HBM naturally leads discussion of SK hynix, a Korean memory producer. Yet an AI server needs system memory and storage as well as stacked memory, which is why the company presents itself as a full-stack AI memory supplier.
Core idea SK hynix should be judged on its ability to commercialize a broader AI memory stack, not on HBM announcements alone.
The Core Business Is DRAM and NAND. SK hynix develops and manufactures DRAM and NAND flash. HBM is a premium DRAM product that stacks multiple dies for high bandwidth; the company cannot be treated as a stand-alone HBM business.
Growth in AI servers can support HBM, DDR5, SOCAMM-family memory, and enterprise SSD demand. Conversely, customer server spending and memory inventory adjustments can affect several product categories at once.
HBM4E Is Not Yet the Same as Production Revenue. On June 18, 2026, SK hynix announced that it had supplied major customers with samples of its 48GB, 12-layer HBM4E, rated at 16Gbps per pin. Sample delivery begins customer evaluation; it is not a large production contract with finalized volume and price.
Customer qualification, manufacturing yield, orders, and shipments must follow. Naming undisclosed customers or describing supply as exclusive would go beyond the public facts.
Heat and Power Will Shape the Next Performance Step. As HBM stack height and speed rise, heat control and signal stability become more important. SK hynix said the same HBM4E sample uses Advanced MR-MUF and lowers thermal resistance by 17% from the prior generation.
The company separately announced the start of SOCAMM2 mass production on April 19, 2026. That puts a non-HBM AI system-memory product at the production stage; it does not establish HBM4E qualification or mass production.
NAND Is Also Part of AI Infrastructure. Data created during training and inference must be stored and read again. The company has presented 321-layer QLC NAND-based cSSD supply and higher-capacity eSSD expansion as parts of its AI storage strategy.
QLC increases storage per cell but must balance performance and endurance for its use case. Starting supply of a product does not guarantee profitability across NAND; shipment volume, product mix, and inventory still matter.
Trace the Connection from Technology to Cash. Product development, sample supply, qualification, production readiness, and shipment are separate stages. Labeling the stage in each announcement reduces the risk that expectations run ahead of revenue.
In quarterly reporting, follow HBM and server-memory shipments, NAND mix, capital spending, and cash flow. Customer-level shares and market shares should remain unstated when the company has not disclosed them.
Testing the idea that HBM is only half the picture. Suppose hypothetical quarterly memory revenue of KRW 10 trillion consists of KRW 4 trillion from HBM, KRW 4 trillion from conventional DRAM, and KRW 2 trillion from NAND. Even if HBM grows 50%, a 20% price decline across the other KRW 6 trillion leaves total revenue only KRW 0.8 trillion higher.
If product-level revenue is not disclosed, or if categories use different pricing bases or include intersegment transactions, stop estimating shares and compare only disclosed DRAM and NAND revenue and shipment volumes.
HBM shipment growth must be placed in the same table as conventional-memory prices and inventory to avoid seeing only half of the company.
Check your understanding
- Did you distinguish HBM from the company’s overall memory business?
- Did you keep sample delivery separate from production shipments?
- Did you avoid guessing undisclosed customers and supply shares?
- Did you consider DRAM and NAND investment together with cash flow?
Verification date: July 24, 2026. Checked against official company releases, investor materials, and issuer disclosures. Undisclosed customers, volumes, prices, market shares, and forecasts are not presented as facts. This article is not investment advice. This English article is a translated learning resource, not investment advice.