In the era of AI large models, the competition for computing power has extended from the chips themselves to the memory layer. HBM (High Bandwidth Memory)凭借其 high bandwidth and low latency characteristics, has become the core bridge connecting GPUs and data. Compared with traditional DDR memory, HBM uses 3D stacking technology to vertically stack multiple DRAM chips and connects them through a silicon interposer, greatly increasing data transmission rates to meet the needs of massive data parallel processing in AI training. For example, the HBM3 memory bandwidth of NVIDIA's H100 GPU reaches 3.35TB/s, which is twice that of the previous generation A100, and the upcoming H200 will further increase to 4.8TB/s, which is precisely supported by the iteration of HBM technology.
For AI chip manufacturers, the supply capacity of HBM directly determines the performance ceiling of their products. NVIDIA CEO Jensen Huang once said: 'HBM is the 'blood' of AI chips. Without sufficient HBM, even the most powerful GPU cannot exert its full computing power.' Currently, AI large model training needs to process trillions of parameters, and the amount of data transmitted per second reaches the PB level. The bandwidth and capacity of HBM have become key bottlenecks restricting the improvement of computing power. Therefore, the production capacity, price, and technological innovation of HBM storage have become one of the most concerned links in the AI computing industry chain.
I. SK Hynix Q3 Pre-increase: Signal of HBM Capacity Release
Recently, SK Hynix announced its Q3 2026 performance pre-increase announcement, expecting Q3 revenue to increase by about 20% year-on-year, of which HBM business revenue increased by 50% year-on-year, becoming the core driver of the company's performance growth. This data has attracted widespread market attention, as SK Hynix is the world's second-largest HBM supplier (second only to Samsung), and its capacity release means that the supply pattern of the HBM market is changing.
In terms of production capacity, SK Hynix is accelerating the production of HBM3E and HBM4. HBM3E is the current mainstream product, used in high-end AI chips such as NVIDIA H100 and AMD MI300; while HBM4 is the next-generation product, expected to be mass-produced in 2027, with bandwidth increasing to 6.4TB/s to further meet the needs of future AI large models. SK Hynix stated that its monthly production capacity of HBM3E has increased from 100,000 pieces in 2025 to 150,000 pieces in Q3 2026, and is expected to reach 250,000 pieces in 2027, with a production capacity expansion speed far exceeding market expectations.
In addition to SK Hynix, Samsung is also increasing its HBM production capacity. Samsung plans to reach a HBM production capacity of 200,000 pieces/month in 2026 and further increase to 300,000 pieces/month in 2027. Micron, through acquiring Elpida Memory's HBM technology, is gradually entering the HBM market, with an expected HBM production capacity of 50,000 pieces/month in 2026. Overall, global HBM production capacity is transitioning from 'shortage' to 'tight balance', and it is expected to achieve supply-demand balance in 2027.
II. HBM Price Trend: From 'Sky-High Prices' to 'Rational Return'
Over the past two years, HBM prices have experienced fluctuations of 'surge - fall'. In 2024, due to the AI demand explosion, the price of HBM3E once reached over $1,000 per piece, while in 2025, with the release of production capacity, the price fell back to $600-700 per piece. SK Hynix's Q3 pre-increase shows that the gross profit margin of the HBM business still remains above 40%, indicating that even with the price fall, high gross profit still supports its performance growth.
For US stock investors, the stability of HBM prices is crucial to the profit margins of AI chip manufacturers. NVIDIA mentioned in its 2025 financial report that HBM costs account for more than 30% of its GPU costs. If HBM prices continue to fall, it will directly improve its profit margin. AMD stated that the HBM cost of its MI300X GPU accounts for about 25%, and the price fall will enhance its product competitiveness. Therefore, the stabilization of HBM prices will be beneficial to the performance of US AI chip leaders.
