The US stock market closed in the early morning, with the Philadelphia Semiconductor Index SOX breaking through 14,000 points for the first time, setting a historical high.
Historically, there have only been two periods when SOX rose more than 230% within 14 months: from December 1998 to February 2000, and from April 2025 to now.
The returns in this round of the semiconductor bull market are highly concentrated and significant. The year-to-date increases for the top three memory companies—Micron, SK Hynix, and Samsung—have reached approximately 141%, 186%, and 114%, respectively. TSMC's US ADR has risen more than 50% year-to-date.
NVIDIA hit a historical high of $235.47 on May 14. Broadcom, Marvell, and ASML have all refreshed or approached records in their respective segments. The 52-week low for the entire SOXX ETF is $148, with a high close to $369, showing a swing of nearly 150%.
Goldman Sachs raised its forecast for the 2026 DRAM supply-demand gap from 3.3% to 4.9% in April, calling it the most severe memory shortage in 15 years. The price of HBM is even more exaggerated, with the HBM3E stacking at approximately $300 per unit, and the soon-to-be-mass-produced HBM4 estimated at $500 per unit. SK Hynix's HBM capacity for 2026 has long been fully reserved by Microsoft, Google, and NVIDIA, with some customers even paying full deposits in advance to secure production capacity.
Clearly, the speed of AI data center construction is far outpacing the speed of chip capacity expansion.
"Choke Point" Bull Market
Scarcity is the most profitable product.
Understanding this statement essentially reveals the core logic of this round of semiconductor bull market. Whoever controls the choke point of AI infrastructure holds the hardest pricing power. Conversely, whoever's link can be replaced or pressured will see their stock prices struggle to rise, regardless of demand.
Optical modules are a typical example of the latter. Photon Capital's report in April pointed out that seven out of the top ten global optical module positions are held by China, yet they haven't made much money, with the profits still going to chip companies. Zhongji Xuchuang and Newyeasen have achieved global leading levels in shipment volumes and cost control for 800G and 1.6T optical modules, directly squeezing the profit margins of US optical module companies like Coherent and Lumentum. Demand has doubled, but profit margins have thinned out. The reason is simple: the assembly process of optical modules is not scarce enough.
On the other hand, storage has become the hardest main line in this round of the US stock semiconductor market. Essentially, it is because it holds the choke point, and the choke becomes tighter.
HBM is not ordinary DRAM. 3D stacking, TSV silicon via, specialized packaging processes—each layer of technological barrier is the result of over a decade of heavy asset investment. Only three companies globally are capable of mass-producing HBM, with SK Hynix taking about half of the market share.
Interestingly, this logic applies at the macro national level as well.
The real winners of AI data center infrastructure are not "all semiconductor countries," but rather those countries and regions that have established scarce industrial clusters in certain irreplaceable links over the past few years or even decades. Scarcity is the key.
Every Region Has Its Own Main Track
It is quite interesting to see someone propose this viewpoint in the US stock community.
The United States remains at the top of the value chain.
NVIDIA, AMD, and Broadcom's ASIC designs, Synopsys and Cadence's EDA tools, and Arista’s AI networks—all three major cloud companies package computing power as a service for the whole world. Google, Amazon, and Microsoft are all accelerating self-developed ASICs. Broadcom and Marvell have collectively captured about 95% of the custom ASIC design market, with just Google spending about $8 billion each year on TPU development with Broadcom.
The core nodes in manufacturing are in Taiwan and Korea, but they are eating completely different foods.
On the Taiwan side, it revolves around TSMC and advanced packaging. Only TSMC can achieve mass production at the 3nm and 2nm process nodes. All three of TSMC's CoWoS backend factories are fully loaded, with lead times of 52 to 78 weeks, and NVIDIA alone has secured 60% to 70% of CoWoS capacity. TSMC is expanding its monthly capacity from 35,000 wafers by the end of 2024 to 130,000 wafers by the end of 2026, nearly quadrupling it. However, even with this much expansion, capacity remains tight. Taiwan's server foundry system, including Hon Hai, Quanta, and Wistron, is also ramping up alongside AI server shipments.
