SK hynix and Samsung Race to Ship 12-Layer HBM4E

The next AI-memory war is being fought in sample shipments. Samsung shipped the industry-first 12-layer HBM4E samples to major customers on May 29, 2026, and SK hynix answered on June 18, both delivering 48GB per stack at close to 4 terabytes per second, the memory that will feed 2027 AI accelerators like Nvidia's Rubin Ultra. Neither has cleared customer qualification or started volume production, so this is a race for position, not a finished win, but the order of arrival will help decide who locks in next year's biggest orders.

• HBM4E is the next high-bandwidth memory generation, lifting per-stack capacity from HBM4's 36GB to 48GB and per-pin speed toward 16 Gbps, with 20%-plus better energy efficiency.
• Samsung shipped first on May 29 (claimed 3.6TB/s per stack); SK hynix followed June 18, ahead of its own schedule, at roughly 4TB/s.
• The target is next-gen AI silicon: Nvidia Rubin Ultra is slated for 384GB of HBM per GPU (eight 48GB stacks), up from 288GB on HBM4-based Vera Rubin, with AMD Instinct MI500 also in line.
• Both have only shipped samples. Qualification and volume production, the stages that decide real orders, are still ahead.

What is HBM4E and why does it matter?

High-bandwidth memory is the stacked DRAM that sits beside an AI chip's compute die and feeds it data. Modern accelerators are rarely starved for math, they are starved for bandwidth, so the memory generation often matters as much as the GPU it serves. HBM4E is the refinement of HBM4: it swaps in a denser sixth-generation DRAM core die to push per-stack capacity from 36GB to 48GB, raises per-pin speed from the 11-to-13 Gbps range toward 16 Gbps, and improves energy efficiency by more than 20 percent, all while keeping the same 2,048 input-output pins. In an era where a single training run can cost tens of millions of dollars, a 30-percent capacity jump and a 20-percent efficiency gain per stack are not incremental, they change what a rack can hold and what it costs to run.

RelatedOpenAI's First Chip, Jalapeno, Targets Cheaper Inference

Who shipped first, and does it matter?

Samsung got there first. It began shipping the industry's first 12-layer HBM4E samples on May 29, 2026, claiming a stable 14 Gbps pin speed that scales to 16 Gbps under load and a single-stack bandwidth of 3.6TB/s at 48GB per stack. SK hynix followed on June 18, ahead of the second-half timeline it had given in April, with 48GB 12-layer samples reaching up to 16 Gbps and roughly 4TB/s, built using its Advanced MR-MUF packaging with 17 percent better heat resistance than HBM4. First-to-sample matters because qualification cycles are long and customers prefer to design around memory they already have in hand. But it is not decisive: neither company has said it cleared qualification or begun volume production, and those are the gates that actually convert samples into revenue.

The share numbers explain the urgency. SK hynix led the HBM market with about 58 percent in the first quarter of 2026, but that is down from 69 percent a year earlier as Samsung and Micron closed the gap. A key technical fork sits in the base die: SK hynix is understood to use a TSMC 3nm-class base die, which it has not officially confirmed, while Samsung builds its base die on its own 4nm-class foundry process. That choice affects performance, yield and cost, and it is one reason Nvidia's endorsement still matters more than raw specs.

Who is actually affected?

The buyers are the story. HBM4E is expected to power 2027's flagship AI chips, and the marquee customer is Nvidia. Rubin Ultra is slated to carry 384GB of HBM per GPU, eight of the 48GB 12-layer parts, more than 30 percent above the 288GB on the HBM4-based Vera Rubin. AMD's Instinct MI500 is also in line. Nvidia's preference was on open display at Computex, where Jensen Huang visited SK hynix's booth and wrote "Please make more" on an HBM4E wafer, while SK Group's chairman noted the company currently had a single HBM4E customer. That single-customer detail is the whole game: whoever qualifies first for Nvidia's next platform captures the volume that funds the following generation.

RelatedIntel's Crescent Island GPU Skips HBM for 160GB of LPDDR5X

1. May 29Samsung ships first 12-layer HBM4E samples. Claims 3.6TB/s, a lead of several months.
2. Jun 2Vera Rubin enters full production. Samsung, SK hynix and Micron named HBM4 suppliers.
3. Jun 18SK hynix ships 12-layer HBM4E samples. Ahead of its own second-half timeline.
4. 2H 2026Customer qualification. The stage that decides who wins Rubin Ultra orders.
5. 2027Rubin Ultra and MI500. 384GB per GPU built on eight 48GB HBM4E stacks.

Our take

Sample races are easy to overhype, and this one deserves a cool head: shipping a first sample is a milestone, not a contract, and both Samsung and SK hynix are still weeks or months from the qualification that actually moves money. That said, the trajectory is what matters. HBM has quietly become the choke point of the entire AI buildout, and the company that consistently qualifies first for Nvidia's next platform earns a franchise that compounds for years. Samsung winning the first-sample sprint is a real signal that it is done ceding the high end to SK hynix, and SK hynix beating its own schedule shows it feels the pressure. For everyone downstream, the takeaway is simpler: the 48GB stack is what makes 384GB GPUs possible, and 2027's AI hardware is already being decided in these sample shipments today.

Original analysis by GenZTech. Figures current as of July 2026; bandwidth figures are vendor-stated. Source: skhynix.com