HWBRIEF
HBM is the stacked DRAM placed next to AI accelerators; Samsung says it must evolve 10x by 2030 to keep AI assistants streaming replies at conversational speed.
Samsung says AI systems need to hit 1,000 tokens per second per user by 2030, a 10x jump from today's roughly 100, and is pitching a 3D memory roadmap to get there, per EE Times' report on a Samsung pre-briefing for the FMS 2026 keynote, the annual Future of Memory and Storage summit.
The roadmap centers on high-bandwidth memory (HBM), the stacked DRAM placed next to AI accelerators to feed them data. "AI accelerators are bottlenecked on how fast they can pull data from nearby memory," Leno Park, Samsung's VP of flash solutions, told the pre-briefing, according to Samsung's own write-up of the same FMS keynote. Samsung's answer: denser stacks, true 3D layouts, and a "heat pipe block," a chimney over hot spots in the stack.
Two forecasts anchor Samsung's case. The 1,000-tokens-per-second-per-user target is roughly the speed at which an AI assistant streams a reply. Samsung also projects HBM will be more than half of all DRAM sales by 2030. Both figures are Samsung's own projections, not industry consensus.
The hardware is pre-production. HBM4E, built on a 4-nm base die with roughly 4x more through-silicon vias (vertical interconnects inside each stack) and 300,000+ microbumps, is in sampling with partners evaluating devices. HBM5 targets a 2-nm base die with gate-all-around transistors, a next-gen architecture that wraps the gate around the channel for tighter current control. A concept called zHBM would move from today's side-by-side layout to true 3D stacking of an accelerator on top of its memory.
Samsung is betting its roadmap is what AI infrastructure needs to look like by 2030. Whether the targets hold is the open question.