U.S. stocks closed overnight, with the Philadelphia Semiconductor Index (SOX) surpassing 14,000 points for the first time, setting a new all-time high.

Historically, SOX has risen more than 230% over a 14-month period only twice: from December 1998 to February 2000, and from April 2025 to the present.

The returns from this round of the semiconductor bull market have been highly concentrated and significant. This year, the top three memory manufacturers—Micron, SK Hynix, and Samsung—have seen year-to-date gains of approximately 141%, 186%, and 114%, respectively. TSMC’s U.S. ADR has risen over 50% this year.

NVIDIA reached a historic high of $235.47 on May 14. Broadcom, Marvell, and ASML have all set or neared their respective record highs. The SOXX ETF's 52-week low was $148, and its high approached $369, representing a range 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. HBM prices are even more extreme: a single HBM3E stack costs approximately $300, and the upcoming HBM4 is estimated at $500 per unit. SK Hynix’s 2026 HBM capacity has already been fully reserved by Microsoft, Google, and NVIDIA, with some customers even paying full upfront deposits to secure production slots.

Clearly, the pace of AI data center construction far exceeds the rate of chip production capacity expansion.

A bull market with supply constraints

Scarcity is the most profitable product.

Understanding this sentence is essentially understanding the core logic behind this semiconductor bull market. Whoever controls the bottleneck in AI infrastructure holds the strongest pricing power. Conversely, whoever’s segment can be substituted or pressured on price—even with massive demand—will see little to no stock price appreciation.

Optical modules are a classic example of the latter. According to Photon Capital’s April report, Chinese optical module manufacturers hold seven of the top ten spots globally, yet they earn very little profit—while chip companies reap the majority of the profits. Companies like InnoLight and Eoptolink have achieved world-class levels in shipping volume and cost control for 800G and 1.6T optical modules, directly squeezing the profit margins of U.S.-listed optical module firms such as Coherent and Lumentum. Despite demand doubling, profit margins have been compressed. The reason is simple: the assembly segment of optical modules is not scarce.

Storage has become the strongest and most resilient theme in this round of U.S. semiconductor stocks, essentially because the supply chain bottleneck has tightened further.

HBM is not ordinary DRAM. Each layer of technological barrier—3D stacking, TSV through-silicon vias, and specialized packaging processes—is the result of decades of heavy capital investment. Only three companies worldwide can currently mass-produce HBM, and SK Hynix holds approximately half of the market share.

Interestingly, this logic holds true when scaled up to the macro level of nations.

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

Each region has its own main track.

I came across this viewpoint in a U.S. stock market community, and it's very interesting.
The United States remains at the top of the value chain.

NVIDIA, AMD, and Broadcom design ASICs; Synopsys and Cadence provide EDA tools; Arista delivers AI networking; the three major cloud providers bundle computing power into services sold worldwide. Google, Amazon, and Microsoft are all accelerating their own ASIC development. Broadcom and Marvell together hold approximately 95% of the custom ASIC design manufacturing market, with Google alone spending about $8 billion annually on TPU development at Broadcom.

The core nodes in manufacturing are in Taiwan and South Korea, but they are operating in entirely different sectors.

On this side, the focus is on TSMC and advanced packaging. Only TSMC can mass-produce 3nm and 2nm processes globally. All three of TSMC’s CoWoS backend factories are operating at full capacity, with lead times of 52 to 78 weeks; NVIDIA alone has secured 60% to 70% of CoWoS capacity. TSMC is expanding its monthly production from 35,000 wafers at the end of 2024 to 130,000 wafers by the end of 2026—nearly a fourfold increase. Yet even with this massive expansion, capacity remains tight. Taiwan’s server contract manufacturing ecosystem, including Hon Hai, Quanta, and Wistron, is also scaling up in tandem with the surge in AI server shipments.

South Korea’s story is entirely centered on storage. SK Hynix holds approximately 50% to 55% of the global HBM market share, Samsung holds 19% to 35%, and Micron holds about 5% to 20%. HBM is not the same as conventional memory—3D stacking, TSV silicon vias, and specialized packaging processes represent layers of technological barriers that Korean companies have consistently invested in over the past decade.

Japan and the Netherlands also play crucial roles. Tokyo Electron manufactures semiconductor equipment, Shin-Etsu Chemical and SUMCO produce silicon wafers, and Ajinomoto makes ABF substrate materials. Although Japan has long exited the competition in final chip products, its position in materials and precision processing remains irreplaceable to this day.

The Netherlands is even more direct: ASML monopolizes EUV lithography machines. In January, Morgan Stanley significantly raised its target price for ASML to €1,400, forecasting that 2027 will be the year with the highest profit growth for ASML, with EPS increasing by 57% year-over-year. This outlook is based on three key drivers: stronger-than-expected expansion in advanced logic foundry capacity, large-scale capacity increases in the DRAM storage sector, and overall demand outperforming expectations. Dutch packaging equipment companies like BESI have also secured substantial orders amid the surge in AI chip packaging demand.

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

Infinera and Eoptolink have global leading-level shipment volumes and pricing power in 800G and 1.6T optical modules. However, Photon Capital’s analysis also highlights a critical time window: the current high profit margins of optical module companies stem from temporary pricing power driven by阶段性 shortages in 800G production capacity. Once 1.6T mass production ramps up in the second half of 2026 through 2027, and secondary and tertiary manufacturers expand their capacity, pricing pressure on modules will quickly emerge.