The Korean story revolves entirely around storage. SK Hynix has about 50% to 55% of the global HBM market share, Samsung holds about 19% to 35%, and Micron approximately 5% to 20%. HBM is not the same as ordinary memory; the 3D stacking, TSV silicon via, and specialized packaging processes represent years of sustained investment by Korean companies.
The roles of Japan and the Netherlands are also crucial. Tokyo Electron manufactures semiconductor equipment, Shin-Etsu Chemical and SUMCO produce silicon wafers, and Ajinomoto produces ABF substrate materials. Japan has long exited the competition in chip end products, but its position in materials and precision processing remains irreplaceable to this day.
As for the Netherlands, it is even more direct, with ASML monopolizing EUV lithography machines. In January, Morgan significantly raised its target price for ASML to 1,400 euros, predicting 2027 will be ASML's year of highest profit growth, with EPS expected to grow by 57% year-on-year. They base this judgment on three driving forces: greater-than-expected expansion of advanced logic foundry capacity, large-scale expansions in the DRAM storage sector, and overall demand performance exceeding expectations. Dutch packaging equipment companies like BESI have also secured numerous orders in the explosion of AI chip packaging demand.
China and Europe approach things differently, but the logic is similar; both have established cost advantages or delivery capabilities in specific links of AI infrastructure.
Zhongji Xuchuang and Newyeasen have achieved global leading levels in shipment volumes and price control for 800G and 1.6T optical modules. However, Photon Capital's analysis also warns of an important time window: the current high profit margins of optical module companies stem from the temporary pricing power due to the phased shortage in 800G capacity. Once 1.6T enters mass production in the second half of 2026 and second- and third-tier companies catch up with capacity, pricing pressures on the module side will quickly arise.
In Europe, companies like Schneider Electric, ABB, and Vertiv, which specialize in power distribution and cooling, have received far more orders than expected amidst a surge in electricity consumption in data centers. Wedbush estimates that in 2026, hyperscalers' AI infrastructure spending will be about $725 billion, an increase of 77% year-on-year, with power infrastructure being one of the fastest-growing sub-items.
AI Reshapes the Semiconductor "Smile Curve"
If we summarize this graph with a smile curve: the US at the left end is responsible for "definition and design," Taiwan, Korea, the Netherlands, and Japan in the middle are responsible for "manufacturing advanced chips," Taiwan, China, and Southeast Asia handle "mass assembly" at a lower middle level, and the US and China at the right end manage "cloud platforms, models, and customer entry points."
The curve's originator is Acer founder Stan Shih, who used this model in 1992 to explain why PC assembly profits are the thinnest.
But thirty years later, AI data centers are rewriting the shape of this curve.
Value chain analyses from FourWeekMBA and a paper published by Atlantis Press this year point to the same conclusion: AI has raised the middle segment of the traditional smile curve. TSMC's advanced packaging CoWoS, SK Hynix's HBM stacking, and ASML's EUV lithography machines—these links, which traditionally belonged to the thinnest profit "middle manufacturing segment," have become the most scarce resources in the AI era, with profit margins and pricing power not lower than those at the design and application ends.
Data from the paper shows that NVIDIA's gross margin for 2023 to 2024 is 72.72%, and the net margin is 48.85%. Meanwhile, TSMC's gross margin for Q1 2026 has reached 66.2%, with a net margin of 50.5%. The profit margin gap between design and manufacturing is shrinking, which is unprecedented in the history of the semiconductor industry.
The traditional smile curve suggests that the manufacturing segment has the thinnest profit. AI has transformed the most difficult manufacturing segment into the scarcest resource.
A Morgan report from March summarizes similar conclusions: the AI cycle from 2023 to 2024 is primarily focused on GPUs, while from 2025 to 2026, demand begins to spread to a broader industry chain, with storage, advanced packaging, custom ASICs, and data center networks taking over.
Each cycle of bottleneck rotation pushes a batch of previously overlooked companies to the forefront while allowing those that surged the most in the previous round to enter a digestion phase.
