Business · Companies · 4 hrs ago
Samsung and SK Hynix expand AI memory race beyond HBM
As AI models grow and AI services spread, memory needs to move data faster and use power efficiently.
South Korea’s Samsung Electronics and SK Hynix are competing to develop memory technologies that go beyond improving high-bandwidth memory, or HBM.
Samsung’s proposed zHBM would stack memory directly on an AI accelerator to shorten the distance data travels; the company says it could deliver up to eight times the performance per GPU of HBM5.
SK Hynix is developing 3D-stacked DRAM and high-bandwidth flash, or HBF, which aims for about ten times the capacity of same-generation HBM at half its bandwidth.
Both companies have also started mass-producing SOCAMM2, a low-power memory module for servers, and are developing CXL-based ways to expand server memory.
These technologies target limits in AI systems, including power use, data bottlenecks and growing demand for storage.
When the new technologies reach the market and whether they cut costs will help determine the companies’ competitiveness.
Samsung Electronics and SK Hynix are expanding their AI memory competition beyond high-bandwidth memory (HBM) to technologies aimed at improving speed, power efficiency and storage capacity.
Samsung is promoting zHBM, which stacks HBM directly on an AI accelerator, while SK Hynix is developing HBF, a high-bandwidth flash memory.
Samsung says zHBM could deliver up to eight times the performance per GPU and three times the performance per watt of HBM5.
SK Hynix aims for HBF to offer about 10 times the capacity of HBM from the same generation, with half its bandwidth.
The companies are also competing in low-power server memory modules and CXL-based memory expansion.
- Who
- Samsung Electronics and SK Hynix.
- What
- They are developing memory technologies for AI infrastructure beyond HBM, including stacked memory, high-bandwidth flash and low-power modules.
- When
- The report was published on October 23, 2026, and dated October 24 in Seoul.
- Where
- Not stated
- Why
- Larger AI models and the spread of inference and AI agent services are increasing demand for memory speed, power efficiency and data storage and management.
Samsung Electronics
SK Hynix
Stacked memory
Samsung Electronics
Samsung’s zHBM would stack HBM directly on an AI accelerator to reduce data travel and improve performance and power efficiency.
SK Hynix
SK Hynix proposes 3D-stacked DRAM above processors such as GPUs and NPUs to bring data closer to computing units.
Next-generation storage
Samsung Electronics
Samsung is presenting zNAND-O, a high-performance NAND solution for on-device AI.
SK Hynix
SK Hynix is developing HBF, combining HBM-like data speed with NAND-like storage capacity.
Samsung Electronics and SK Hynix began mass production of SOCAMM2 low-power server memory modules.
Samsung said it had completed validation of 16GB LPDDR6 on Qualcomm’s Snapdragon 8 Elite Extreme 6th-generation platform.
The companies presented or highlighted next-generation memory technologies at semiconductor technology exhibitions.
- zHBM
- Samsung’s design stacks HBM directly on an AI accelerator.
- zHBM performance claim
- Up to 8 times the performance per GPU and 3 times the performance per watt compared with HBM5.
- HBF capacity target
- About 10 times the capacity of same-generation HBM.
- HBF bandwidth target
- Half the bandwidth of same-generation HBM.
- SOCAMM2
- A low-power server memory module based on LPDDR.
1 source
What people are saying
The center of AI memory competition is moving from the stage of improving the performance of individual products to the stage of improving the overall structure and operational efficiency of servers and other systems.