[Relevant Companies Mentioned in This Article] DapuStor, SK Hynix, YMTC, Micron, Samsung, TSMC (Taiwan Semiconductor Manufacturing Company), Longsys
The Super Cycle of Memory
Why Memory Is Surging So Dramatically
There has always been a question in the capital market: Is AI actually making money? Patel (founder of SemiAnalysis) directly stated that Anthropic is practically printing money. In the second quarter of this year, it has already achieved positive free cash flow and entered a profitable state. Among startup AI computing power companies, it is almost in a league of its own. Its Annual Recurring Revenue (ARR) has exceeded $50 billion, with a gross margin of over 70%. In his own words, "Anthropic is printing money." Furthermore, YipitData (a well-known US data analysis company) reported that its ARR has increased to $69 billion. Compared to May, its ARR in June saw a rapid month-over-month growth, with the daily average increase rising from about $400 million to as much as $550 million.
With AI large model services being so profitable, the hardware equipment in the computing power industry chain naturally follows suit. Patel's breakdown of the entire industry chain began as early as the end of 2024. At that time, after OpenAI released o1, it pushed the reasoning model into the AI Agent era for the first time. In the past, the context length of AI conversations was only a few thousand tokens, whereas in the era of reasoning models and Agents, the context length is easily over 100,000 tokens. The model reading weights have undergone a fundamental shift, and the original bottleneck has suddenly shifted from interconnects to memory (VRAM, main memory, and storage). The demand for VRAM and main memory from KV Cache (context cache) has experienced explosive growth. SemiAnalysis concluded at the time that future memory demand and profits will be far higher than the CPO (Co-Packaged Optics) sector for optoelectronic interconnects, making it the biggest winner in this round of the AI cycle.
SemiAnalysis also gave a clear answer to whether a 100% price increase in NAND, DRAM, and HBM in the market has already reached its peak: the market simply does not understand the supply and demand structure. Its analysis shows that over the next three years (2026-2029), the average annual growth in memory capacity will be limited to only 20%, and all new capacity will be high-end capacity such as HBM, DDR5, or DDR6. However, the demand from AI computing equipment is doubling, especially now in the stage of the large model competition, where manufacturers are flocking to make capital investments in AI. The supply and demand imbalance naturally leads to a rapid price increase. This supply and demand divergence is not only not narrowing but is intensifying.
On the other hand, the consumer electronics end is at the other extreme. Since global mainstream manufacturers have switched their existing capacity to advanced capacity, there is a severe shortage of mid-to-low-end capacity in HBM, DDR5/6, and high-capacity 3D NAND. The mobile phone, computer, and white goods industries are forced to raise prices and scramble for goods. Manufacturers with weak bargaining power will see their gross margins completely eroded if they do not raise prices, but raising prices will reduce terminal shipments by more than 30%. Patel also frankly stated that the prices of products like iPhones, Android phones, and computers have risen significantly by $100 or more. It seems that $100 has been "snatched" by AI, but in reality, $100 is far from enough for the grand AI computing power industry chain. In this round, memory prices have risen by nearly 400%, but there will still be a 2x to 3x increase in the future, and the gross profit margin of memory manufacturers will reach a staggering 80% or even 90%. For memory, a bulk commodity, prices and supply-demand relationships rely more on market determination, with great price elasticity and particularly severe fluctuations in the capital market.
The IPOs of SK Hynix and CXMT
Last week's big news in the AI industry chain was the US stock listing of SK Hynix and the upcoming IPO of domestic DRAM leader CXMT (ChangXin Memory Technologies). On Friday, July 10, in the US stock market, the South Korean memory giant SK Hynix quickly listed on the Nasdaq via ADR (American Depositary Receipts, which allow non-US companies to be traded by US investors in the US market in the form of transferable certificates). It closed up 12.76% on its first day, with a closing price of $168.01. During the trading session, it once rose nearly 19%, with its market value approaching $1.2 trillion. That day, it directly surpassed the $1.11 trillion market value of the US local memory giant Micron Technology, once exceeding it by $200 billion. In addition, the financing scale of this capital carnival was also unprecedented, surpassing the $21.8 billion set by Alibaba in 2014, reaching $26.5 billion, refreshing the historical high for foreign companies going public in the US, and ranking second only to SpaceX's $75 billion financing last month nationwide.
