According to the Financial and Economic APP, Choi Tae-yuan, the chairman of SK Group, one of South Korea's large chaebols, stated in an interview on Monday.$NVIDIA (NVDA.US)$CEO Huang Renxun had previously requested SK Group's storage chip manufacturing giant SK Hynix to launch its next-generation high-bandwidth storage product HBM4 six months ahead of schedule.

In its financial report in October, SK Hynix had indicated plans to provide HBM storage systems to major customers (speculated to be NVIDIA and AMD) in the second half of 2025. An SK Hynix spokesperson mentioned on Monday that this timetable is indeed faster than the original target, but did not provide further details.

Huang Renxun personally requested SK Hynix to expedite the delivery speed of the next generation HBM – HBM4 storage system, highlighting the strong demand from NVIDIA for more advanced AI GPU systems with higher capacity and more energy-efficient HBM storage systems, further emphasizing OpenAI, Anthropic, and$Microsoft (MSFT.US)$,$Amazon (AMZN.US)$and $Meta Platforms (META.US)$As artificial intelligence, cloud computing, and internet giants have almost limitless "explosive demand" for AI training/inference computing power, they are forcing Nvidia's core chip foundry$Taiwan Semiconductor (TSM.US)$To work overtime to expand Blackwell AI GPU capacity, and to request Nvidia to speed up the research and development process of the next generation AI GPU with higher performance, larger storage capacity, stronger inference efficiency, and more energy-efficient.

Nvidia plans to launch the next generation AI GPU architecture, Rubin, in 2026, with Rubin AI GPU expected to be equipped with HBM4 storage systems. According to incomplete statistics, Nvidia currently holds an 80%-90% market share in the global data center AI chip market, considered almost a monopoly, with AMD accounting for nearly 10%, and other shares held by Google TPU, Amazon's in-house ASIC, and other major self-developed AI chip manufacturers.

As the core HBM storage system supplier for Nvidia's H100/H200 and recently mass-produced Blackwell AI GPU, SK Hynix has been leading the global storage chip capacity race to meet the explosive demand for HBM storage systems from major customers such as Nvidia, AMD, Google, and other enterprises, and the increasing demand for data center SSDs and other enterprise-level storage products. These storage-level chip products are considered core hardware for handling massive data to train increasingly powerful AI models and meet the skyrocketing demand for cloud AI inference computational power.

But SK Hynix also faces challenges from Samsung Electronics and the American storage giant$Micron Technology (MU.US)$Increasingly intense competitive pressures. Samsung announced in last week's financial report that it has made positive progress in reaching a supply agreement with a major customer (possibly NVIDIA), after the company has long been plagued by product testing delays and failed to pass NVIDIA's qualification test in the last test. Samsung added that the company is in negotiations with a major customer and could mass produce 'improved' HBM3E products in the first half of next year. Samsung also plans to produce the next generation HBM - HBM4 memory in the second half of next year to keep up with competitor SK Hynix's pace.

In the USA, Micron, the American storage giant, is another HBM supplier approved by Nvidia, in addition to SK Hynix. In February this year, Micron began mass production of HBM3E storage systems designed specifically for AI and high-performance computing, and stated that some of Nvidia's H200 and Blackwell AI GPUs will be equipped with Micron's HBM3E. Following that, Micron's CEO has repeatedly mentioned that Micron's HBM capacity for the next two years has been sold out. Micron also stated that they are expediting the development process of the next generation HBM4 and HBM4e.

While being interviewed almost at the same time as Cui Taiyuan, SK Hynix CEO Guo Lu stated at the 2024 SK AI Summit in Seoul that the company plans to provide the latest 12-layer HBM3E to a major customer (market speculation suggests it is nvidia) by the end of this year, and plans to deliver even more advanced 16-layer HBM3E storage samples early next year. Additionally, the CEO revealed in the interview that nvidia's AI GPU supply still cannot meet the demand, and nvidia has repeatedly requested SK Hynix to accelerate the scale of HBM3E supply.

