Samsung’s HBM position is improving fast. The harder question is whether a richer AI-memory mix can cushion a broader DRAM cycle that brokers see very differently in 2027.
Samsung Electronics is heading into the next phase of the memory cycle with two stories moving at the same time.
The first is structural: high-bandwidth memory, or HBM, is becoming a larger part of Samsung’s AI-memory business. HBM stacks multiple DRAM dies to feed AI accelerators with much more data than conventional memory can deliver. As AI chips become more demanding, the value of HBM increasingly depends not only on the memory dies themselves, but also on the base die, logic design and advanced packaging around them.
The second story is cyclical: conventional DRAM pricing may not stay as strong as it is in 2026. That matters because even a better HBM mix does not make Samsung independent of the broader memory market.
HBM is becoming a bigger part of Samsung’s memory mix
KB Securities estimates Samsung’s HBM market share at 33% in the second quarter of 2026 and expects it to approach roughly 40% in the fourth quarter. The same report expects third-quarter HBM4 revenue to rise to more than three times the previous quarter and HBM4 to represent more than 60% of Samsung’s HBM revenue in the second half of 2026.
Those are forecasts rather than reported outcomes, but they illustrate why Samsung’s HBM trajectory matters. A larger HBM position could improve the mix of its memory business even if growth in ordinary DRAM becomes less uniform.
The product roadmap also points to more integration. Mirae Asset Securities highlights custom HBM and Samsung’s zHBM concept as examples of how memory, logic and packaging are becoming more closely linked. In simple terms, future HBM products may compete on how well the memory stack works with the logic underneath it, not only on the amount of DRAM capacity supplied.
That does not guarantee share gains. zHBM remains a forward-looking architecture concept, and a more integrated design is not the same thing as proven volume production. Still, the direction makes HBM mix more strategically important than a simple commodity-memory price view would suggest.
The real disagreement starts in 2027
The most useful contrast in the current research is that the brokers largely agree on the strength of the 2026 DRAM pricing surge, then split sharply on what happens next.
Mirae Asset forecasts Samsung’s annual DRAM average selling price, or ASP, to rise 277.1% in 2026 and another 22.6% in 2027. BNK Investment & Securities is similarly strong on 2026, forecasting a 274% increase, but expects DRAM ASP to fall 8% in 2027.
That difference is central to the HBM question. If DRAM pricing remains firm into 2027, HBM growth would add to an already supportive memory backdrop. If conventional DRAM prices roll over, HBM would have to do more of the work.
The comparison should not be treated as a controlled, apples-to-apples model test. Each broker uses its own assumptions and methodology. The value of the comparison is the direction of the disagreement: both models see a powerful 2026 cycle, but they do not agree on how durable it is.
BNK’s more cautious view is already visible in its near-term revisions. The broker reduced its third-quarter 2026 DRAM ASP growth assumption from 18% quarter on quarter to 14% as part of a broader earnings estimate cut. It also argues that memory demand is becoming more price-sensitive as system memory costs rise, while suppliers continue to expand capacity.
AI demand is strong, but demand elasticity is the swing factor
The bullish case still has a large demand engine behind it. KB Securities estimates China’s AI data-center capacity will rise from 40 gigawatts in 2026 to 80 gigawatts in 2030. That is a broker estimate, not an official capacity forecast, but it captures the scale of the infrastructure build-out that underpins the AI-memory thesis.
More AI infrastructure usually means more accelerators, and those accelerators require more memory bandwidth. That supports HBM demand directly and can also lift server DRAM content.
The counterargument is about demand elasticity: how buyers react as memory becomes more expensive. BNK argues that rising system memory costs and a stronger focus on AI efficiency could make customers more sensitive to price, especially if model providers prioritize lower inference costs over raw performance growth.
Those two forces can coexist. AI infrastructure can keep expanding while buyers become more selective about how much memory they deploy per unit of compute. The result would be a market in which HBM remains structurally attractive even as conventional memory pricing becomes more cyclical.
What matters most is mix versus cycle
The evidence in these reports does not support a precise calculation of how much HBM growth would be needed to offset weaker conventional DRAM or NAND. The selected research does not provide one directly comparable HBM revenue series that can be netted against the rest of Samsung’s memory business.
What it does show is a clear analytical split. Samsung’s HBM position is expected to improve through the second half of 2026, and the product roadmap is moving toward more customized and integrated architectures. At the same time, the durability of the broader DRAM cycle is becoming more uncertain as the market moves toward 2027.
That makes the next phase less about whether AI memory is growing and more about whether HBM can grow fast enough to change the mix before conventional memory pricing normalizes.
