In 2026, the global AI industry will enter an explosive growth phase, with demand for cloud data centers to edge computing devices, and the demand for AI chips rising exponentially. HBM (High Bandwidth Memory) performance and supply capacity, as the "memory heart" of AI chips, directly determine the upper limit of AI computing power. Recently, SK Hynix, a giant in the Korean memory market, announced its Q2 2026 earnings report, once again putting HBM in the spotlight—its HBM revenue grew by 300% year-over-year, reaching a new historical high, and capacity expansion plans are also accelerating. This data not only reflects the boom in the HBM market but also signals a critical juncture for breaking through AI computing bottlenecks. ## HBM: The "Memory Heart" of AI Computing Power In the architecture of AI chips, HBM plays a crucial role. Compared to traditional DRAM, HBM achieves higher bandwidth and lower power consumption by stacking multiple DRAM chips and connecting them through an interposer. For AI training and inference tasks, higher bandwidth means faster data transmission speeds, reducing waiting times during computation and thereby improving overall computing efficiency. For example, NVIDIA's H100 and A100 GPUs, as well as AMD's MI300 series, rely on HBM to support their large-scale parallel computing capabilities. With the explosive growth in AI model parameters (such as GPT-4 parameters exceeding 1.7 trillion), the demand for memory bandwidth is rapidly increasing. Industry research institutions predict that the global HBM market size will reach $15 billion in 2026, with a compound annual growth rate exceeding 40%. However, HBM supply capacity is one of the key bottlenecks constraining AI chip production capacity. Previously, due to the complexity of HBM manufacturing technology and limited capacity, chip manufacturers like NVIDIA and AMD faced the challenge of "difficulty in securing one chip," even affecting the delivery cycle of AI servers. ## SK Hynix: Behind the Threefold Increase in HBM Revenue As the world's second-largest HBM supplier, SK Hynix's strong performance in its Q2 earnings report confirms the strong demand in the HBM market. The report showed that in the second quarter of 2026, SK Hynix's HBM revenue reached $1.2 billion, a year-on-year increase of 300%, accounting for 15% of its total revenue, an increase of 10 percentage points from the same period last year. This growth was mainly due to two factors: a surge in orders from AI chip manufacturers and the advancement of capacity expansion plans. In terms of capacity, SK Hynix has started construction of its second HBM factory in Cheongju, South Korea, expected to become operational in 2027, which will increase HBM capacity by 50%. Furthermore, the company plans to launch an upgraded version of HBM3e—HBM4—in the second half of 2026, which offers a 20% increase in bandwidth and a 15% reduction in power consumption, further meeting the high-performance demands of AI chips. CEO Lee Seok-hee stated during the earnings conference call, "HBM is the core driver of SK Hynix's future growth, and we will continue to increase investment to maintain a leading position in the era of AI computing power." ## Industry Competitive Landscape: Samsung and Micron Catch Up, Technological Iteration Accelerates Despite SK Hynix's important position in the HBM market, competition remains fierce. Samsung Electronics, as the world's largest memory manufacturer, has also increased investment in the HBM field in recent years, and its HBM3e product has secured orders from customers such as NVIDIA and AMD. Micron Technology is gradually increasing its market share in HBM by acquiring Cypress Semiconductor's HBM technology through acquisitions. Furthermore, Chinese memory manufacturers such as Yangtze Memory Technologies and ChangXin Memory are actively developing HBM technology, attempting to break the monopoly of foreign manufacturers. In terms of technological iteration, HBM4 is becoming the industry focus. Compared to HBM3e, HBM4 adopts more advanced stacking processes (such as 12-layer stacking), increases bandwidth to 3.2TB/s, reduces power consumption to 1.2W/Gb/s, and can better support the demands of next-generation AI chips. NVIDIA's next-generation GPU architecture Rubin is expected to adopt HBM4, and AMD's MI400 series also plans to feature HBM4. Manufacturers such as SK Hynix, Samsung, and Micron are accelerating the R&D and mass production of HBM4, with large-scale commercialization expected in 2027. ## Breakthroughs in AI Computing Power: From Capacity to Technology SK Hynix's capacity expansion and technological iteration provide important support for breaking through the AI computing power bottleneck. Previously, the main bottlenecks faced by AI chip manufacturers included insufficient HBM capacity, packaging technology limitations, and supply chain risks. Now, with the increase in HBM capacity (such as the commissioning of the Clearzhou factory by SK Hynix), advancements in packaging technology (such as doubling of TSMC's CoWoS packaging capacity), and supply chain diversification (such as the entry of Samsung and Micron), these bottlenecks are gradually being alleviated. However, challenges still remain. HBM production involves complex stacking processes and silicon interlayer technology, requiring extremely high precision in equipment and process control, leading to low yields (current HBM yield is about 70%-80%, far below the 95% or more of traditional DRAM). Furthermore, HBM has a high cost, accounting for 30%-40% of the total AI chip cost, so reducing costs is an issue the industry needs to solve. SK Hynix has reduced the unit cost of HBM by 15% through process optimization and yield improvement, but further efforts are still needed. ## Investment Opportunities: Potential of the HBM Industry Chain For investors, the HBM industry chain holds rich investment opportunities. Firstly, HBM suppliers such as SK Hynix, Samsung Electronics, and Micron Technology will directly benefit from the growth in market demand. Secondly, HBM packaging manufacturers such as TSMC and ASE, the expansion of their CoWoS packaging capacity will drive demand for related equipment manufacturers (such as Applied Materials, Lam Research). Furthermore, HBM raw material suppliers (such as silicon wafers, packaging materials) will also see growth. Looking at the US stock market, HBM concept stocks such as SK Hynix (although a Korean company, it has an ADR in US stocks), Micron Technology (MU), and Samsung Electronics (SSNLF) are attracting investor attention. Additionally, memory manufacturers in the A-share market such as Yangtze Memory Technologies and ChangXin Memory (although not listed, but related industry companies such as SMIC and Huatian Technology) are also worth watching. ## Future Outlook: HBM Trends in the Era of AI Computing Power Looking ahead, HBM development will show three trends: first, continuous improvement in capacity and bandwidth, with HBM5 expected to launch in 2028, achieving bandwidth exceeding 4TB/s; second, cost reduction through process optimization and economies of scale, with the unit cost of HBM expected to decrease by 20%-30%; third, expansion of applications, where HBM will be applied to fields beyond AI chips, such as High-Performance Computing (HPC) and autonomous driving, further expanding market demand. SK Hynix's Q2 earnings report is not only a reflection of its own performance growth but also a microcosm of the AI computing era. As HBM capacity expands and technological iteration progresses, breakthroughs in AI computing bottlenecks are gradually being achieved, and investors in the HBM industry chain will welcome long-term growth opportunities.
Detail Page Advertisement