#分享美股交易赢英伟达股票

Memory stocks receive good news—Nvidia's Jensen Huang denies rumors that the company's chips will cut back on high-bandwidth memory (HBM) usage

On June 5th, Nvidia CEO Jensen Huang gave an interview with foreign media in South Korea, when asked about the rumor that "Nvidia will reduce HBM usage due to supply shortages," Huang Huang clearly responded: "We will use a large amount of HBM memory. Of course, supply is indeed limited at the moment. Therefore, we need to use memory more wisely across all systems. We will continue working with our partners here to secure as much supply as possible and to deploy it as cleverly as possible."

Regarding suppliers, Huang Huang confirmed that three memory suppliers (Samsung Electronics, SK Hynix, and Micron Technology) will provide HBM4 for Nvidia's latest AI chips, all of which have passed qualification and are in mass production, "all competing to support the company's Vera Rubin architecture." Previously, Huang Huang stated that the Vera Rubin chips are fully in production and are expected to be delivered in the third quarter of this year.

Huang Huang's response this time was very timely. On June 4th, research firm SemiAnalysis published an article stating that the capacity of Nvidia's next-generation flagship supercomputer rack Vera Rubin NVL72's SOCAMM DRAM (a new memory module standard designed specifically for AI servers) may have decreased from the previously expected approximately 55TB to about 28TB. Meanwhile, most Rubin systems will use 96GB SOCAMM modules instead of the previously market-expected 192GB.

After the news broke, some opinions suggested that Nvidia has started to cut back on memory usage due to shortages. Coupled with the fact that the financial report released by custom AI chip (ASIC) giant Broadcom on the same day failed to meet investors' high expectations, the global storage concept experienced a pullback.

On the 4th, Micron Technology (Nasdaq: MU) temporarily plunged over 10% during trading, erasing more than $100 billion in market value in a single day. On June 5th, the Korean stock market also saw a sharp correction, with semiconductor giants' stocks suffering heavy losses. By the close of that day, SK Hynix fell 9.92%, and Samsung Electronics fell 6.4%.

However, some analysts pointed out that, upon careful reading of the SemiAnalysis report, it can be seen that the main change in memory comes from CPUs using LPDDR5X memory. Each Vera Rubin NVL72 rack is equipped with 72 Rubin GPUs and 36 Vera CPUs, with each GPU using 288GB of HBM4 memory, totaling about 20.7TB per rack, which has not changed.

At the same time, because SOCAMM is a slot-based design, it can be replaced with higher-capacity memory at any time. In other words, the initial memory capacity of the CPUs in the Vera Rubin NVL72 shipment configuration has been reduced.

Dylan Patel, founder of SemiAnalysis, also commented on the X (formerly Twitter) platform: "I really like one thing—that those who repost our report often miss most of the content in the report. This happens quite often."

In this interview, Huang Huang also mentioned that during his trip to South Korea, he arranged meetings with Hyundai Motor, LG, SK Hynix, Samsung Electronics, and South Korea's internet giant NAVER. The main goal was to ensure that supply chain partners "stay aligned and be well prepared." He said he brought "a huge business" to Korea and hinted that there are more surprises yet to be announced.