The Philadelphia Semiconductor Index SOX has historically surpassed 14000 points, setting a new historical high; This round of market has increased by 230% accumulatively in four months, only benchmarking against the Internet foam super market in 2000. Different from previous cycles, this AI semiconductor bull market is no longer a single carnival, but a distributed bull market driven by scarce capacity pricing, global division of labor restructuring, and smile curve reversal: the traditional low profit midstream manufacturing sector has become a high profit highland in the industry chain, and the regional tracks of various countries around the world have been completely solidified. At the same time, the extreme game of long and short on Wall Street, the three variables of cycle turning point, capacity release, and geopolitical risk determine the final outcome of the market.

1、 Market review: replicate the super market of 2000 level, and the valuation of the entire industry chain skyrockets

The US stock market closed overnight, with the Philadelphia Semiconductor Index SOX stabilizing at the 14000 point mark, setting a new historical high. Back in history, the SOX index rose by more than 230% in 14 months, only twice in the world: the first was the bull market of the Internet foam from December 1998 to February 2000, and the second was the AI computing super cycle since April 2025.

In this round of market trends, sector returns are highly concentrated, and the gains of individual stocks in the core computing power sector are crushing the market

1. Storage sector: With a 141% increase within the US light year, SK Hynix 186%, Samsung 114%, storage has become the strongest mainline in semiconductors;

2. Manufacturing leader: TSMC's US stock ADR has risen by over 50% this year, with advanced packaging capacity continuing to command a premium;

3. On the computing power design side, Nvidia reached a historic high of $235.47 in May, while Broadcom, Maiwell, and ASML simultaneously refreshed their stock price highs;

4. Sector ETF: The semiconductor broad-based SOXX ETF has a volatility of up to 150% within the year, with a 52 week price range of $148-369.

Fundamental supply-demand gap continues to widen, computing power premium: Goldman Sachs raises the DRAM supply-demand gap for 2026 to 4.9%, the most severe supply shortage in 15 years; The price of high-end HBM storage continues to soar, with the single HBM3E unit price exceeding $300 and the next generation HBM4 production estimated unit price reaching as high as $500. At present, all HBM production capacity of SK Hynix in 2026 has been fully locked in by Microsoft, Google, and Nvidia, and top cloud manufacturers have pre paid in full to seize production capacity. The current industry situation directly confirms that the growth rate of AI data center computing power construction far exceeds the global chip production capacity expansion speed.

2、 The core of this bull market: scarcity equals pricing power, and extreme bottlenecks give rise to structural market trends

This round of semiconductor bull market is not a general rise in the entire industry chain, but a unilateral bull run in irreplaceable scarce links and continuous pressure on profits in replaceable assembly links, with differentiation logic running through the entire industry chain.

1. Negative case: The demand for optical module industry is increasing and profits are decreasing

According to industry research data from Photon Capital, seven of the top ten optical module companies in the world are Chinese manufacturers, with Zhongji Xuchuang and Xinyisheng firmly ranking in the first tier of 800G/1.6T optical modules worldwide, surpassing overseas Coherent and Lumentum in terms of shipment volume and cost control capabilities. But there is an abnormal phenomenon in the industry: industry demand has doubled, and module manufacturers' profit margins continue to be compressed.

Core essence: Optical module terminal assembly technology has a low threshold, can be quickly replicated in production capacity, has no long-term scarcity, belongs to a fully competitive link, and naturally loses pricing power. Institutions predict that after the concentrated landing of 1.6T optical module production capacity in the second half of 2026-2027, the industry price war will further compress the profits of the middle and lower reaches.

2. Core theme: HBM storage, the strongest barrier link in this round of computing power cycle

Different from ordinary DRAM memory, HBM relies on 3D stacking, TSV through silicon vias, and proprietary advanced packaging processes to form, requiring more than ten years of heavy asset technology accumulation from enterprises and extremely high industry entry barriers. Only three companies in the world have mass production capabilities, and the industry landscape is highly solidified: SK Hynix holds a 50% -55% market share, Samsung holds a 19% -35% market share, and Micron only has a 5% -20% market share.

From the industry to the national dimension, the underlying logic of this AI bull market is unified: only by deeply cultivating for decades, building technological moats, and forming irreplaceable regional industrial clusters can we share the dividends of AI computing power.

