SK Hynix ships first 12-layer HBM4E samples to AI customers

SK Hynix said on Thursday that it had begun shipping samples of HBM4E, its next generation of high-bandwidth memory for AI, to major customers. The number that matters is the one the chipmaker is leading with: a 12-layer stack reaching 48GB of capacity, running at up to 16Gbps per pin, with power efficiency improved by more than 20 per cent over the previous generation.

The announcement is a milestone in a race SK Hynix has been winning. High-bandwidth memory is the component that sits beside an AI accelerator and feeds it data fast enough to keep up, and it has become one of the tightest bottlenecks in the entire AI supply chain.

The South Korean company has spent the current boom as the supplier most closely tied to Nvidia’s memory demand, and shipping HBM4E samples on schedule is its way of signalling that the lead is intact.

The technical claims are specific. SK Hynix said it used its Advanced MR-MUF packaging to achieve the 48GB, 12-layer build while keeping the stack structurally stable, and that heat resistance is up 17 per cent on the preceding HBM4, which matters for keeping memory running reliably inside dense, hot data-centre racks.

The company said it delivered the 12-stack samples on schedule and would work with partners towards mass production “in a timely manner”, its phrasing rather than a fixed date.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!

Investors read it as a win. SK Hynix shares rose on the news, with one report putting the move at close to 5 per cent and a record high. The reaction reflects how much of the company’s valuation now rests on staying first to each new memory node, in a market where being early to qualify with Nvidia can decide years of supply contracts.

That market is unusually concentrated. Three firms, SK Hynix, Samsung, and Micron, control the overwhelming majority of DRAM production, and all three are racing to convert capacity towards HBM as quickly as they can.

Nvidia has certified all three for HBM4, but analyst estimates have repeatedly put SK Hynix well ahead on share for the coming generation. Shipping HBM4E samples first is the kind of move that keeps those estimates where they are.

The broader context is a memory market that has gone from glut to scarcity. AI infrastructure spending has pulled an ever larger share of memory output towards data centres, leaving conventional buyers competing for what is left and pushing prices sharply higher.

HBM4E is built for exactly the customers driving that shift: the operators filling buildings with accelerators and needing memory that can keep them fed without overheating.

That scarcity is now reaching ordinary buyers. The same reallocation of capacity towards HBM has helped trigger a broad shortage of conventional memory, the kind that goes into phones and laptops, and it has begun to show up in retail prices.

Apple chief executive Tim Cook has said the company can no longer absorb the cost and that price increases are coming, a downstream consequence of exactly the demand SK Hynix is racing to serve at the top of the chain.

SK Hynix’s position rests on a tightly interlocked set of partnerships. It has a technical alliance with TSMC, which supplies advanced logic and packaging, while Nvidia supplies the accelerator demand and platform direction the memory is designed around.

That arrangement, more than any single specification, is what has kept the company ahead: being inside the design conversation for the next platform, rather than bidding for it afterwards.

The surge driving HBM4E reaches well beyond the giant training clusters. As inference moves towards the edge, into devices and on-premise hardware, chipmakers are building for smaller, power-constrained footprints, a shift visible in parts like Nvidia’s edge-focused models. HBM4E sits at the opposite, data-centre end of that spectrum, but the same AI workloads pull the whole memory market along.

What remains unsettled is timing. Samples are not mass production, and SK Hynix did not commit to a date for the latter. The company has said only that it will move with partners as readiness allows. For now, the sample shipment is the claim on the table, and it is a strong one: first out with the next node, in the part of the AI supply chain where being first is worth the most.