Summary of Key Points
On the day of the subscription for ChangXin Technology, the storage sector experienced a collective slump (institutional investors sold off their stocks to free up funds for the new share offering; DeMingLi's performance fell short of expectations, exacerbating market sentiment; and the new shares had a dilutive effect on existing stock prices). However, short-term price fluctuations are not the real focus. The crucial aspect is ChangXin's HBM (High Bandwidth Memory) technology: it is the "supply pipeline" for AI computing power and determines whether China's AI industry can break free from foreign dependence. Currently, ChangXin has sent samples of its HBM3 to Huawei, but there is a generation gap compared to Samsung and SK Hynix in terms of yield and mass production progress. The project is supported by the Hefei government's "industrial investment strategy," but with giants accelerating their own production capacity expansion, ChangXin only has about 5 years left to catch up. Additionally, the new shares face risks such as a small circulating supply and valuation that is highly sensitive to market cycles.
Analysis 1: Who Was Behind the Storage Sector's Slump on the Subscription Day?
The decline in the storage and semiconductor sectors was not due to retail investors fleeing the market; rather, it was institutional investors moving their assets:
- Institutional Fundraising for the New Share Offering: ChangXin is the second-largest IPO on the STAR Market, and offline subscriptions require a large amount of capital to be frozen. Institutions had to sell their existing storage stocks to raise the necessary funds, leading to the "dilutive effect" on the market.
- Emotional Outbreak: DeMingLi's performance forecast a few days prior fell short of expectations, already putting pressure on the sector. This triggered panic, causing the entire industry chain to decline.
- Irony of the Situation: A company in the storage industry was raising funds but ended up driving down its own sector's stock price. The combination of negative sentiment and financial pressure resulted in more than 4,000 stocks turning red.
Analysis 2: Why Is HBM the "Lifeline" for AI?
HBM may sound complex, but it essentially acts as a high-speed data conduit for AI chips:
- The Critical Role of HBM: If you disassemble a H100 GPU, the most expensive component is not the chip itself but the stacked HBM, which can transfer 1TB of data per second (whereas traditional memory can only transfer a few hundred GB). A narrow data pipeline means even a powerful GPU will be limited in its computing power, leading to bottlenecks during large model training and inference (referred to as the "video memory wall" in the industry).
- High Cost Impact: HBM accounts for 40%-50% of the material cost in a high-end GPU. Nearly half of the profit from selling a GPU goes to memory manufacturers.
- Foreign Dependency: The US has restricted the export of high-end HBM to China, making it essential for Chinese AI companies to develop their own production capabilities.
Analysis 3: How Close Is ChangXin to Becoming World Leader in HBM?
ChangXin is not just talking the talk; there are still significant gaps:
- Progress Made: Samples of the 16-nanometer HBM3 have been sent to Huawei for joint testing, indicating that the interface and architecture meet international standards and can be integrated with Huawei's Ascend chips. The company plans to start mass-producing the 12-layer HBM3E by 2027.
- Yield Challenges: There is a big difference between sample production and mass production. While a 99% yield for each layer results in an 88% overall yield after stacking, a 95% yield only yields 54% for 12 layers. Samsung and SK Hynix have already started mass-producing the 12-layer HBM4, while ChangXin will only start in 2027, lagging by about two generations.
- Packaging Constraints: After producing the HBM, it needs to be packaged using TSMC's CoWoS technology. However, NVIDIA owns most of the available capacity for this process, so even if ChangXin meets its yield targets, it will still face delays in getting its products packaged.
Analysis 4: Hefei's Support, but a Limited Time Window
ChangXin is part of Hefei's strategy to invest in "hard technology" (with BOE and NIO being the previous success stories):
- How Hefei Is Supporting: The Hefei government has led an investment of 49.4 billion yuan, working with large funds and banks to support ChangXin's growth from its inception until it became the second-largest IPO on the STAR Market.
- Time Pressure: Samsung and SK Hynix plan to double their DRAM production capacity within five years, and their HBM production expansion is even more aggressive. By the time ChangXin starts mass-producing the HBM3E in 2027, their HBM4 may have become widely available, potentially driving down prices due to increased supply.
- Challenges of the Hefei Model: The storage industry requires more sophisticated manufacturing processes than the display or automotive sectors (with yield improvements coming from thousands of trial productions). While patience can help manage initial losses, physical limitations cannot be overcome easily.
Analysis 5: Risks Associated with the New Shares
Don't Just Focus on the "5x PE Ratio"
ChangXin's issue price is 8.66 yuan, with a dynamic PE ratio of around 5 (similar to Samsung and SK Hynix). However, several risks should be considered:
- Valuation Sensitivity: The 5x PE ratio is based on current high DRAM prices. If prices fall in the second half of the year, the valuation will decrease significantly.
- Limited Circulating Supply and Volatility: There were no price limits in the first five days of trading, meaning the stock price could soar or plummet quickly (for example, some new STAR Market stocks rose 200% only to drop by half on the third day).
- Potential Pullback: Once price limits are reinstated on the sixth day, if the stock has been overbid, it may experience consecutive daily limit-downs, making it difficult to sell.
Final Conclusion
Short-term price movements are interesting, but what really matters is whether ChangXin can successfully mass-produce the 12-layer HBM3E by 2027 with the required yield. This is not only a critical test for ChangXin's success but also for the stability of China's AI computing infrastructure.
(The entire analysis avoids technical jargon and uses everyday language to explain complex concepts, covering short-term fluctuations, technological fundamentals, progress gaps, capital support, and risk factors, making it accessible to non-experts.)