Samsung ships first 12-layer HBM4E samples to major AI customers

Samsung Electronics has begun shipping the first 12-layer HBM4E samples to major global customers, an early move in the fight to supply the memory that underpins AI servers and accelerators. The new chip is Samsung’s latest attempt to close the gap in a market where being first to qualify with cloud and chipmakers can decide who wins the next wave of demand from AI builders.

Samsung said the samples reach speeds of up to 16 gigabits per second and are built with its 1c DRAM process technology and a 4-nanometer logic base die. The company said the design improves energy efficiency and thermal performance, two of the most important constraints in dense AI systems where memory bandwidth, heat and power all shape how much work a data center can do. The 12-layer package is meant to push more performance into the same footprint, a key advantage as AI servers become more tightly packed with high-end accelerators.

The shipment also signals a sharper offensive from Samsung in a memory race it entered late in earlier generations. The company missed the earliest momentum in HBM3 and HBM3E, limiting the number of orders it could capture while rivals moved ahead. This time, Samsung is trying to get in front of customers sooner, with major names including AMD, Nvidia and Google among those receiving samples. Samsung had said in April that it planned to ship first samples in the second quarter, and this move shows it met that timetable.

Investor reaction was immediate. Samsung shares rose as much as 6.5% in morning trading, while SK Hynix also advanced, with its stock up about 1.2% at 0207 GMT. The market response reflected how closely traders now watch memory qualification cycles, since sample shipments often mark the beginning of a long process that can end in design wins, volume orders and a stronger position inside the AI hardware stack.

The timing matters because Samsung is already moving fast into the next generation. The company began mass production of HBM4 in February 2026 and said at NVIDIA GTC 2026 that its HBM4 family is designed for NVIDIA’s Vera Rubin platform. SK Hynix, meanwhile, has said it plans to supply HBM4E samples in the second half of 2026 and start mass production in 2027. That leaves Samsung with an opening: not just to ship first, but to shape the next round of AI memory demand before rivals can lock it down.