What Is HBM? High-Bandwidth Memory Explained, and Why Korea Dominates It

High-bandwidth memory (HBM) is a type of DRAM built by stacking memory chips on top of each other and placing the stack right next to an AI processor, so data can move between them extremely fast. It is the reason a single AI accelerator can feed on enormous models — and it is also why two Korean companies, SK hynix and Samsung Electronics, have become central to the global AI boom.

If you’ve wondered why Korean stocks swing every time US tech companies talk about data centers, HBM is most of the answer.

What is HBM, in plain English?

Every processor needs memory to hold the data it is working on. Ordinary memory modules sit some distance from the processor and talk to it through a relatively narrow connection. For AI, that becomes a bottleneck: the processor can calculate faster than memory can feed it.

HBM fixes this in two ways. First, it stacks DRAM dies — thin slices of memory chip — vertically, like floors of a building. SK hynix’s own SEC prospectus describes HBM as a product that “vertically interconnects multiple DRAM chips.” Second, the stack is mounted right next to the GPU in the same package, connected by thousands of tiny wires instead of a few dozen.

The result is a much wider “highway” for data. A standard memory chip is like a two-lane road; HBM is a highway with more than a thousand lanes.

How are the chips stacked together?

The key manufacturing technique is the through-silicon via (TSV) — microscopic vertical holes drilled through each die and filled with metal, so electrical signals can pass straight up and down the stack. SK hynix’s prospectus calls TSV critical to next-generation HBM development.

Stacks are described by height: an “8-Hi” or “12-Hi” stack has 8 or 12 DRAM dies. More layers mean more capacity in the same footprint, but also more difficulty — every die must be thinned, aligned, and bonded without defects. That is why HBM is far harder to make than conventional DRAM and why yields (the share of usable chips) are a constant topic in industry news.

HBM3E vs. HBM4: what changed?

HBM has gone through several generations. Most HBM revenue in 2026 still comes from HBM3E, according to Counterpoint, but HBM4 — the sixth generation — began shipping in volume this year for Nvidia’s next platform, Vera Rubin.

Sources: JEDEC HBM4 standard (JESD270-4) as reported by Tom’s Hardware; Samsung figures as reported by TrendForce. The HBM3E speed is the figure TrendForce uses for comparison.

The big change is the much wider 2,048-bit interface, which helps HBM4 double the bandwidth of HBM3. Running 2,048 data lines at 8 Gb/s each works out to roughly 2 terabytes per second from a single stack. Vendors are also competing to run faster than the standard: TrendForce reports Samsung’s HBM4 reaches 11.7 Gb/s per pin, about 37% above the JEDEC baseline.

Who makes HBM, and who is winning?

Only three companies produce HBM in volume. SK hynix moved first and led for years; Samsung has been catching up fast in 2026.

Source: Counterpoint Research, as reported by EDAILY (September 2026).

Market-share figures vary by research house and method. IDC data cited in SK hynix’s SEC filing puts SK hynix’s share at 56.4% in Q1 2026, lower than Counterpoint’s 58% for the same quarter. Either way, the direction is clear: SK hynix still leads, but Samsung’s HBM4 ramp has narrowed the gap, and Counterpoint expects Samsung’s share to keep rising.

Why does HBM matter so much to Korea?

Two of the three suppliers are Korean, which makes HBM one of the most concentrated — and valuable — industrial positions any country holds in the AI supply chain.

• Stock market weight. Samsung and SK hynix together make up roughly half of the KOSPI’s market capitalization, per Korea JoongAng Daily. When AI sentiment shifts, the whole index moves.
• Exports. Memory chips have driven Korea’s record export run this year; see our breakdown of October’s export surge.
• US customers. SK hynix’s prospectus shows its US sales subsidiaries generated 64.7% of revenue in Q1 2026, and its two largest customers accounted for 14.8% and 12.4%. DRAM, including HBM, made up 77.3% of sales.
• Global capital. SK hynix listed American depositary shares on Nasdaq in July 2026 (ticker SKHY), raising about $26.5 billion — partly to fund more capacity. Our guide on how to invest in Korean stocks from the US explains how the ADRs work.

What are the risks?

HBM’s strength is also its weakness: demand comes overwhelmingly from AI data center spending by a handful of large customers.

SK hynix’s own risk factors list the main ones: a slowdown in AI infrastructure spending by hyperscale cloud providers; new technology that reduces memory needs; customer concentration; and the complexity of HBM production, which can constrain supply.

Policy is a newer risk. On November 5, 2026, SK hynix fell 9.46% and Samsung 7.57% after the US midterm elections raised fears that tighter data center rules could slow AI construction — even though no new rule has been announced. We covered that selloff in today’s KOSPI brief.

There is also the memory cycle. Memory has historically been a boom-and-bust business: when supply catches up with demand, prices can fall quickly. HBM is harder to make than regular DRAM, which has slowed that catch-up, but competition is intensifying as Samsung and Micron ramp HBM4.

FAQ

Is HBM the same as DRAM? HBM is DRAM — the same basic memory cell — packaged differently. The stacking, TSV wiring, and placement next to the processor are what make it “high-bandwidth.”

Why can’t AI chips just use regular memory? They can, but they would sit idle waiting for data. AI training and inference move vast amounts of data, so memory bandwidth often limits performance more than raw computing power.

Does Nvidia make HBM? No. Nvidia designs the GPUs and buys HBM from SK hynix, Samsung, and Micron, then packages the memory next to its processors.

What comes after HBM4? One direction is “custom HBM,” where a customer’s logic is built into the stack’s base die. SK hynix’s prospectus defines custom HBM as HBM with GPU or ASIC functions built into the base die.

Is it a good time to buy HBM makers’ stocks? We don’t give investment advice. What we can say is that these companies combine high growth with high concentration risk, and their share prices react sharply to AI spending signals — as November 5 showed.

Decision guide: how to read HBM headlines

• “Share gains” headlines — check the source (Counterpoint, IDC, TrendForce) and whether it measures revenue or shipments; numbers differ.
• “Qualification” or “certification” headlines — these mean a supplier passed a customer’s tests, which usually comes before volume orders.
• Data center spending news — the most direct driver of HBM demand, and therefore of Korean chip stocks.
• Memory price data — rising contract prices boost profits; a plateau is an early warning sign.