SK hynix Ships 12-Layer HBM4E Samples to AI Customers, Heats Up Battle with Samsung

• Performance: A data processing speed of up to 16 gigabits per second (Gbps) per pin.
• Capacity: 48 gigabytes (GB) per stack, achieved using 12 layers of 32 gigabit (Gb) die. This 32 Gb single-die capacity represents a 33% increase in storage density over the previous HBM4 generation .
• Power Efficiency: An improvement of more than 20 percent compared to the previous HBM4 generation .
• Packaging Technology: The company applied its proprietary Advanced MR-MUF (Mass Reflow Molded Underfill) process. A key benefit is a 17% reduction in heat resistance, which enhances structural stability and thermal performance in high-performance computing environments .
• Bandwidth: While SK hynix described targeting a ~4 TB/s class for next-gen AI chips, the Computex showcase highlighted a 38% bandwidth boost over HBM4 . Industry reports have also cited a bandwidth of up to 4.0 TB/s .

The chip is fabricated using the 1c DRAM process node, according to an industry report, likely in conjunction with TSMC's 3-nanometer process for the base die .

SK hynix vs. Samsung: A HBM4E Specification Showdown

With both companies now sampling their 12-layer HBM4E products, a direct comparison of their claimed specifications highlights a clear competitive dynamic. Samsung was first to ship, but SK hynix’s initial official specs claim an edge in several key performance metrics .

A critical differentiator is base pin speed. SK hynix’s HBM4E operates natively at 16 Gbps, while Samsung’s initial specification is 14 Gbps, with a stated capability to scale to 16 Gbps . Similarly, SK hynix claims a 20%+ power efficiency improvement, compared to a 16% gain for Samsung's HBM4E . The measurable 17% reduction in heat resistance via Advanced MR-MUF could be a major factor for hyperscale customers packing these modules tightly into AI server racks.

The Broader Battle for AI Memory Dominance

The HBM4E sample shipments are the latest chapter in a multi-year struggle to supply Nvidia, the world’s most valuable AI chip designer.

• SK hynix's HBM4 Stronghold: In late January 2026, multiple sources reported that SK hynix had secured approximately 70% of Nvidia’s HBM4 orders for the Vera Rubin AI platform . This commanding lead over Samsung and Micron was built on its early volume production and high yields, particularly with its MR-MUF technology .
• Nvidia Confirms All Three: On June 5, 2026, Nvidia CEO Jensen Huang formally certified all three major memory makers—SK hynix, Samsung, and Micron—as HBM4 suppliers for Vera Rubin . This ended months of speculation that Micron might be excluded entirely from the platform’s initial supply chain . Industry analysts still estimate SK hynix holds a 60–70% volume share for Vera Rubin's HBM4, with Samsung capturing 25–30% and Micron filling the remainder .
• A Pattern of Early Sampling: SK hynix established a pattern of aggressive timelines by shipping the world’s first 12-layer HBM4 samples to customers in March 2025, ahead of its original schedule . That base HBM4 product is now in mass production and is foundational to the HBM4E follow-on revealed at Computex 2026 . At the show, SK hynix highlighted a 33% density increase and 38% bandwidth jump for HBM4E, signaling its roadmap directly to the AI data center market .

What Comes Next

With samples in the hands of major customers—which typically include Nvidia, AMD, and large cloud service providers—the critical qualification process is underway . Customers will rigorously test the memory for signal integrity, thermal behavior, and compatibility with their accelerator designs. Success in this stage will determine final purchase orders and volume ramps for AI systems expected to debut in the 2027 timeframe.

SK hynix’s ability to deliver its HBM4E samples ahead of its original second-half 2026 guidance demonstrates its intent to not just defend its market lead, but to set the pace for the entire industry . The battle is no longer just about who can ship first, but whose technology can deliver the most performance-per-watt in the most stable package—directly fueling the next leap in AI capabilities.