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

AI is reshaping the semiconductor "smile curve"

If we summarize this chart using the smile curve: the U.S. on the left handles "definition and design," Taiwan, South Korea, the Netherlands, and Japan in the higher middle segment handle "manufacturing advanced chips," Taiwan, China, and Southeast Asia in the lower middle segment handle "large-scale assembly," and the U.S. and China on the right handle "cloud platforms, models, and customer access."

The creator of this curve is Stan Shih, founder of Acer, who in 1992 used this model to explain why PC assembly yields the lowest profits.

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 smile curve. Components such as TSMC’s advanced packaging CoWoS, SK Hynix’s HBM stacking, and ASML’s EUV lithography machines—once considered the lowest-margin “midstream manufacturing” segment in the traditional manufacturing smile curve—have become the most scarce resources in the AI era, with profit margins and pricing power no lower than those at the design or application ends.

The data shows that NVIDIA's gross margin from 2023 to 2024 was 72.72%, with a net margin of 48.85%. However, TSMC's gross margin in Q1 2026 also reached 66.2%, with a net margin of 50.5%. The gap in profitability between the design and manufacturing segments is narrowing—a phenomenon unprecedented in the history of the semiconductor industry.

The traditional smile curve holds that manufacturing has the lowest profit margins. AI has turned the most challenging manufacturing step into the most scarce resource.

Morgan’s March report on Asian semiconductors concluded similarly: the AI cycle from 2023 to 2024 was primarily focused on GPUs; starting in 2025 to 2026, demand began to spread across a broader supply chain, with memory, advanced packaging, custom ASICs, and data center networking taking the lead.

Each round of bottleneck rotation brings previously overlooked companies into the spotlight, while pushing the top performers from the previous round into a consolidation phase.

How much further can the bull run? The battle between long and short views

Let’s first hear from the bulls. In May, Dan Ives of Wedbush directly stated on CNBC that the Nasdaq could reach 30,000 points within the next year, citing that demand for AI chips still far exceeds supply. Goldman Sachs provided even more specific figures, estimating global AI capital expenditures at approximately $765 billion in 2026, rising to $1.6 trillion by 2031.

In its March Asia semiconductor report, Morgan clearly stated: AI computing power investment remains in an expansion phase, and the semiconductor industry is entering a new structural demand cycle.

The bullish outlook on storage has become more aggressive. Goldman Sachs recently revised its forecasts for the DRAM supply-demand gap from 2026 to 2028 deeper into the shortage range, with 2027’s estimate adjusted from -2.5% to -5.9%, nearly doubling. Their assessment is that this storage cycle differs from previous ones: demand from AI servers offers greater visibility, while supply growth is constrained by long-term contract commitments, leading to a longer duration of price increases than the market expects.

Goldman Sachs even raised Kioxia’s operating profit forecasts for 2027 through 2029 by 16% to 48% across the three years, citing the expectation that this wave of high profitability could last two to three years. For a company in the highly cyclical storage business, such a forecast suggesting sustained high profits for three years is extremely rare on Wall Street.

JPMorgan’s shift in stance is more interesting. In 2024, they were still proclaiming a “DRAM winter,” forecasting prices to decline for years starting in Q4 2024. Yet by 2025, they completely reversed course, embracing the supercycle theory and predicting a 62% increase in DRAM prices by 2026, with SK Hynix and Samsung’s profits exceeding consensus estimates by 30% to 50%.

But the bearish voices are also strong—and influential.

Michael Burry publicly warned in May that the current semiconductor rally bears strong similarities to the final months of the internet bubble from 1999 to 2000. The SOX has risen 65% this year and 10% in a single week, while the SOXX ETF is trading 60% above its 200-day moving average—a level of technical stretch rarely sustained in history. SEC filings reveal that he has purchased large amounts of put options on SOXX, QQQ, NVIDIA, Palantir, and Oracle, with expiration dates set for January 2027 and strike prices significantly below current share prices.

Man Group (one of the world’s largest publicly traded hedge funds) published a lengthy article in June analyzing the risks of the AI bubble. Their core argument is that the financial architecture surrounding AI has become too large, overly leveraged, and excessively reliant on a small number of interconnected participants.

They specifically noted that the massive construction of AI data centers has been financed through private credit, with collateral consisting of "hardware that depreciates as quickly as smartphones, not 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 becomes unavoidable.

Looking ahead, several key milestones are worth noting.

Micron will release its earnings on June 24; forward guidance on HBM demand and capacity allocation will determine the direction of the entire storage sector this summer. NVIDIA’s next earnings report is equally critical—any sign of even a slight slowdown in AI chip demand could lead to a renewed repricing of sentiment across the sector.

Looking further ahead, the timeline for capacity ramp-up is the true turning point. SK Hynix’s M15X facility is expected to reach full output by mid-2027, with its Yongin new factory accelerated to February 2027. Samsung’s P5 facility is scheduled to begin production in 2028. Micron’s Idaho Fab 1 is expected to contribute output by mid-2027.

Combined, these factors will increase industry capacity by 20% to 30% between the second half of 2027 and the first half of 2028. The issue is that HBM demand is also growing at a compound annual rate of over 40%. Whether supply can keep up with demand depends on whether AI capital expenditures slow 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-node foundry capacity—a figure that represents efficiency in bull markets but systemic risk in conflict scenarios. Factors such as the Taiwan Strait situation, the escalation path of U.S. export controls on China, and the degree of coordination between Japan and the Netherlands on equipment restrictions are rarely discussed during favorable market conditions, but once a crisis emerges, they will influence pricing faster than any fundamental change.