4 months up 230%, how much longer can the semiconductor "bottleneck" rally run?

U.S. stocks closed in the early hours, with the Philadelphia Semiconductor Index (SOX) breaking through the 14,000-point mark for the first time in history, reaching a new all-time high.

Historically, there have only been two periods when the SOX surged over 230% within 14 months: from December 1998 to February 2000, and from April 2025 to the present.

The returns in this semiconductor bull market have been highly concentrated and significant. The year-to-date gains for the memory trio—Micron, SK Hynix, and Samsung—have reached approximately 141%, 186%, and 114%, respectively. TSMC's US-listed ADR has risen over 50% year-to-date.

Nvidia hit a record high of $235.47 on May 14. Broadcom, Marvell, and ASML have all set or approached new records in their respective niches. The 52-week low for the SOXX ETF was $148, with a high near $369, representing an amplitude of close to 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 extreme, with a single HBM3E stack costing around $300, and the upcoming HBM4 estimated at $500 per unit. SK Hynix's 2026 HBM production capacity has already been fully booked by Microsoft, Google, and Nvidia, with some clients even paying the full deposit upfront to secure capacity.

Clearly, the pace of AI data center construction far outstrips the speed of chip capacity expansion.

The "Bottleneck" Bull Market

Scarcity is the most profitable commodity.

Understanding this statement essentially unlocks the core logic of this semiconductor bull market. Whoever controls the "bottleneck" of AI infrastructure holds the strongest pricing power. Conversely, any link in the chain that can be substituted or forced into price cuts, regardless of how strong the demand, will see its stock price stagnate.

Optical modules are a prime example of the latter. A report from Photon Capital in April noted that while Chinese companies occupy seven of the top ten global spots in optical modules, they haven't made much money. The real profits are captured by chip companies. Zhongji Innolight and Eoptolink are world leaders in the shipment volume and cost control of 800G and 1.6T optical modules, directly squeezing the profit margins of U.S. optical module companies like Coherent and Lumentum. Despite demand doubling, profit margins are being compressed. The reason is simple: the assembly phase of optical modules lacks scarcity.

Memory, however, has become the strongest narrative in this round of U.S. semiconductor stocks. Essentially, this is because it has a stranglehold on the bottleneck, and that grip is tightening.

HBM is not ordinary DRAM. Technologies like 3D stacking, TSV (Through-Silicon Via), and specialized packaging processes represent layers of technical barriers built through over a decade of heavy capital investment. Only three companies globally can mass-produce HBM, and SK Hynix has captured roughly half of the market share.

Interestingly, this logic also holds true when scaled up to the macro, national level.

The true winners of 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 happened to build scarce industrial clusters in some irreplaceable link of the chain. Scarcity is the key.

Every Region Has Its Own Main Track

This viewpoint, seen proposed in the U.S. stock community, is very interesting.

The United States remains at the very top of the value chain.

Nvidia, AMD, and Broadcom dominate ASIC design; Synopsys and Cadence lead in EDA tools; Arista powers AI networking; and the three major cloud providers package computing power as a service to sell worldwide. Google, Amazon, and Microsoft are all accelerating their in-house ASIC development. Broadcom and Marvell together command approximately 95% of the custom ASIC design market, with Google alone spending around $8 billion annually on TPU development, largely for Broadcom.

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

Taiwan's story revolves around TSMC and advanced packaging. TSMC is the only company in the world capable of mass-producing 3nm and 2nm processes. All three of TSMC's CoWoS backend fabs are running at full capacity, with lead times of 52 to 78 weeks. Nvidia alone locks down 60% to 70% of CoWoS capacity. TSMC is expanding its monthly capacity from 35,000 wafers at the end of 2024 to 130,000 wafers by the end of 2026, nearly a fourfold increase. However, even with this massive expansion, capacity remains tight. Taiwan's server ODM ecosystem, including Hon Hai (Foxconn), Quanta, and Wistron, is also shipping in volume alongside AI server demand.

