NVIDIA CEO Jensen Huang personally hosted a dinner last week for engineers from a partner company, a rare move underscoring the strategic importance of next-generation high-bandwidth memory HBM4 for the AI chip giant. With Samsung Electronics taking the lead by shipping HBM4 samples, whether SK Hynix can deliver high-performance products on schedule will directly impact the market performance of NVIDIA's new AI accelerator, Vera Rubin, set to launch in the second half of this year.

On the evening of February 14, Huang appeared at a Korean fried chicken restaurant near NVIDIA's headquarters, spending approximately two hours moving between tables, preparing soju cocktails, and toasting with over 30 engineers from SK Hynix and NVIDIA. He repeatedly emphasized "We are one team" and "I am proud of you," urging the engineers to "deliver extraordinary results through relentless challenge and effort," specifically mentioning the sixth-generation HBM4 products committed by SK Hynix.

This impromptu gathering sent a clear signal: HBM4 is viewed by NVIDIA as a critical differentiating component for the Vera Rubin accelerator. NVIDIA has set specifications for HBM4 suppliers requiring an operating speed above 11 Gbps and bandwidth exceeding 3.0 TB/s, which is over 30% higher than similar requirements from competitor AMD. Industry analysts believe Huang's personal appearance increases the likelihood of SK Hynix maintaining its position as the largest HBM4 supplier.

According to industry sources, NVIDIA preliminarily allocated this year's HBM supply shares in December last year, with SK Hynix receiving over 55%, Samsung Electronics taking between 20% and nearly 30%, and Micron Technology accounting for approximately 20%. While Samsung's progress in the technology race has raised the possibility of share changes, the industry widely believes SK Hynix is expected to secure the largest quota after completing quality optimization in the first quarter.

**Uncommon 'Engineer Diplomacy'** The semiconductor industry views Huang's personal hosting of a dinner for partner engineers as a highly unusual gesture. The gathering was quickly arranged after Huang instructed NVIDIA staff last week to "organize a dinner to encourage the SK Hynix HBM engineers," highlighting the critical nature of SK Hynix's HBM4 for NVIDIA's future business.

Huang arrived at the 99 Chicken restaurant in Santa Clara around 5:20 PM. During the final toast near the end of the dinner, he stated: "AI accelerators and HBM4 represent extraordinary, world-class challenging technologies. I am proud of all of you working day and night, and I believe you will deliver outstanding results." He added: "I know the development timelines for HBM4 and Vera Rubin are tight, but I believe in you. Now is the time for SK Hynix and NVIDIA to show the world something great together."

SK Hynix formally entered NVIDIA's supply chain in July 2020 with its HBM2E (third-generation) products, subsequently becoming the de facto exclusive supplier for HBM3 (fourth-generation) and HBM3E (fifth-generation), forming a tight relationship with TSMC known as the "AI Semiconductor Tripartite Alliance."

**Technical Hurdles and Time Pressure** HBM4 is seen as a key component that will determine the performance of NVIDIA's next-generation AI accelerator, Vera Rubin. Scheduled for launch in the second half of this year, HBM4 is a high-performance memory module made by stacking 12 layers of advanced DRAM chips, responsible for delivering massive amounts of data to the computing GPU in a timely manner.

NVIDIA's specification requirements for HBM4 far exceed those of its competitors. Its demanded operating speed and bandwidth metrics are over 30% higher than AMD's requirements for HBM4, effectively positioning HBM4 as a key differentiator for Vera Rubin.

Unlike the HBM3E market, which is largely dominated by SK Hynix, the HBM4 market, set to fully open in the second half of the year, presents a different competitive landscape. Samsung Electronics shipped its first official HBM4 samples to NVIDIA on February 12, becoming the industry's first mover, with its product achieving an operating speed of 11.7 Gbps (up to 13 Gbps) and bandwidth of 3.3 TB/s.

SK Hynix has currently ensured HBM4 performance reaches 11.7 Gbps or higher, is supplying paid samples to NVIDIA in volume, and is concurrently working on performance optimization. The industry expects SK Hynix to receive formal "volume supply" approval from NVIDIA in the near future.

At the dinner, Huang urged SK Hynix engineers to "deliver top-performance HBM4 without delay." This statement was interpreted as a clear requirement regarding the supply timeline.

**HBM Share Battle Intensifies** As Samsung makes progress in the HBM4 technology race, this year's supply share allocation could see changes. However, the industry generally believes that SK Hynix still has a strong chance of securing the largest quota after completing quality optimization in the first quarter.

A semiconductor industry source explained to media: "The final HBM4 optimization work for the three memory manufacturers won't be complete until around March. There has also been a recent trend of prioritizing increased general DRAM production to improve profitability, rather than fighting for HBM market share."

Evercore's research has also pointed out differences in product generational evolution, with some industry insiders emphasizing the distinction between "minor" and "major" evolutions. The transition from H100 to B100/B200 is seen as minor, while the move to GB-Series rack-scale systems is a major evolution. The shift from GB-Series to VR-Series is considered minor, but the transition to Rubin Ultra, accompanied by a cluster scale expansion from 144 to 576, represents a major upgrade.

The dinner hosted personally by Jensen Huang served both as recognition for a long-term partner and as a督促 for a critical supply chain link. For SK Hynix, the ability to maintain its leading position amidst fierce competition will hinge on its execution over the coming months.