This is the AI bottleneck nobody argues about, and Samsung thinks stacking fixes it
Samsung unveiled zHBM at the FMS 2026 conference, claiming four times HBM4's transfer speed at a quarter of the power. SK hynix and Sandisk used the same stage to publish the first open High-Bandwidth Flash spec.
The thing slowing AI down right now is not the processor. It is the distance data travels between the processor and its memory, and that distance exists because in current designs the memory sits next to the chip rather than on top of it.
Samsung’s answer, shown on 4 August at the Future of Memory and Storage conference in Santa Clara, is called zHBM: stack the high-bandwidth memory vertically, directly above the AI accelerator. The company says zHBM delivers four times HBM4’s data transfer speed while drawing a quarter of the power, and that a next-generation interface system built around it should reach roughly eight times HBM5’s performance.
The rival route, using flash instead of DRAM
At the same conference SK hynix and Sandisk published the first open technical specification for High-Bandwidth Flash through the Open Compute Project. It defines modules with capacities up to 512GB and bandwidth of roughly 0.4 to 3 terabytes per second, developed with partners including Google DeepMind and Tenstorrent.
Two different bets on one problem. HBF swaps DRAM for flash to get far more capacity per module, which matters for models that no longer fit comfortably in available memory. zHBM keeps DRAM and attacks the geometry instead, shortening the physical path the data has to travel. Samsung, notably, sits outside the HBF consortium.
Why it suddenly matters
Because demand for compute has not slowed and the choke point has moved. Companies that used to simply buy GPUs are now designing their own inference silicon to control cost and performance, and on the manufacturing side TSMC has just posted its strongest July on record.
None of this reaches data centres this year. But it decides what it will cost, in electricity, to run the models of 2028.
By Oliver Grant
Image: ScotXW / Wikimedia Commons (CC BY-SA 4.0)