3、 Global division of computing power: Six major economies, six exclusive tracks

AI is reconstructing the global semiconductor supply chain, and major economies rely on their own industrial accumulation to lock in exclusive high profit tracks, forming a stable distributed computing power division of labor pattern, with each performing its own duties and clear barriers:

United States: Top level design+cloud ecosystem, controlling the top value of the industrial chain

Firmly control the top-level benefits of the computing power industry chain: Nvidia AMD、 Broadcom controls GPU/ASIC chip design; Xinsi Technology and Kaideng Electronics monopolize the global EDA industry software; Arista leads AI high-speed network equipment; The three major cloud vendors in North America, Amazon, Microsoft, and Google, independently develop computing power chips and pay Broadcom $8 billion annually for TPU custom development. Broadcom and Maiwell together monopolize 95% of the custom ASIC OEM design market and control the global computing power definition rights.

Taiwan, China, China: advanced manufacturing process+CoWoS packaging, absolute core of AI OEM

TSMC monopolizes over 90% of the world's 3nm/2nm advanced process production capacity, with three CoWoS advanced packaging factories operating at full capacity and order lead times of 52-78 weeks. NVIDIA exclusively locks in 60% -70% of packaging capacity. TSMC is accelerating its production expansion and plans to increase the monthly production capacity of AI chips from 35000 to 130000 by the end of 2026, nearly four times the capacity expansion, but still unable to meet the demand for computing power. Meanwhile, Foxconn, Guangda, and Wistron are leveraging their local OEM systems to fully enjoy the benefits of AI server shipments.

South Korea: Specializes in HBM storage and monopolizes computing power storage bases

The national industrial resources are focused on the high barrier storage track, relying on more than ten years of research and development investment in stacking technology, firmly establishing the global HBM supply chain leader, binding global AI giant orders, and enjoying the benefits of computing power storage shortage, becoming the most profitable regional sector in this cycle.

Netherlands+Japan: Material and equipment bottleneck, invisible oligopoly of essential needs

ASML in the Netherlands exclusively monopolizes EUV lithography machines, and Morgan has raised its target price to 1400 euros, predicting a year-on-year EPS growth rate of 57% in 2027. Advanced process and storage expansion will boost equipment orders in both directions; Japan controls the entire semiconductor chain with essential materials such as Tokyo Electronic Process Equipment, Shin Etsu/SUMCO silicon wafers, and Ajinomoto ABF packaging substrates. Despite losing the competition for terminal chips, the basic materials are irreplaceable.

Chinese Mainland: midstream and downstream modules+power infrastructure, seize the cost advantage track

Relying on the advantages of large-scale manufacturing, leading the global high-speed optical module track; Europe is focusing on AI data center supporting infrastructure, with orders from Schneider, ABB, and Vertiv power distribution and cooling companies exploding. According to Wedbush data, in 2026, global cloud vendors will spend $725 billion on AI infrastructure, a year-on-year increase of 77%, with power temperature control being one of the fastest-growing segments.

4、 Paradigm change: AI completely rewrites the classic smile curve of semiconductors

In 1992, Acer founder Shi Zhenrong proposed the semiconductor smile curve: the profits of R&D, design and terminal application at both ends of the industrial chain are extremely high, and the profits of midstream manufacturing, assembly and processing are low. The PC era and mobile Internet era all follow this set of value laws.

In the era of AI computing power, the smile curve has completely reversed and reshaped: the valuation and profit margins of high barrier manufacturing in the midstream have risen sharply, smoothing out the profit gap with the design end.

Industry profit data confirms the transformation: Nvidia's gross profit margin for 2023-2024 is 72.72%, and net profit margin is 48.85%; TSMC achieved a gross profit margin of 66.2% and a net profit margin of 50.5% in the first quarter of 2026, with its advanced manufacturing profitability equaling that of a leading chip design company, marking a historic change in the industry.

Synchronized switching of track rotation rules: AI market single speculation GPU design end in 2023-2024; From 2025 to 2026, demand will overflow and HBM storage, advanced packaging, customized ASIC, and computing power network will continue to rise. The bottleneck of the industrial chain is turning upwards, becoming the core feature of this distributed bull market.

