The South Korean technology group has unveiled denser, vertically integrated memory designs intended to reduce the speed, power and heat constraints confronting the next generation of artificial-intelligence systems.
Samsung Electronics has introduced a new generation of artificial-intelligence memory technology, intensifying the global race to overcome one of the most important obstacles facing advanced computing: moving enormous quantities of data quickly enough to keep increasingly powerful AI processors operating efficiently.
The world’s largest memory-chip manufacturer presented its V10 Bonding V-NAND prototype at the Future of Memory and Storage conference in California. The chip contains more than 400 layers and uses a wafer-bonding architecture that separates and then joins its memory-cell and peripheral circuitry, allowing Samsung to increase storage density while improving performance.
Samsung said the new BV-NAND design delivers approximately 58 percent greater memory density than its previous V9 generation. The company also reported improvements in read, write and input-output performance, which could make the technology particularly valuable for AI servers that must rapidly retrieve large datasets while generating responses, images, video or software code.
The announcement reflects a broader shift in the artificial-intelligence industry. Early investment centred primarily on the processing chips used to train large language models. As AI systems move from training into widespread commercial use, however, memory and data storage are becoming equally important.
Every request sent to an AI service requires information to be retrieved, transferred to an accelerator and processed before an answer can be returned. When millions of users interact with the same models simultaneously, delays in accessing memory can leave expensive processors waiting for data. The industry describes this constraint as the “memory wall”—a widening gap between the speed of modern computing engines and the systems that supply them with information.
Samsung is seeking to address that bottleneck by changing not only the number of memory layers inside a chip but also their position relative to the processor.
Alongside BV-NAND, the company displayed concepts known as zHBM and zNAND-O. Conventional high-bandwidth memory is normally installed beside an AI accelerator and connected through advanced packaging. Samsung’s zHBM concept would instead place memory vertically above the accelerator, shortening the distance that electrical signals must travel.
The company said its wafer-bonding approach could eventually provide more than 10 times the memory density of conventional HBM5, while tripling energy efficiency and reducing thermal resistance by more than half. Those claims concern a concept architecture rather than a mass-produced commercial product, but they indicate the direction in which semiconductor manufacturers believe AI infrastructure must develop.
Reducing energy consumption has become a critical objective for the technology industry. AI data centres require vast quantities of electricity not only to operate processors but also to cool them. Placing memory and computing components closer together can reduce the energy consumed when data moves between chips, although vertically stacking components also creates complex manufacturing and heat-management challenges.
Samsung’s announcement therefore represents both a technological proposal and a strategic declaration. The company is competing with South Korean rival SK Hynix and US manufacturer Micron for leadership in high-bandwidth memory, one of the most profitable and strategically significant segments of the semiconductor market.
Demand has risen sharply as Nvidia and other chip designers require increasingly large quantities of advanced memory for their AI accelerators. Samsung has already begun selling HBM4 products for Nvidia’s Vera Rubin platform and has said it intends to more than triple its high-bandwidth-memory revenue this year.
The company has also expanded its cooperation with American chip designer Broadcom. Under an agreement announced in July, the two groups plan to collaborate on next-generation high-bandwidth memory, advanced packaging and the manufacture of specialised AI processors using Samsung’s sub-two-nanometre production technology.
Such partnerships illustrate how AI is restructuring the semiconductor supply chain. Technology companies are increasingly designing custom accelerators for specific workloads rather than relying exclusively on general-purpose graphics processors. That trend creates opportunities for manufacturers capable of combining logic chips, memory and advanced packaging within a single integrated system.
Samsung is positioning itself as one of the few companies able to provide all three.
Its new storage designs could ultimately affect consumer technology as well as data centres. NAND flash memory is used in smartphones, tablets, computers and solid-state drives to store applications, photographs, videos and other information. More compact and energy-efficient NAND could permit mobile devices to carry greater storage capacity while running more sophisticated AI models locally.
Samsung has already introduced UFS 5.0 mobile storage capable of transferring data at up to 10.8 gigabytes per second, describing it as a platform for next-generation on-device AI applications. Faster storage can help smartphones retrieve information for real-time translation, generative photo editing, voice assistants and other functions without sending every request to a remote server.
On-device processing offers potential advantages in speed, privacy and reliability. A phone that can perform an AI task locally may respond more quickly, function without a permanent internet connection and avoid transmitting sensitive personal information to the cloud. The difficulty is fitting sufficient processing power, memory and battery capacity into a small device that must remain affordable and cool enough to hold.
Advances such as BV-NAND could ease some of those constraints, although Samsung has not announced when the prototype will enter mass production or which commercial products will use it first.
The launch comes at a favourable moment for the company’s semiconductor business. AI-related demand has driven record profits and encouraged major customers to secure memory through long-term supply agreements. Samsung has said such contracts could eventually account for between 60 and 70 percent of its memory sales, reducing its dependence on the semiconductor industry’s traditionally volatile spot market.
Inventories of both DRAM and NAND remain below historical averages, according to analysts cited by Reuters, despite concerns that weaker smartphone demand could eventually reduce consumption of conventional memory. AI infrastructure has so far more than compensated for softness in some consumer markets.
That imbalance could also carry consequences for consumers. As chipmakers devote more production capacity to lucrative AI data-centre components, supplies of conventional memory may remain tight and prices could rise. Samsung has warned that the global memory shortage could persist through 2028 because building new fabrication plants is costly and takes several years.
The result is an unusual division within the technology market. The AI boom is generating extraordinary profits for memory manufacturers while simultaneously increasing component costs for smartphones, computers and other consumer electronics.
Samsung’s latest designs are intended to expand capacity and efficiency over the longer term. Yet their importance extends beyond the specifications of a single chip. They demonstrate that the next phase of artificial intelligence will depend not only on creating faster processors, but also on redesigning the physical architecture surrounding them.
The future of AI may therefore be built vertically. By stacking storage and memory closer to computing engines, Samsung is betting that the industry can extract more performance from every unit of space and electricity.
Whether its experimental architectures become commercial standards remains uncertain. What is increasingly clear is that memory—once treated as a supporting component—is becoming one of the central battlegrounds of the global AI economy.