What the market pays more attention to is SK Hynix's nearly exaggerated financial report. In the first quarter of 2026, its operating revenue was 52.58 trillion Korean won, a year-on-year increase of 198%; net profit was 40.35 trillion Korean won, a year-on-year increase of 398%. Its net profit margin of 72% far exceeds Nvidia's 66% and TSMC's 58%. Moreover, with its global market share of HBM exceeding 55% and long-term agreement orders fully booked until 2028, maintaining an ultra-high profit margin is a sure thing, driving the entire market further upward.
Coincidentally, good news also came from the domestic capital market. As the strongest leading domestic DRAM company, CXMT will also enter the capital market recently. (The Semiconductor Industry Research Team previously published "CXMT: On the Eve of Dawn," which also analyzed in detail the contents of CXMT's prospectus, as well as its strength and positioning.) Before this round of the AI hardware cycle, the market was more discussing whether CXMT could survive. The DRAM industry had been in a situation of low prices and production cuts before 2025. CXMT's financials before 2025 showed a cumulative loss of about 36.6 billion RMB. However, the financial data for the first quarter of 2026 showed a single-quarter revenue of 50.8 billion RMB, with a net profit of about 33 billion RMB. Although the statistical calibers of the two are different and mechanical addition and subtraction cannot offset each other, it is obvious that they have capitalized on this super cycle of memory, and the company has also gained true self-sustaining capability for the first time.
What the capital market is even more optimistic about is that CXMT's growth is not due to the periodic premium brought by price increases, but rather the structural transformation behind it, including increased capacity and yield rates, and market share growth. The company's full-year revenue in 2025 was 61.7 billion RMB, a year-on-year increase of 155%. The shipping unit price of its main DDR series products increased by 60%, and the shipment volume, calculated by memory particle capacity, grew by 2.8 times. The revenue growth was even more impressive, at about 515%. In other words, a price increase for a single product is not enough to support the huge growth in revenue and the turnaround of net profit. It is more about the substantial increase in shipment volume and cost improvement after the mass production of DDR/LPDDR. Behind this is also the continuous blood supply of patient capital. According to the data, the Hefei state-owned capital behind CXMT holds a 36% stake. Between 2022 and 2025, Hefei state-owned capital provided a total of 6 billion RMB in government subsidies. Continuous subsidies allowed the enterprise not to be forced to exit during the long-term loss stage. Now, market improvements have given CXMT its voice back. Cash flow growth and the 29.5 billion RMB raised from the IPO on the STAR Market will further enhance its strength.
HBM Is the Real Vast Ocean
According to market news, CXMT has currently launched HBM3E, with the goal of sending samples by the end of 2026. Although there are various speculations in the market, the use of raised funds and future plans in the prospectus are the solid enterprise development trends. CXMT plans to invest all the raised funds in this IPO around its main business in three major directions: 7.5 billion RMB for the technical upgrading and transformation project of the memory wafer manufacturing mass production line, 13 billion RMB for the DRAM memory technology upgrading project, and 9 billion RMB for the forward-looking technology research and development project of dynamic random access memory. The final 9 billion RMB is considered by multiple media and industry insiders to be for building a new HBM production line.
At the same time, the global mobile phone giant Apple has also begun testing CXMT's DRAM, while actively communicating with the US government and commerce departments to minimize political risks. For CXMT, Apple's procurement is not just about adding a new customer and making money. Apple has always been known for its strictness in the industry. Passing its supply chain assessment and gaining more industry voice is the most core aspect. The DRAM bulk commodity industry competes on market share and voice, and progress in high-quality standards, high performance, and product consistency is equally important for CXMT. New orders and new quality evaluation standards are the necessary path for technological improvement. The funds raised from the IPO in the capital market and the profits from the memory market are the cornerstone for its future entry into the HBM track. By using the profits from products such as DDR and LPDDR to feed back into HBM, which is still in the critical stage, from firmly sitting in the fourth place globally to slowly starting HBM sampling, small-batch, and then mass production, the road is long and arduous, but CXMT presses on.
Conclusion
Returning to the title, the listing of the two giants should not be the peak yet. Battle reports may lie, but the battle line will not. Memory hardware prices remain high, and price increase letters from terminal manufacturers are coming one after another. This round of the "chip" shortage has just begun, and the acceleration of the AI hardware industry is just in time.
The illustrations in the article are generated by generative AI
References:
"Invest Like The Best" The Supply and Demand of AI Tokens | Dylan Patel Interview
"CXMT: On the Eve of Dawn"
"Is CXMT's IPO the Peak? No Matter How Extreme the Current Game Is, It Cannot Stop It from Continuing to Grow Stronger"