South Korea's story, however, is entirely focused on memory. SK Hynix holds roughly 50% to 55% of the global HBM market share, Samsung accounts for 19% to 35%, and Micron around 5% to 20%. HBM is not the same as regular memory. Technologies like 3D stacking, TSV, and specialized packaging represent layers of technical barriers that South Korean companies have spent over a decade building through sustained investment.

The roles of Japan and the Netherlands are also significant. Tokyo Electron makes semiconductor equipment; Shin-Etsu Chemical and SUMCO produce silicon wafers; and Ajinomoto provides ABF substrate materials. While Japan might be out of the race for end-chip products long ago, its position in materials and precision processing remains irreplaceable to this day.

The Netherlands' role is even more direct: ASML has a monopoly on EUV lithography machines. In January, Morgan Stanley significantly raised its price target for ASML to €1,400, predicting that 2027 will be the year of the highest profit growth for ASML, with EPS up 57% year-over-year. They base this on three drivers: advanced logic foundry capacity expansion exceeding expectations, large-scale capacity expansion in the DRAM memory sector, and overall demand performing better than anticipated. Dutch packaging equipment companies like BESI have also secured massive orders from the surge in AI chip packaging demand.

The entry points for China and Europe differ, but the logic is similar: both have built cost advantages or delivery capabilities in specific segments of AI infrastructure.

Zhongji Innolight and Eoptolink are world leaders in the shipment volume and pricing power of 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 come from temporary pricing power due to a phase-specific shortage of 800G capacity. Once 1.6T mass production ramps up in the second half of 2026 or 2027, and second and third-tier manufacturers increase their capacity, price pressure on the module side will arrive quickly.

In Europe, companies like Schneider Electric, ABB, and Vertiv, which specialize in power distribution and cooling, have received orders far exceeding expectations amid the surge in data center electricity consumption. Wedbush estimates that hyperscaler AI infrastructure spending 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 Reshapes the Semiconductor "Smile Curve"

If we use the smile curve to summarize this landscape: the left end (USA) handles "definition and design"; the upper-middle part (Taiwan, South Korea, Netherlands, Japan) is responsible for "advanced chip manufacturing"; the lower-middle part (Taiwan, China, Southeast Asia) handles "large-scale assembly"; and the right end (USA and China) owns the "cloud platforms, models, and customer access points."

The originator of this curve was Acer founder Stan Shih, who used this model in 1992 to explain why PC assembly had the thinnest profit margins.

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

Both the value chain analysis from FourWeekMBA and a paper published by Atlantis Press this year point to the same conclusion: AI is re-elevating the middle section of the traditional smile curve. TSMC's advanced CoWoS packaging, SK Hynix's HBM stacking, and ASML's EUV lithography machines—these segments were considered the "thin-margin middle manufacturing" in the traditional manufacturing smile curve. But in the AI era, they have become the scarcest resources, commanding profit margins and pricing power no less than the design and application ends.

Data from the paper shows Nvidia's gross margin for 2023-2024 was 72.72%, with a net margin of 48.85%. However, TSMC's gross margin for Q1 2026 reached 66.2%, with a net margin of 50.5%. The profit margin gap between the design and manufacturing ends is narrowing, something unprecedented in the history of the semiconductor industry.

The traditional smile curve posits that manufacturing has the thinnest margins. AI is turning the most difficult parts of manufacturing into the most scarce resources.

A March report from Morgan Stanley on Asian semiconductors reaches a similar conclusion: the AI cycle from 2023 to 2024 was primarily focused on GPUs, but from 2025 to 2026, demand is beginning to spread across a broader industrial chain—memory, advanced packaging, custom ASICs, and data center networking are all taking the baton.

Each rotation of bottlenecks pushes a new set of previously overlooked companies into the spotlight, while the biggest gainers from the previous cycle enter a digestion phase.

