HBM Storage: The Core Engine of the AI Computing Era
With the rapid development of artificial intelligence technology, the demand for AI computing power is growing exponentially, and High Bandwidth Memory (HBM), as a critical component of AI chips, has become the core engine driving the enhancement of AI computing power. In the third quarter of 2026, the global HBM market size will exceed $10 billion, a year-on-year growth of over 80%. The market size is projected to reach $25 billion in 2027, maintaining a compound annual growth rate of over 60%. Behind this astonishing growth rate is the continuous explosion in demand for high-bandwidth memory from large AI model training and inference.HBM Technology Analysis: Evolution from Generation 1 to Generation 4
High Bandwidth Memory (HBM) is an advanced DRAM memory technology that achieves higher bandwidth and lower power consumption by stacking multiple DRAM chips and interconnecting them through Through Silicon Via (TSV) technology. Compared to traditional GDDR memory, HBM can offer higher bandwidth at the same power consumption, or lower power consumption at the same bandwidth, which is crucial for compute-intensive AI applications. Since HBM 1 was introduced in 2013, HBM technology has undergone four major evolutions: HBM 1, HBM 2, HBM 2e, and HBM 3, while the latest HBM 4 technology is currently under accelerated development. Each generation brings significant performance improvements. HBM 3's bandwidth is already more than six times that of HBM 1, and the upcoming HBM 4 is expected to achieve bandwidth exceeding 8TB/s, further meeting the stringent storage performance requirements of large AI models.Market Landscape: Three Pillars and Emerging Challengers
The current global HBM market is dominated by three South Korean companies: SK Hynix, Samsung Electronics, and Micron Technology. Together, these three companies account for over 95% of the global HBM market share, forming a highly concentrated market structure. SK Hynix, as the pioneer in HBM technology, currently holds about 45% of the global HBM market share, and its HBM 3 product is widely used in NVIDIA's latest AI chips. Samsung Electronics follows closely with a market share of about 35%, while Micron Technology holds the remaining about 15% share. It is noteworthy that although Micron Technology performs strongly in the traditional DRAM market, it entered the HBM field later and has relatively insufficient technological accumulation, and is currently actively catching up. Besides these three giants, companies such as Changjiang Storage and Changjiang Storage are actively developing HBM technology, attempting to break the current market monopoly. However, due to technological barriers and patent restrictions, these emerging companies find it difficult to shake the existing structure in the short term.Demand Drivers: The Revolution of Large AI Models and Data Centers
The rapid growth in the HBM market is mainly due to the explosion in demand for large AI model training and inference. As the parameter scale of large language models like ChatGPT and Claude expands from billions to trillions, the demand for computing and storage grows exponentially. Taking GPT-4 as an example, its training process requires the collaboration of thousands of high-performance GPUs, and each GPU needs large-capacity, high-bandwidth HBM memory to support model training and inference. Besides large AI models, demand for HBM in areas such as data centers, high-performance computing, and autonomous driving is also growing rapidly. Market research agencies predict that by 2028, AI applications will account for over 70% of the total demand for HBM, becoming the main driver for HBM market growth. ### Technical Challenges: Yield, Cost, and Capacity Bottlenecks Despite the broad prospects of the HBM market, technical challenges remain severe. First, HBM production processes are complex, and yield control is difficult. HBM requires the precise stacking of multiple DRAM chips and interconnection via TSV technology, which demands extremely high precision in process and process control. Currently, HBM3 yield is only about 60% of traditional DRAM, and HBM4 yield is expected to decrease further, which will directly affect product cost and supply capability. Second, the production cost of HBM is much higher than that of traditional DRAM. Taking HBM3 as an example, its production cost is 5 to 8 times that of traditional DDR5 memory, keeping HBM prices high and limiting its application in some price-sensitive areas. Furthermore, HBM capacity expansion faces bottlenecks. Currently, global HBM capacity is mainly concentrated in South Korea, and investment in new foundries is huge and long-term, making it difficult to meet the demand for rapid market growth. Industry analysis suggests that the global HBM market will continue to face supply shortages from 2026-2027, with supply and demand gaps potentially reaching over 30%. ### Investment Opportunities: Industry Chain Layout and Value Differentiation Different companies