SK hynix eyes Q4 rollout of 12-layer HBM4 as it locks in Nvidia supply deal

HBM4 chips with double the bandwidth are coming, and SK hynix is betting on its early lead and deep expertise

SK hynix is close to finalizing a supply deal with Nvidia for its sixth-generation high-bandwidth memory (HBM4), as the two companies gear up for the launch of Nvidia’s next-gen AI accelerator, the Rubin platform, according to people familiar with the matter.

The South Korean chipmaker plans to begin mass production of 12-layer HBM4 chips in the fourth quarter of 2025, supplying initial volumes to Nvidia once Rubin’s launch timeline and order size are confirmed. Production will take place at SK hynix’s nearly completed M15X fab in Cheongju, where new HBM4-specific equipment is already being installed.

The aggressive timeline underscores SK hynix’s push to lock in its leadership in the fast-growing HBM segment, where it already commands nearly 70% of the market, according to Counterpoint Research. The company shipped the industry’s first HBM4 samples earlier this year and is racing to stay ahead as AI-related demand surges.

Nvidia CEO Jensen Huang, who visited SK hynix’s booth at Computex 2025 in Taipei earlier this month, reportedly asked the chipmaker to “fully support” HBM4 as Nvidia prepares Rubin for launch.

Nvidia CEO Jensen Huang signed an HBM memory display at SK hynix’s booth during Computex 2025 at the Nangang Exhibition Center in Taipei, Taiwan, on May 20, 2025./Yonhap

HBM4 represents a major evolution in memory design. It’s the first generation to feature a foundry-grade logic die and double the number of I/O channels, significantly increasing bandwidth. Since SK hynix lacks a foundry business, it’s outsourcing logic die production to TSMC.

Analysts say HBM4 signals the beginning of more customized memory stacks, although the 12-layer versions remain relatively close to HBM3E in manufacturing complexity. Still, SK hynix appears confident that its experience gives it an edge.

Rivals are closing in. U.S.-based Micron is already supplying Nvidia with 8-layer HBM3E chips, and Samsung Electronics is currently running sample tests for its own HBM3E lineup.

One potential snag: Rubin’s launch timeline. Sources say the upcoming AI chip—expected to debut with HBM4—comes with a premium price tag, and final volume commitments may depend on demand from major cloud and AI customers. Any delay from Nvidia could push back SK hynix’s production ramp. That said, some industry sources believe volume manufacturing could begin as early as October, with equipment orders possibly placed in July or August.

SK hynix remains on track. It expects to complete the M15X fab in the second half of this year and start HBM4 mass production shortly after—positioning itself to lead as the AI memory arms race intensifies.