U.S. stocks closed in the early hours of the morning, with the Philadelphia Semiconductor Index (SOX) breaking through 14,000 points for the first time, setting a new all-time high.

Historically, SOX has only experienced two periods of growth exceeding 230% within 14 months: from December 1998 to February 2000, and from April 2025 to the present.

The returns from this semiconductor bull market have been highly concentrated and significant. The three memory giants, Micron, SK Hynix, and Samsung, have seen year-to-date gains of approximately 141%, 186%, and 114%, respectively. TSMC's US-listed ADRs have risen by over 50% year-to-date.

Nvidia hit an all-time high of $235.47 on May 14. Broadcom, Marvell, and ASML all refreshed or approached records in their respective segments. The SOXX ETF as a whole had a 52-week low of $148 and a high of nearly $369, representing a range of almost 150%.

In April, Goldman Sachs raised its 2026 DRAM supply-demand gap forecast from 3.3% to 4.9%, calling it the most severe memory shortage in 15 years. HBM prices are even more dramatic, with HBM3E stacks costing around $300 per chip, and the soon-to-be-mass-produced HBM4 estimated at $500 per chip. Hynix's 2026 HBM capacity has already been fully booked by Microsoft, Google, and Nvidia, with some customers even paying full deposits in advance to secure the capacity.

Clearly, the construction of AI data centers is progressing much faster than the expansion of chip production capacity.

A bull market that is "strangled"

Scarcity makes products the most profitable.

Understanding this statement essentially reveals the core logic behind this semiconductor bull market. Whoever controls the bottleneck of AI infrastructure gains the most powerful pricing power. Conversely, whoever's segment can be replaced or their prices can be lowered, no matter how high the demand, their stock price will not rise.

Optical modules are a prime example of the latter. A Photon Capital report in April pointed out that while Chinese optical module manufacturers occupy seven of the top ten spots globally, they haven't made much money; instead, chip companies are the ones profiting. Companies like InnoLight Technology and Eoptolink have achieved world-class levels in terms of shipments and cost control for 800G and 1.6T optical modules, directly squeezing the profit margins of US-listed optical module companies like Coherent and Lumentum. Demand has doubled, yet profit margins have been squeezed. The reason is simple: the assembly stage of optical modules is not scarce enough.

Storage has become the most compelling theme in the US semiconductor sector this round. Essentially, this is because the bottleneck is being tightened, and the bottleneck is only getting tighter.

HBM is not ordinary DRAM. 3D stacking, TSV through-silicon vias, and specialized packaging processes—each technological barrier represents the result of over a decade of heavy investment. Globally, only three companies can mass-produce HBM, with SK Hynix holding approximately half the market share.

Interestingly, this logic also holds true when applied to the macro-level of the nation.

The real winners in AI data center infrastructure are not "all semiconductor countries," but rather those countries and regions that have built scarce industrial clusters in a particular irreplaceable segment over the past few years or even decades. Scarcity is the key.

Each region has its own main track.

I saw someone raise this point in the US stock community, and it's very interesting.
The United States still stands at the very top of the value chain.

Nvidia, AMD, and Broadcom provide ASIC designs; Synopsys and Cadence offer EDA tools; and Arista provides AI networks. The three major cloud providers package computing power as a service and sell it worldwide. Google, Amazon, and Microsoft are all accelerating their own ASIC development. Broadcom and Marvell together hold approximately 95% of the custom ASIC design market share, with Google alone spending about $8 billion annually on TPU development for Broadcom.

The core manufacturing nodes are in Taiwan and South Korea, but they operate on completely different models.

Meanwhile, in Taiwan, the focus is on TSMC and advanced packaging. Globally, only TSMC can mass-produce 3nm and 2nm processes. TSMC's three CoWoS back-end factories are operating at full capacity, with lead times ranging from 52 to 78 weeks. Nvidia alone has secured 60% to 70% of the CoWoS capacity. TSMC is expanding its monthly capacity from 35,000 wafers at the end of 2024 to 130,000 wafers at the end of 2026, nearly quadrupling. Even with this significant expansion, capacity remains tight. Taiwan's server foundry system, including Foxconn, Quanta, and Wistron, is also ramping up shipments along with AI server demand.