How Much Further Can the Bull Run? The Bull vs. Bear Debate

Let's start with the bulls. Dan Ives of Wedbush directly called for the Nasdaq to reach 30,000 points over the next year on CNBC in May, arguing that demand for AI chips still far outstrips supply. Goldman Sachs provides more specific numbers, projecting global AI capital expenditure at approximately $765 billion in 2026, rising to $1.6 trillion by 2031.

In its March report on Asian semiconductors, Morgan Stanley explicitly stated that AI computing investment is still in an expansion phase, and the semiconductor industry is entering a new cycle of structural demand.

The bullish case for memory is even more aggressive. Goldman Sachs recently revised its DRAM supply-demand gap forecasts for 2026-2028, deepening the projected shortfall. The forecast for 2027 was almost doubled, moving from a previous -2.5% to -5.9%. Their assessment is that this memory cycle is different from the past: AI server demand offers higher visibility, supply growth is constrained by long-term lock-in agreements, and price increases will last longer than the market expects.

Goldman Sachs even significantly raised its operating profit forecasts for Kioxia for the three years from 2027 to 2029, by 16% to 48%, based on the belief that this period of high profitability can last for two to three years. Issuing a "high profits for three years" judgment for a company in the strongly cyclical memory business is extremely rare on Wall Street.

Morgan Stanley's change of heart is even more interesting. In 2024, they were still predicting a "DRAM winter," forecasting multi-year price declines starting in Q4 2024. By 2025, they had completely flipped to a super-cycle thesis, predicting DRAM prices would rise 62% in 2026 and that SK Hynix and Samsung's earnings would beat consensus expectations by 30% to 50%.

However, the bearish voices are also significant and come from notable figures.

Michael Burry publicly warned in May that this semiconductor rally bears a strong resemblance to the final months of the 1999-2000 dot-com bubble. The SOX is up 65% year-to-date, gaining 10% in a single week, and the SOXX ETF is trading 60% above its 200-day moving average. Such technical extensions are rarely sustainable historically. SEC filings show he purchased a large number of put options on SOXX, QQQ, Nvidia, Palantir, and Oracle, with expiration dates set for January 2027 and strike prices well below current share prices.

Man Group (one of the world's largest publicly listed hedge funds) published a lengthy piece in June specifically deconstructing the AI bubble risk. Their core argument is that the financial architecture surrounding AI has become too large, overly leveraged, and excessively dependent on a small number of interconnected participants.

They specifically noted that a large portion of AI data center construction is financed through private credit, with loans collateralized by "hardware that depreciates as quickly as a mobile phone, rather than long-lived assets like buildings." The first wave of defaults could occur between 2027 and 2028, when initial leases expire and the gap between financing assumptions and reality becomes unavoidable.

Looking ahead, several key dates warrant our attention.

Micron's earnings report on June 24 will provide forward guidance on HBM demand and capacity allocation, likely setting the tone for the memory sector for the entire summer. Nvidia's next earnings report is also critical; even a slight signal of deceleration in AI chip demand could lead to a re-pricing of sentiment across the entire sector.

Looking further out, the timeline for capacity release is the true watershed moment. SK Hynix's M15X factory is expected to ramp up production by mid-2027, with the Yongin new fab brought forward to February 2027. Samsung's P5 plant will start production in 2028. Micron's Idaho Fab 1 is expected to contribute output by mid-2027.

Combined, this means industry capacity will increase by 20% to 30% between the second half of 2027 and the first half of 2028. The problem is that the compound annual growth rate (CAGR) for HBM demand is also above 40%. Whether supply can catch up to demand depends on whether AI capital expenditure slows down before then.

The final variable is geopolitics. The more concentrated the semiconductor supply chain, the greater the impact of a black swan event. TSMC alone accounts for over 90% of global advanced process foundry capacity—this is efficiency in a bull market but a systemic risk in a conflict scenario. Factors like the situation in the Taiwan Strait, the trajectory of U.S. export controls on China, and the level of cooperation from Japan and the Netherlands on equipment restrictions are topics no one wants to discuss when markets are good. But if something happens, the speed of price adjustment will outpace any fundamental change.