SK Hynix Pushes HBM4 Into Volume as Samsung Scrambles to Close the Gap
The Icheon chipmaker says its first HBM4 stacks have cleared qualification at a marquee US accelerator customer, hardening a lead that has become the most lucrative bottleneck in the AI build-out.
- ·Capital rotates out of US/EU equities into hard ASEAN infrastructure.
- ·Data centres, power transmission and ports are the three priority lanes.
- ·Vietnam, Indonesia and the Philippines absorb the largest allocations.
ICHEON, South Korea — SK Hynix said on Wednesday that its first 12-layer HBM4 memory stacks had passed qualification at a leading American accelerator maker, clearing the way for volume shipments in the third quarter and extending the company's grip on the single scarcest input in the artificial-intelligence supply chain.
The announcement, made at a briefing on the sidelines of the firm's Icheon campus, lands at a moment when the price of high-bandwidth memory has become the closest thing the industry has to a tax on every frontier model trained this year. Executives put the qualified part at 2.0 terabytes per second of bandwidth per stack, a roughly 40 percent jump over the HBM3E generation now sold out through 2027.
For a product that did not exist commercially three years ago, HBM has become the line item that decides who can build at the frontier and who waits in a queue. SK Hynix now derives more operating profit from a few thousand square metres of advanced-packaging floor than from the entire commodity DRAM business that defined it for a generation.
What qualification means
Qualification is the unglamorous gate that separates a press release from revenue. It means the customer has run the part through months of thermal cycling, signal-integrity testing and burn-in, and accepted it for a named product. In HBM, where a single stack bonds a dozen wafer-thin dies onto a logic base, a yield miss of a few percentage points can erase a quarter's margin.
SK Hynix declined to name the customer, but two people briefed on the roadmap said the part is destined for a 2027 training accelerator and that initial allocations are already spoken for. The company said yields on the qualified line were 'comfortably above' its internal target, without giving a figure.
HBM4 is not a faster version of the same thing. The base die is now effectively a custom logic chip, and that changes who you have to partner with to build it, said Jae-won Sohn, a memory analyst at Hangang Securities in Seoul.
That shift — from a memory part to a co-designed logic-and-memory assembly — is why HBM4 has scrambled the competitive order. The base die that sits beneath the memory stack is now fabricated on a leading-edge foundry node, pulling TSMC into a market that used to belong wholly to the three memory makers.
Samsung's catch-up problem
Samsung Electronics, the world's largest memory maker by volume, has spent two years trying to close a gap it did not expect to open. Its HBM3E parts were slow to qualify at the same marquee customer, and people familiar with the effort say its HBM4 samples are still working through thermal qualification rather than clearing it.
The stakes are measured in tens of billions of dollars. The accelerator market is supply-constrained on memory, not logic; every additional qualified HBM source loosens a bottleneck that has held back datacentre operators from Singapore to Virginia. A second qualified vendor would also give buyers pricing leverage they currently lack.
Samsung said it remained 'on track' for HBM4 mass production in the second half of 2026 and pointed to its in-house foundry as an advantage in co-designing the base die. Micron, the third player, has guided to a similar timeline and has quietly won allocation at a hyperscaler building its own silicon.
The packaging chokepoint
Even with stacks qualified, the constraint moves downstream to advanced packaging — the through-silicon vias and hybrid bonding that fuse memory to compute. SK Hynix is spending to add hybrid-bonding capacity at Icheon and at a new line near Cheongju, but the equipment lead times run past a year.
That is the quiet reason this announcement matters beyond Korea. Every gigawatt of AI datacentre being financed across Asia this year assumes a memory supply curve that bends the right way. SK Hynix just signalled that, for now, it bends in its favour.
Whether that holds depends on a customer the company will not name and a rival it cannot afford to underestimate. For the next several quarters, though, the most valuable bottleneck in computing has a single dominant gatekeeper — and it sits an hour southeast of Seoul.
