Broker research increasingly frames Samsung Electronics’ AI opportunity as a full stack spanning HBM, advanced base dies, foundry and packaging—not memory alone.
Samsung Electronics’ AI story is starting to look broader than a memory-cycle trade. Recent Korean broker research argues that the company’s opportunity is expanding from high-bandwidth memory itself into the surrounding layers that determine how AI accelerators are built: base dies, advanced foundry processes, packaging and eventually more customized HBM architectures.
That shift matters because the economics of AI hardware are moving closer to the package. As accelerators demand more bandwidth and consume more power, the memory stack, the logic underneath it and the way those components are integrated can matter almost as much as the raw DRAM capacity.
HBM4 is the near-term bridge
KB Securities estimates Samsung’s HBM market share at 33% in the second quarter of 2026 and expects it to approach 40% in the fourth quarter. The same report forecasts that Samsung’s HBM4 revenue in the third quarter will rise by more than three times quarter on quarter, with HBM4 accounting for more than 60% of second-half HBM revenue.
Those figures are broker estimates rather than independently verified market-share data. But the direction of the argument is important: Samsung is expected to move further into the highest-performance portion of the memory market just as HBM becomes more tightly linked with advanced logic and packaging decisions.
KB also describes a roadmap extending from HBM4 mass production and HBM4E samples to an HBM5 architecture using a 1c core die and a 2nm base die, followed by a next-generation zHBM concept. The report does not establish commercial timing or customer qualification for every generation, so the roadmap should be read as a strategic path rather than a confirmed revenue schedule.
Foundry becomes part of the HBM story
The important change is that foundry is no longer separate from the HBM thesis. KB Securities estimates Samsung Foundry’s 4nm yield at above 80% and its 2nm GAA yield at above 70%, and argues that improving advanced-node performance could help both foundry economics and HBM base-die competitiveness.
Those yield figures are broker estimates and are not independently verified in the selected evidence. Even so, the strategic link is straightforward: as HBM uses more advanced logic in its base die, Samsung’s ability to manufacture both memory and logic creates a path to capture more of the value inside an AI package.
Custom HBM raises the value of integration
Mirae Asset Securities makes the same point from a more technical angle. Its report argues that memory bandwidth and I/O power are becoming critical constraints as AI accelerators support longer context windows and higher throughput. One ASIC example in the report uses a 16-GiB HBM4 configuration with 5.4 TB/s of bandwidth.
The broker also points to custom HBM and zHBM as architectures designed to reduce controller-die area, lower I/O power and increase bandwidth. These examples do not prove that every AI system will move in the same direction, but they show why memory design is becoming more deeply integrated with accelerator and package design.
The earnings case still depends on the memory cycle
The broader stack thesis does not remove the industry’s cyclical risk. Mirae Asset forecasts DRAM average selling prices to rise 277.1% in 2026 and another 22.6% in 2027. The extraordinary 2026 increase reflects a sharp comparison base and should not be treated as a normal long-term growth rate.
That pricing assumption is a major reason the broker’s earnings outlook is so strong. Mirae Asset forecasts Samsung operating profit at about $266.6 billion in 2026 and $385.9 billion in 2027, based on KRW forecasts converted using the article’s single-date reference FX rate. These are broker forecasts, not company guidance, and they are highly sensitive to memory prices, product mix and exchange rates.
A broader AI exposure, but execution still matters
The emerging case for Samsung is therefore not simply that HBM demand is rising. It is that the company could participate in more layers of the AI hardware stack at once: memory cells, advanced base dies, foundry processes and packaging. If those layers become more customized and tightly integrated, Samsung’s breadth could become more strategically important.
The remaining uncertainties are equally important. The selected reports do not independently verify the cited foundry yields, do not confirm customer qualification for every HBM generation, and do not establish a company-approved commercial timeline for zHBM. The opportunity is broader than memory alone, but the commercial outcome still depends on technology execution and a memory market that remains unusually strong.
