According to reports, Samsung Electronics and SK Hynix are in the final stages of preparations to supply NVIDIA with the 6th-generation High Bandwidth Memory (HBM4). The focus lies on when HBM4 might be integrated into NVIDIA’s next-generation AI chips, the “Vera Rubin.” Both companies have provided paid samples of HBM4 to NVIDIA, and the product is in the final optimization phase. Key performance metrics such as power efficiency, heat management, and processing speed are being adjusted to meet NVIDIA’s standards. As NVIDIA plans to incorporate eight units of HBM4 per Vera Rubin chip debuting later this year, the market is closely watching when mass production and integration will commence, potentially in the second quarter instead of the initially expected first quarter.
NVIDIA had previously increased the data rate requirements for HBM4 to over 11Gbps last year, which could extend testing periods as companies work to meet these specifications. There is also speculation that NVIDIA might further increase these requirements. CEO Jensen Huang emphasized during the CES 2026 keynote that Vera Rubin will have over four times the efficiency of the previous Blackwell product.
Meanwhile, demand for Blackwell is expected to surge in the first half of the year as major tech companies continue to incorporate it into AI data center infrastructures. Due to its capabilities, Blackwell remains sufficient to drive AI operations, potentially affecting the schedule for HBM4’s incorporation into Vera Rubin. However, demand for HBM4 is anticipated to rise quickly in the latter half of the year, given its expected use in NVIDIA’s upcoming AI chip “Rubin Ultra” in 2027.
In their pursuit of leadership in the next-generation memory market, Samsung has resumed construction at its semiconductor production site in Pyeongtaek, and SK Hynix plans to expedite the start of operations at its Cheongju M15X plant by the first half of this year.