5、 Extreme game between the sky and the sky: a trillion dollar bull market, a foam or a bull market?

Multi head logic: Long term upward trend in computing power capital expenditure, establishment of super storage cycle

1. Macro prediction of institutions: Wedbush analyst Dan Ives sees a yearly increase of 30000 points for Donaldrak; Goldman Sachs estimates that global AI capital expenditures will reach $765 billion in 2026 and climb to $1.6 trillion in 2031, demonstrating long-term growth potential in computing power investment.

2. Storage super cycle tuning: Goldman Sachs continues to lower DRAM supply and demand data, with the supply and demand gap reduced to -5.9% by 2027, almost doubling the shortage; Contrary to the pessimistic prediction of a DRAM winter in 2024, Morgan Stanley has changed its bullish stance and predicts a 62% increase in DRAM prices by 2026, with Korean storage companies exceeding market expectations by 30% -50% in profits. At the same time, Wall Street rarely gives a three-year high profit forecast for storage companies, locking in the industry's high prosperity from 2027 to 2029.

3. Industrial chain transmission: AI computing power investment is in the early stage of expansion, and the industry has entered a structural incremental cycle. The order locking mode locks in the medium and long-term production capacity gap, and the supply-demand imbalance is difficult to reverse in the short term.

Short position logic: benchmarking the Internet foam, credit leverage hidden collapse risk

1. Legendary short position warning: Michael Burry made a big move in putting options on SOXX, NASDAQ and Nvidia in January 2027, and spoke frankly about this round of semiconductor market, with the trend highly reminiscent of the end of the Internet foam in 1999-2000; The current SOXX has a premium of up to 60% relative to the 200 day moving average, with a severe technical oversold and extremely poor historical sustainability.

2. Top hedge fund risk warning: Global listed hedge fund leader Man Group warns that the financial leverage in the AI track is too high and participants are highly homogeneous; A large number of data centers rely on private credit financing, with the collateral being computing hardware that depreciates rapidly, rather than fixed assets. After the concentrated expiration of leases in 2027-2028, the first wave of industry defaults may occur.

6、 Three turning point variables: determining the duration of the semiconductor bull market

1. Short term performance milestone: sentiment of the designated sector in the second half of June financial report

On June 24th, Micron released its financial report, and HBM orders and production capacity guidance directly determine the summer market situation of the storage sector; Nvidia's subsequent financial report on computing power shipments will directly affect the valuation and pricing of the entire sector, and a slowdown in demand margins can trigger a sector correction.

2. Mid term production capacity milestone: Concentrated release of production capacity from the second half of 2027 to 2028

SK Hynix's new factory, Samsung P5 factory, and Micron Idaho factory will be put into operation from mid-2027 to 2028, with an overall expansion of 20% -30% in the industry's computing chip production capacity. But the annualized demand growth rate of HBM in the industry remains above 40%, and whether supply can catch up with demand depends on whether global AI capital expenditures cool down in advance.

3. Long term geopolitical node: high concentration of supply chain breeds black swan

TSMC monopolizes 90% of the world's advanced processes, and the concentration of its supply chain has reached a historical peak. During the upward phase of the market, concentrated production capacity represents efficient division of labor, while during the downward phase of the market and geopolitical fluctuations, it will directly trigger systemic risks. The situation in the Taiwan Strait, chip regulations in Europe and America, and restrictions on equipment coordination between Japan and the Netherlands may break the existing supply and demand pricing system at any time.

Conclusion

The AI era has completely rewritten the century old value distribution logic of semiconductors: the traditional smile curve has failed, high barrier midstream manufacturing has crowned profit high ground, and the world's six major economies have formed a distributed computing bull market with each performing its own duties.

In the short term, the gap between computing power supply and demand still supports high-level fluctuations in the sector, and the logic of bullish fundamentals remains unbroken; In the medium to long term, the three major risks of capacity release, credit risk, and geopolitical games continue to accumulate, and the market is about to enter the stage of performance verification and valuation digestion. This round is not a one-sided sustainable bull market, but a structurally differentiated market where high barriers and scarce links continue to rise while fully competitive links continue to weaken. Choosing barriers and abandoning internal competition are the only core ideas for subsequent semiconductor trading.