What Company Makes HBM? The 3 Chipmakers Behind AI Memory
Photo by TECNIC Bioprocess Solutions on Unsplash
Related video: "SK Hynix's AI Memory Monopoly: The $47B HBM Market No One Talks About" (via Asia Tech Macro)
The short answer
If you've read that "Nvidia can't get enough HBM" and wondered who actually manufactures the stuff, here it is: SK hynix, Samsung Electronics, and Micron Technology. That's the complete list. There is no fourth supplier waiting in the wings, and there hasn't been for years.
The reason the list is so short becomes obvious once you know it. HBM is a type of DRAM — the working memory inside computers — and after three decades of consolidation, those same three companies are the only ones left that can make leading-edge DRAM at scale. The HBM list is just the DRAM list. Two of the three are Korean, and their main memory campuses sit a short drive apart in Gyeonggi Province, the ring of cities around Seoul.
What HBM actually is, in plain English
HBM stands for High Bandwidth Memory. It's a standardized memory format defined by JEDEC, the industry body that writes memory specifications.
Ordinary computer memory (DDR5) lives on modules a few centimetres from the processor and talks to it over a relatively narrow electrical path. HBM does something different: it stacks DRAM chips vertically, like a tower of pancakes, and drills electrical channels straight down through them — a technique called through-silicon vias, or TSV. That tower is then placed right beside the GPU die inside the same package, connected by a slab of silicon called an interposer.
The payoff is a very wide, very short data path, which means enormous bandwidth per watt. The cost is that HBM is far harder to build and far more expensive per gigabyte than normal memory.
Why does AI need it? Running a large language model means constantly streaming billions of parameters through the GPU's compute units. These workloads are typically limited by memory bandwidth rather than raw math. A GPU that can't be fed fast enough simply sits idle — which is how memory, of all things, became the bottleneck of the AI boom.
The generations
HBM → HBM2 → HBM2E → HBM3 → HBM3E → HBM4. Each step roughly doubles bandwidth per stack and adds height (8-high, 12-high, 16-high stacks). HBM4 is the generation tied to Nvidia's platforms after Blackwell, and it's where the current competitive fight is happening.
Who supplies Nvidia — the honest version
This is the part people search for, so it's worth being careful, because the answer moves.
SK hynix has been the lead HBM supplier to Nvidia since the HBM3 generation and is widely reported to hold the largest share of the market. Micron arrived later, qualified for HBM3E, and serves as a second source — notable partly because it's the only non-Korean option, which matters a great deal to US policymakers. Samsung Electronics is the world's biggest DRAM maker overall, but was conspicuously late to qualify HBM3E with Nvidia, a turn that became a long-running story in the Korean business press and a source of visible soul-searching inside the company.
One nuance that most coverage skips: "Nvidia's supplier" isn't a single fixed slot. Nvidia qualifies vendors per product, per generation, per stack height. A company can be approved on the 8-high stack and shut out of the 12-high one. So treat any snapshot of "who's winning" as exactly that — a snapshot.
Also worth knowing: HBM isn't sold like a commodity. Industry reporting suggests it is generally committed under long-horizon agreements, effectively pre-sold a year or more ahead, and often co-designed with the customer. That's close to the opposite of how DRAM traditionally worked, and it's a big reason the economics look so different now.
Why one Korean company ended up on top
1. SK hynix got there first and refused to quit
HBM was co-developed by SK hynix with AMD, and the first commercial HBM product shipped in the mid-2010s inside an AMD Radeon graphics card. For most of the following decade, HBM was a niche curiosity — useful for high-end graphics and supercomputers, irrelevant to the mass market, and not especially profitable.
SK hynix kept a team on it anyway. When ChatGPT landed and Nvidia's data-centre demand exploded, SK hynix was the one company with a mature, yield-proven HBM line and an existing relationship. Its lead is largely the payoff of a decade of patience on an unglamorous product.
2. It's a packaging fight, not a lithography fight
The intuitive assumption is that whoever makes the smallest transistors wins. For HBM, that's wrong. The hard part is bonding and stacking: aligning eight to sixteen paper-thin chips, connecting them through TSVs, and keeping the whole tower from cooking itself — all at an acceptable yield.
SK hynix's signature method is MR-MUF (mass reflow–molded underfill), which bonds the whole stack at once and fills the gaps with a heat-conducting molding compound. Samsung pursued TC-NCF (thermo-compression non-conductive film), bonding layer by layer. The broad industry view is that MR-MUF gave SK hynix an edge in yield and heat dissipation as stacks grew taller.
3. Yield is the whole business
An HBM stack contains many DRAM dies plus a logic base die at the bottom. One bad die can scrap the entire stack, so defect rates compound with height. That's why falling a year behind is so painful — you don't just need the design, you need years of accumulated manufacturing learning.
