SK Hynix, already the leading supplier in the high-bandwidth memory (HBM) market, was first to provide HBM3E chips to Nvidia in 2024
In sum – what to know:
HBM4 development completed – SK Hynix finalized the industry’s first HBM4 chip and is preparing for mass production, extending its lead in AI memory technology.
Performance and efficiency gains – HBM4 doubles bandwidth, improves power efficiency by 40%, and could boost AI service performance by nearly 70% while reducing energy demand in data centers.
Competitive race ahead – Analysts see initial HBM4 pricing 60–70% higher than HBM3E, with Samsung and Micron expected to bring rival products to market soon.
Korean company SK Hynix announced it has completed development of what it claims to be the world’s first HBM4 chip for artificial intelligence systems and is prepared to begin mass production.
The Korean company, already the leading supplier in the high-bandwidth memory (HBM) market, was first to provide HBM3E chips to Nvidia in 2024. Analysts say the debut of HBM4 strengthens SK Hynix’s position in the AI memory sector as rivals prepare to follow.
“Completing HBM4 development sets a new milestone for the industry,” said Cho Joo-hwan, head of HBM development at SK Hynix. “By delivering performance, power efficiency, and reliability that match customer needs, we will ensure timely supply and maintain our competitiveness.”
HBM technology stacks DRAM chips vertically to enable much faster data transfer than conventional memory, critical for AI servers and other compute-intensive workloads. Nvidia is expected to integrate eight of SK Hynix’s 12-layer HBM4 chips in its upcoming Rubin GPU platform, slated for the second half of 2026, according to press reports.
According to the Korean company, HBM4 doubles bandwidth with 2,048 input/output connections and achieves speeds above 10 Gbps, outperforming the JEDEC standard of 8 Gbps. Power efficiency improves more than 40% compared to the prior generation, potentially lifting AI service performance by up to 69% while cutting energy consumption in data centers.
For mass production, SK Hynix has applied its advanced MR-MUF stacking process and its fifth-generation 1b 10nm technology node, designed to minimize risks and improve heat dissipation. “We are unveiling the world’s first mass production system for HBM4,” said Kim Ju-seon, president and head of AI Infra at SK Hynix. “HBM4 marks a symbolic turning point beyond AI infrastructure limits.”
Analysts expect HBM4 prices to launch 60–70% above HBM3E, though costs will likely ease as Samsung Electronics and Micron Technology ramp up their own versions. Samsung has said its HBM4 will use its more advanced 1c, sixth-generation 10nm process.
The U.S. government is weighing a plan to let Korean companies Samsung Electronics and SK Hynix bring American-made chipmaking equipment into their plants in China under a limited approval framework.
According to Bloomberg, the U.S. Department of Commerce has proposed issuing annual permits for the two South Korean memory chip giants, replacing the indefinite authorizations previously granted under the Biden administration. The idea of a “site license” was recently presented to Korean officials, according to sources familiar with the matter.
Samsung and SK Hynix were previously covered by the U.S. Validated End User (VEU) program, which allowed certain factories in China to import U.S.-made tools without additional licenses. That privilege ended after the Trump administration removed their Chinese plants from the VEU list, raising concerns about supply disruptions and difficulties in maintaining operations.