South Korea's story revolves entirely around memory. SK Hynix holds approximately 50% to 55% of the global HBM market share, Samsung 19% to 35%, and Micron approximately 5% to 20%. HBM is not the same as regular memory; 3D stacking, TSV (Through Silicon Vias), and specialized packaging processes—each technological barrier is the result of South Korean companies' continuous investment over the past decade.

Japan and the Netherlands also play important roles. Tokyo Electron manufactures semiconductor equipment, Shin-Etsu Chemical and SUMCO produce silicon wafers, and Ajinomoto manufactures ABF substrate materials. Japan has long been out of the running in the competition for chip end products, but its position in materials and precision machining remains irreplaceable to this day.

The situation is even more direct in the Netherlands, where ASML monopolizes EUV lithography machines. In January, Morgan Stanley significantly raised its target price for ASML to €1400, predicting 2027 would be ASML's highest profit growth year, with EPS increasing by 57% year-on-year. They based this assessment on three driving forces: better-than-expected expansion of advanced logic foundry capacity, large-scale expansion in the DRAM memory sector, and better-than-expected overall demand. Dutch packaging equipment companies like BESI have also secured a large number of orders amid the surge in demand for AI chip packaging.

China and Europe have different entry points, but the logic is similar: both have established cost advantages or delivery capabilities in a specific aspect of AI infrastructure.

InnoLight Technology and Eoptolink Technology are among the world's leading companies in terms of shipment volume and price control for 800G and 1.6T optical modules. However, Photon Capital's analysis also highlights a crucial time window: the current high profit margins of optical module companies stem from temporary pricing power resulting from a temporary shortage of 800G capacity. Once 1.6T mass production begins in the second half of 2026 to 2027, and second- and third-tier manufacturers fill their capacity gaps, price pressure on the module side will quickly materialize.

In Europe, companies like Schneider Electric, ABB, and Vertiv, which specialize in power distribution and cooling, have received far more orders than expected amid a surge in data center electricity consumption. Wedbush estimates that hyperscaler AI infrastructure spending will reach approximately $725 billion in 2026, a 77% year-on-year increase, with power infrastructure being one of the fastest-growing sub-items.

AI is reshaping the semiconductor "smile curve".

If we use a smile curve to summarize this diagram: the United States on the left is responsible for "definition and design", Taiwan, South Korea, the Netherlands, and Japan in the middle-high section are responsible for "manufacturing advanced chips", Taiwan, China, and Southeast Asia in the middle-low section are responsible for "mass assembly", and the United States and China on the right are responsible for "cloud platforms, models, and customer access".

The originator of this curve is Stan Shih, the founder of Acer. In 1992, he used this model to explain why PC assembly had the lowest profit margin.

But thirty years later, AI data centers are rewriting the shape of this curve.

Both FourWeekMBA's value chain analysis and a paper published this year by Atlantis Press point to the same conclusion: AI has re-elevated the middle segment of the traditional manufacturing smile curve. TSMC's advanced packaging CoWoS, SK Hynix's HBM stacking, and ASML's EUV lithography machines—these segments in the traditional manufacturing smile curve represent the thinnest-profit "middle manufacturing segment." However, in the AI era, they have become the scarcest resources, with profit margins and pricing power no lower than those in the design and application sectors.

The data in the paper shows that Nvidia's gross margin was 72.72% and net margin was 48.85% in 2023 and 2024. However, TSMC's gross margin also reached 66.2% and net margin 50.5% in Q1 2026. The profit margin gap between the design and manufacturing ends is narrowing, which is unprecedented in the history of the semiconductor industry.

The traditional smile curve suggests that the manufacturing stage has the lowest profit margin. AI, however, transforms the most difficult manufacturing stage into the scarcest resource.

Morgan Stanley's Asian semiconductor research report in March summarized a similar conclusion: the AI cycle from 2023 to 2024 will mainly focus on GPUs, and from 2025 to 2026, demand will begin to spread to a wider range of industries, including storage, advanced packaging, custom ASICs, and data center networks.

Each cycle of bottlenecks brings a batch of previously overlooked companies to the forefront, while allowing the stocks that saw the biggest gains in the previous round to enter a period of digestion.