Exploding demand for HBM, dubbed as a "money-making machine."

nvidia's founder and CEO Jensen Huang recently revealed in an interview that the Blackwell architecture AI GPU has been fully mass-produced, and the demand is extremely "crazy"; at the same time, as AI GPU demand skyrockets, the demand for HBM storage systems also sharply increases, and may continue to outstrip supply in the coming years.

According to the well-known technology industry chain analyst Guo Mingchi from TF International Securities, the latest industry chain order information for the NVIDIA Blackwell GB200 chip indicates that Microsoft is currently the largest GB200 customer in the world, with Q4 orders surging 3-4 times, exceeding the total orders of all other cloud service providers.

Guo Mingchi stated in a recent report that the capacity expansion of the Blackwell AI GPU is expected to start in the early fourth quarter of this year, with shipments in the fourth quarter expected to be between 0.15 million and 0.2 million units. It is projected that shipments in the first quarter of 2025 will significantly increase by 200% to 250%, reaching between 0.5 million and 0.55 million units. This implies that nvidia may achieve its sales target of millions of AI server systems in just a few quarters.

According to the latest forecast data from the Wall Street financial giant Citigroup, by 2025, the datacenter-related capital expenditures of the four largest technology giants in the United States are expected to increase by at least 40% year-on-year, and these massive capital expenditures are basically linked to generative artificial intelligence, meaning that the computational demands of AI applications like ChatGPT remain significant. Citigroup states that this implies that the giants are expected to further expand their datacenter spending beyond the already strong spending levels of 2024, and the institution expects this trend to continue to provide very significant positive catalysts for the AI GPU undoubted dominator Nvidia as well as datacenter interconnect (DCI) technology providers' stock prices.

The HBM storage system, in conjunction with the core hardware provided by nvidia, the AI GPU such as H100/H200/GB200, and a wide range of AI ASIC chips (such as Google TPU), together with HBM and nvidia's AI GPU, combine to form the so-called "NVIDIA AI GPU Server." HBM and AI GPU are essential for driving heavyweight AI applications such as ChatGPT and Sora, the stronger the demand for AI GPU, the more intense the demand for HBM storage.

HBM is a high-bandwidth, low-energy storage technology specifically used in the high-performance computing and graphics processing fields. HBM, through 3D stacked storage technology, connects multiple stacked DRAM chips together comprehensively, utilizing fine Through-Silicon Vias (TSVs) for data transmission, achieving high-speed, high-bandwidth data transfer. HBM, through 3D stacking technology, stacks multiple memory chips together, significantly reducing the storage system's space footprint, lowering energy consumption for data transfer, while high bandwidth can significantly enhance data transfer efficiency, allowing AI large models to run more efficiently 24 hours a day.

Especially, HBM storage systems have unparalleled low-latency characteristics, able to quickly respond to data access requests. Generative AI models like GPT-4 often need to frequently access large datasets and perform extremely heavy model inference workloads. The powerful low-latency characteristics can greatly improve the overall efficiency and response speed of AI systems. In the field of AI infrastructure, HBM storage systems are fully integrated with Nvidia's H100/H200 AI GPUs, AMD's MI300X AI GPUs, as well as Nvidia's mass-produced B200 and GB200 AI GPUs, and AMD's MI325X.

Goldman Sachs, a major Wall Street firm, released a research report stating that the exceptionally strong demand for Generative Artificial Intelligence (Gen AI) by enterprises has driven higher shipments of AI servers and higher HBM density per AI GPU. The institution significantly increased its total HBM market size estimate, now expecting the market size to grow significantly from 2023 to 2026 at a compound annual growth rate of 100%, reaching $30 billion by 2026. Goldman Sachs predicts that the undersupply situation in the HBM market will continue in the coming years, with major players such as SK Hynix, Samsung, and Micron continuing to benefit.

Editor/ping