will face different investment opportunities in the HBM industry chain. From an industry chain perspective, the HBM chain mainly includes upstream semiconductor equipment and material suppliers, midstream HBM manufacturers, and downstream AI chip manufacturers. In the upstream segment, semiconductor equipment manufacturers such as Application Materials and Tokyo Electronics will benefit from the growing equipment demand driven by HBM capacity expansion. Meanwhile, equipment and material suppliers related to TSV will also see development opportunities. In the midstream segment, HBM manufacturers such as SK Hynix, Samsung Electronics, and Micron Technology will directly benefit from the growth in market demand. The HBM business revenue ratio for these three companies is expected to increase from 15% in 2025 to over 30% by 2028, becoming a significant driver for their performance growth. In the downstream segment, AI chip manufacturers such as NVIDIA, AMD, and Intel will benefit from enhanced product competitiveness due to improved HBM performance. Especially NVIDIA, as the leader in the AI chip market, will see continued growth in demand for HBM, indirectly driving the development of the HBM market. ### Risk Warning: Technological Iteration and Geopolitics Although the HBM market has broad prospects, investors still need to pay attention to related risks. First, the risk of technological iteration cannot be ignored. With the continuous development of storage technologies, new storage technologies such as MRAM and ReRAM may pose challenges to HBM in the future, affecting its market position. Second, geopolitical factors may affect the stability of the HBM supply chain. Currently, global HBM capacity is highly concentrated in South Korea, while countries such as the United States and Japan are actively promoting the localization of storage capacity, which could change the global HBM industry landscape. In addition, trade restrictions and technological sanctions may also affect the global flow of HBM technology. For example, the US export restrictions on semiconductors to China have impacted Chinese enterprises' ability to acquire advanced HBM products, which could reshape the global HBM market landscape.Future Outlook: HBM4 and New Storage Technologies
Looking ahead, HBM technology will continue to develop in the direction of high performance, low power consumption, and high integration. HBM4 is expected to be commercialized between 2027 and 2028, with bandwidth exceeding HBM3 by more than 50%, reaching over 8TB/s, while power consumption will further decrease. This will significantly enhance the training and inference efficiency of AI large models. In addition to traditional HBM technology, new storage technologies such as Processing-in-Memory and In-Memory Computing are also developing rapidly. These technologies are expected to integrate with HBM technology in the future, further breaking through the bottlenecks of AI computing power. Overall, HBM as a core component of AI computing power will continue to maintain a rapid growth trend over the next 3-5 years. Investors should focus on leading companies in the HBM industry chain, especially HBM manufacturers with technological advantages and sufficient capacity, as well as AI chip manufacturers deeply tied to HBM, to share in the boom of the AI computing era.Investment Strategy: Select Leaders, Diversify Layout
For investments related to HBM, it is recommended to adopt a strategy of "select leaders, diversify layout." First, focus on leading companies in the HBM field, such as SK Hynix, Samsung Electronics, etc., these companies possess technological advantages and economies of scale, and can continue to benefit from industry growth. Second, focus on high-quality companies in the upstream and downstream parts of the HBM industry chain, including semiconductor equipment manufacturers, material suppliers, and AI chip manufacturers. These companies will create synergy in the HBM market and jointly share in the boom of AI computing growth. Finally, investors are advised to pay attention to HBM-related ETF products, such as semiconductor ETFs and AI computing power ETFs, to diversify investment risks while grasping overall growth opportunities in the HBM market.Conclusion
HBM storage, as a core component of AI computing power, is in a golden growth period. With continuous progress in AI technology and expanding application scenarios, the HBM market will maintain high-speed growth, bringing substantial returns to companies in the industry chain. However, investors must also pay attention to risks such as technological iteration and geopolitical factors, adopting a prudent investment strategy to seize investment opportunities in the HBM era. In the era of AI-driven digital economy, HBM storage is not just a technology product but a key engine for enhancing AI computing power. Investors should view the development of the HBM market from a long-term perspective, select high-quality companies, and share in the growth boom of the AI computing era.
