in giant data centers. It focuses on fitting more memory into less space and responding quickly. The company showcased LPDDR5X-PIM at its booth as well, describing it as the industry's first mobile memory chip that can do small amounts of computing directly inside the memory itself. Normally, memory just stores data and a separate processor does the math; this new approach reduces the back-and-forth between the two. Samsung's FMS 2026 booth, designed to look like an AI cloud server, displayed about 30 memory and storage technologies in total.

These announcements come at a time of enormous financial success for Samsung's chip business. The company's second-quarter 2026 operating profit jumped 19-fold year on year, with chip profit rising more than 250-fold to $61.7 billion, Reuters reported on July 30, 2026. Samsung has signed multi-year supply deals with major data center operators lasting into 2028 and expects the memory chip shortage to continue until then. In March 2026, Samsung's CEO pushed for these long-term contracts, sending shares to record highs. Samsung also began shipping samples of its latest high-bandwidth memory chip to customers in June 2026.

Samsung has been climbing this particular ladder for over a decade. The company started mass-producing its first 32-layer vertical flash memory in 2014, reached 48 layers in 2015, and introduced its sixth generation for personal computers in 2019. Each generation roughly doubled the layer count. V10 BV-NAND's 400-plus layers are where that growth curve has been heading, and the shift to bonded layers suggests Samsung is now combining wafers at the factory level rather than simply stacking higher.

The broader context here is that the biggest challenge in AI computing is no longer just how fast memory can send data. It is about how far that data has to physically travel, how much power that travel costs, and how much hardware you can fit in a building. zHBM's claimed 8x performance and 3x energy efficiency would, if Samsung can actually build it at scale, address all three problems at once by removing the physical gap between memory and processor. The wafer bonding that makes this possible is the hardest part to manufacture reliably, and it is the piece most likely to cause production problems.

It is also worth noting that Samsung is calling these concept models, not products with shipping dates. The distance between a demo at a trade show and mass production of bonded 3D chip stacks could be measured in years. Samsung's own forecast that memory shortages last through 2028 is both a supply problem and a financial reason to push these designs into production as quickly as the manufacturing allows.

The LPDDR5X-PIM announcement is quieter but could reach more people. The idea of computing inside memory has been researched for over a decade. Putting it into a mobile memory chip targets phones and other devices where the battery cost of moving data back and forth between processor and memory is the biggest drain. Samsung's separate focus on zNAND-O for edge devices shows a clear strategy: data centers get the vertical memory stacking, and everyday devices get memory that does some of its own computing.

Samsung's flash-based SSDs have historically delivered about twice the lifespan and 20 percent lower power use compared to the older flat memory technology. Those advantages helped flash storage replace traditional hard drives in data centers, where flash is now the standard. V10 BV-NAND's 58 percent density improvement over the previous generation keeps driving down the cost of storage, which matters for the data storage layer of AI systems even as zHBM targets the memory sitting closest to the processors.

Taken together, the FMS 2026 announcements sketch a vision where memory technology spans from data-center AI processors down to the phone in your pocket, all built using the same underlying approach of bonding and vertical stacking. Samsung's record profits and contracts through 2028 give it the money and the reason to turn these concepts into real products.