III. Impact on US AI Sector: From 'Bottleneck' to 'Opportunity'
1. NVIDIA: Improved Supply Chain Stability
As the world's largest AI chip manufacturer, NVIDIA's HBM supply mainly relies on Samsung and SK Hynix. SK Hynix's Q3 production capacity increase means that NVIDIA's HBM supply is more stable, which helps it expand the production capacity of high-end products such as H200 to meet market demand. In addition, NVIDIA is developing the 'HBM-on-Logic' technology, which integrates HBM and GPU on the same chip to further increase bandwidth. This technology requires the cooperation of HBM suppliers, and SK Hynix's production capacity expansion provides support for it.
2. AMD: The Key to Catching Up with NVIDIA
As NVIDIA's main competitor, AMD's HBM supply for its MI300 series GPUs also relies on SK Hynix and Samsung. SK Hynix's Q3 production capacity increase will help AMD expand the production capacity of MI300 and narrow the gap with NVIDIA. In addition, AMD is developing the MI400 series, which is expected to be equipped with HBM4, and SK Hynix's HBM4 production capacity plan will provide a guarantee for it.
3. Semiconductor Equipment Manufacturers: Benefiting from Production Capacity Expansion
The production of HBM requires advanced packaging technology, such as CoWoS (Chip-on-Wafer-on-Substrate) packaging. As the main supplier of CoWoS, TSMC's production capacity expansion will benefit from the growth of HBM demand. In addition, semiconductor equipment manufacturers such as Applied Materials and Lam Research, whose lithography machines, etching machines and other equipment are indispensable in HBM production, will benefit from the expansion of HBM production capacity.
IV. Investment Strategy: Focus on HBM Industry Chain Leaders
For US stock investors, the investment opportunities in the HBM storage track mainly focus on the following aspects:
- Upstream Material Manufacturers: The production of HBM requires high-purity silicon wafers, packaging materials, etc., such as Shin-Etsu Chemical, SUMCO, etc. The performance of these manufacturers will benefit from the expansion of HBM production capacity.
- Packaging Manufacturers: Packaging manufacturers such as TSMC and ASE, whose CoWoS production capacity expansion will directly benefit from the growth of HBM demand.
- HBM Suppliers: SK Hynix, Samsung, Micron, etc. The revenue growth of these manufacturers will directly reflect the prosperity of the HBM market.
- AI Chip Manufacturers: NVIDIA, AMD, etc. The improvement of their supply chain stability will enhance their product competitiveness and drive stock price increases.
It should be noted that the competitive landscape of the HBM market is changing. Samsung, SK Hynix, and Micron form a tripod, and more manufacturers may enter in the future. Therefore, it is necessary to pay attention to the technical advantages and production capacity expansion speed of each manufacturer. In addition, fluctuations in HBM prices may still affect the profit margins of related manufacturers, and investors need to closely follow the price trend.
V. Future Outlook: HBM4 and the Next Leap of AI Computing Power
As the parameter scale of AI large models continues to expand, the demand for HBM4 will gradually emerge. SK Hynix plans to mass-produce HBM4 in 2027, with bandwidth increased to 6.4TB/s, which will meet the needs of future AI large models. In addition, the power consumption issue of HBM is also a key concern of the industry. The current power consumption of HBM3E is about 300W, while HBM4 will reduce power consumption to below 250W through technological improvements, improving energy efficiency.
For US stock investors, the mass production of HBM4 will bring new investment opportunities. Manufacturers such as NVIDIA and AMD will be the first to adopt HBM4, and related manufacturers in their supply chains will benefit. In addition, the production of HBM4 requires more advanced packaging technology, such as 2.5D/3D packaging, which will drive the performance growth of packaging manufacturers.
In short, as a core component of AI computing power, the release of HBM storage capacity marks the gradual solution of the AI computing bottleneck. SK Hynix's Q3 pre-increase shows that the HBM market is transitioning from 'shortage' to 'tight balance', which will be beneficial to US AI chip leaders and the semiconductor sector. Investors should focus on the leading manufacturers in the HBM industry chain and seize this golden investment window.