How far can the bulls run? A battle between bullish and bearish opinions.

Let's first hear from the bulls. In May, Wedbush's Dan Ives directly predicted on CNBC that the Nasdaq would reach 30,000 points within the next year, arguing that demand for AI chips still far exceeds supply. Goldman Sachs provides more specific figures: global AI capital expenditure will be approximately $765 billion in 2026, climbing to $1.6 trillion by 2031.

In its Asian semiconductor research report released in March, Morgan Stanley clearly stated that AI computing power investment is still in the expansion phase, and the semiconductor industry is entering a new structural demand cycle.

Bullish views on storage are becoming more aggressive. Goldman Sachs recently lowered its DRAM supply-demand gap forecasts for 2026-2028 to a deeper shortage range, revising its 2027 forecast from -2.5% to -5.9%, almost doubling it. Their assessment is that this storage cycle is different from the past; demand for AI servers is more visible, supply growth is constrained by long-term lock-in agreements, and the price increase will last longer than the market expects.

Goldman Sachs even raised its operating profit forecasts for Kioxia by 16% to 48% for the three years from 2027 to 2029, arguing that this round of high profits could last for two to three years. Giving a prediction of "high profits lasting three years" for a company in a highly cyclical business like storage is extremely rare on Wall Street.

Morgan Stanley's change of attitude is even more interesting. In 2024, they were still predicting a "DRAM winter," forecasting that prices would fall for many years starting in Q4 of 2024. But by 2025, they had completely reversed their stance and adopted a supercycle theory, predicting that DRAM prices would rise by 62% in 2026, and that SK Hynix and Samsung's profits would exceed consensus expectations by 30% to 50%.

But the short sellers are also quite vocal, and they come from a powerful background.

In May, Michael Burry publicly warned that this semiconductor rally was highly similar to the final months of the dot-com bubble in 1999-2000. SOX is up 65% year-to-date, with weekly gains of 10%, and the SOXX ETF is 60% above its 200-day moving average—a level of technical upward momentum rarely sustained in history. SEC filings show he has purchased a large number of put options on SOXX, QQQ, Nvidia, Palantir, and Oracle, expiring in January 2027 with strike prices significantly lower than current share prices.

In June, Man Group (one of the world's largest publicly traded hedge funds) published a lengthy article specifically addressing the risks of an AI bubble. Their core argument is that the financial architecture surrounding AI has become too large, over-leveraged, and overly reliant on a few interconnected participants.

They specifically mentioned that a large amount of AI data center construction was financed through private credit, and the collateral for these loans was "hardware that depreciates rapidly like a mobile phone, rather than long-term assets like buildings." The first wave of defaults may occur in 2027 or 2028, when the initial leases expire, and the gap between financing assumptions and reality will become unavoidable.

Looking ahead, there are several key milestones worth noting.

Micron will release its earnings report on June 24th, and its forward guidance on HBM demand and capacity allocation will determine the direction of the memory sector throughout the summer. Nvidia's next earnings report is equally crucial; even a slight slowdown in AI chip demand will trigger a repricing of the entire sector's sentiment.

Looking further ahead, the timeline for capacity release is the real watershed. SK Hynix's M15X plant is expected to ramp up production in mid-2027, while Yongin's new plant is projected to begin production in February 2027. Samsung's P5 plant will begin production in 2028. Micron's Idaho Fab 1 is expected to contribute output in mid-2027.

All of these factors combined will increase industry capacity by 20% to 30% from the second half of 2027 to the first half of 2028. The problem is that the compound annual growth rate of HBM demand is also over 40%. Whether supply can keep up with demand depends on whether AI capital expenditure slows down before then.

The final variable is geopolitics. The higher the concentration of the semiconductor supply chain, the greater the impact of black swan events. TSMC alone accounts for over 90% of global advanced process foundry services. In a bull market, this represents efficiency; in a conflict scenario, it becomes a systemic risk. Factors such as the Taiwan Strait, the escalation of US export controls on China, and the degree of cooperation between Japan and the Netherlands on equipment controls are factors that no one wants to discuss when the market is good, but once something happens, pricing in these factors will be faster than any fundamental changes.