How Far Can the Bull Run? Bull and Bear Perspectives
Let’s first hear from the bulls. Dan Ives from Wedbush directly stated on CNBC in May that he sees the Nasdaq hitting 30,000 points in the next year, citing that AI chip demand still far exceeds supply. Goldman Sachs provided more specific figures, forecasting global AI capital expenditure to be about $765 billion in 2026, rising to $1.6 trillion by 2031.
Morgan's Asian semiconductor report from March explicitly states: AI computing power investment is still in the expansion phase, and the semiconductor industry is entering a new structural demand cycle.
The bullish sentiment on storage is even more aggressive. Goldman recently adjusted its 2026 to 2028 DRAM supply-demand gap predictions down to deeper shortage ranges, revising 2027 from -2.5% to -5.9%, nearly doubling it. Their judgment is that this storage cycle is different from previous ones; the visibility of AI server demand is much higher, supply growth is locked down by long-term contracts, and the duration of price increases will be longer than the market expects.
Goldman even gave Kioxia significant upward adjustments to their operating profit forecasts for the three years from 2027 to 2029, ranging from 16% to 48%, reasoning that this cycle of high profits can last for two to three years. For a company engaged in such a cyclical storage business, forecasting "high profits for three years" is quite rare on Wall Street.
Morgan's change of attitude is even more interesting. They were still predicting a "DRAM winter" in 2024, forecasting prices to decline from the fourth quarter of 2024 for several years. However, by 2025, they had flipped to the super cycle theory, predicting a 62% rise in DRAM prices in 2026, with earnings from SK Hynix and Samsung exceeding consensus expectations by 30% to 50%.
However, there are also significant bearish voices, and some high-profile ones.
Michael Burry publicly warned in May that the current semiconductor market resembles the last few months of the 1999 to 2000 dot-com bubble. The SOX has risen 65% this year, with a weekly increase of 10%, and the SOXX ETF is 60% above its 200-day moving average; this level of technical stretching is rarely sustainable in history. SEC holdings disclosures show he has bought a large amount of put options on SOXX, QQQ, NVIDIA, Palantir, and Oracle, with expiration in January 2027 and strike prices far below current stock prices.
Man Group, one of the largest listed hedge funds globally, published a lengthy article in June specifically dissecting AI bubble risks. Their core viewpoint is that the financial structure around AI has become oversized, over-leveraged, and excessively reliant on a handful of interrelated participants.
They specifically noted that massive AI data center constructions are being financed through private credit, and the collateral for these loans consists of "hardware that depreciates as quickly as a smartphone, rather than long-term assets like buildings." The first wave of defaults may occur between 2027 and 2028 when initial leases expire, and the gap between financing assumptions and reality will become unavoidable.
Looking ahead, several key time points are worth paying attention to.
Micron will announce its earnings report on June 24, providing forward guidance on HBM demand and capacity allocation that will determine the direction of the storage sector for the entire summer. NVIDIA's next earnings report is also critical; if there is even a slight sign of deceleration in AI chip demand, the sentiment across the sector will be re-evaluated.
Further out, the timeline for capacity release is the real turning point. SK Hynix's M15X factory is expected to ramp up production by mid-2027, with the Yongin new plant moving forward to February 2027. Samsung's P5 factory will go into production in 2028. Micron's Idaho Fab 1 is expected to contribute output by mid-2027.
All these factors combined may lead to an industry capacity increase of 20% to 30% in the second half of 2027 to the first half of 2028. The question is whether HBM demand can keep up with this; the compound annual growth rate of demand is also above 40%. Whether supply can catch up with demand depends on whether AI capital expenditures slow down before that.
The final variable is geopolitics. The more concentrated the semiconductor supply chain, the greater the impact of a black swan event will be. TSMC alone accounts for over 90% of global advanced process foundry, which in a bull market represents efficiency but becomes systemic risk in conflict scenarios. The escalation paths of the Taiwan Strait, US export controls on China, and the degree of cooperation between Japan and the Netherlands regarding equipment controls are factors that no one wants to discuss when the market is doing well, but once an issue arises, pricing will adjust faster than any change in fundamentals.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