It also explains two side effects you may have felt in your wallet. HBM reportedly carries margins memory makers have rarely, if ever, seen before. And HBM production consumes wafer capacity that would otherwise produce ordinary DRAM, a squeeze widely cited as one reason conventional memory prices have been under pressure.
4. Korea's memory ecosystem made it possible at all
Here's the backstory outsiders rarely get. Korean conglomerates — the chaebol, large family-controlled business groups such as Samsung, SK, LG and Hyundai — pushed into DRAM in the 1980s. They survived the savage price wars of the 1990s and 2000s that wiped out nearly every Japanese, American and European memory maker, and emerged as two of the three survivors.
SK hynix is itself a product of that shakeout. It began life as Hyundai Electronics, was restructured into Hynix, spent years under creditor control after the 1997 Asian financial crisis, and was acquired by SK Group in 2012. A company that nearly died twice is now one of the most strategically important suppliers in the entire AI industry.
The three makers, side by side
| SK hynix | Samsung Electronics | Micron | |
|---|---|---|---|
| Country | South Korea | South Korea | United States |
| Position in HBM | Market leader since HBM3 | Largest DRAM maker overall, HBM laggard so far | Second source, capacity-constrained |
| Signature stacking process | MR-MUF | TC-NCF | Its own TSV stacking process |
| Structural advantage | First mover, yield, Nvidia relationship | Only firm doing memory + foundry + packaging in-house | Non-Korean supply, US fabs, policy favour |
| Structural weakness | Heavy customer concentration | Must displace an entrenched incumbent | Smallest capacity base |
The interesting line in that table is Samsung's. It is the only company that can make the DRAM, fabricate the logic base die in its own foundry (a foundry is a factory that builds chips designed by other companies), and package the finished stack — one vendor, one purchase order. SK hynix has no foundry of its own, so for HBM4-era base dies it works with TSMC of Taiwan.
This matters more from HBM4 onward, because that generation moves the logic base die off memory processes and onto an advanced logic node at a foundry. Samsung builds its base die on its own 4nm foundry process — making it the only HBM4 supplier running memory, foundry and packaging as a single turnkey operation — while SK hynix outsources to TSMC's 12nm process today, is weighing a shift to TSMC 3nm from HBM4E, and is also looking at Intel Foundry for some supply-chain diversification.
Whether Samsung's vertical integration ever converts into actual market share is, in my view, the single most important open question in the memory industry. On paper it's a compelling pitch. In practice, customers buy proven yield — and proven yield is exactly what Samsung has been fighting to demonstrate.
Three things people get wrong
- "Nvidia makes its own memory." No. Nvidia designs GPUs and buys HBM from the three makers. It's a customer.
- "TSMC makes HBM." No. TSMC manufactures the GPU die and performs the advanced packaging (its CoWoS process) that joins the GPU and the HBM stacks together. Different job entirely.
- "HBM is just faster RAM you can buy." No. You can't buy an HBM module at retail. It's sold into packaged accelerators under long-term contracts, and supply is generally spoken for well in advance.
What to watch next
Two things. First, HBM4 qualifications — watch which vendors get approved at which stack heights, because that's where the next share shift will show up. Second, whether the industry moves from today's bonding methods toward hybrid bonding as stacks get taller; that transition could reshuffle the yield advantage that currently defines the leader.
And if you take one thing away: the AI boom's scarcest physical input isn't the GPU. It's the memory bolted next to it — and most of it is made in Gyeonggi Province.
FAQ
Is SK hynix the only company that makes HBM for Nvidia?
No, but it has been the lead supplier since the HBM3 generation. Micron qualified as a second source for HBM3E. Samsung finally cleared Nvidia's HBM3E 12-Hi quality certification in September 2025 — the third supplier to do so, after SK hynix and Micron — then moved into mass production, began actual shipments in October 2025, and has already sold out its 2026 supply. Nvidia approves suppliers per product and per stack height, so the lineup can differ between one accelerator and the next.
Why is HBM so expensive compared to normal RAM?
Because it's built by stacking 8 to 16 ultra-thin memory chips and wiring them together vertically through the silicon. A single defective die can ruin the whole stack, so yields are lower and harder to hold. HBM also consumes wafer capacity that would otherwise produce conventional DRAM, tightening supply across the board.
Could a Chinese company start making HBM?
Chinese memory makers have publicly stated ambitions here, but leading-edge HBM requires both advanced DRAM manufacturing and extremely difficult stacking and packaging expertise — plus equipment that is subject to export controls. Closing that gap is widely seen as a multi-year problem, which is a big part of why Korean supply is treated as strategically critical.
Comments
Post a Comment