Domestically developed AI large models are gaining pricing power—will the industry chain undergo a r

尺度商业

joined discussion ·

One-Day Rebound or the Start of a New Rotation? Has the Tech Bull Narrative Changed? [Proprietor’s View]

By Huang Liming, Chief Analyst

In July, tech stocks suffered a violent sell-off. Even with the sharp rebound on the last day of the month, 54 individual stocks posted declines exceeding 50% over the past month. With August about to open, everyone is asking the same question: Is this the beginning of a major market bottom, or the start of another leg down?

Introduction: A Journey from Heaven to Hell—and Now Confusion

At 3 p.m. on July 31, the A-share market closed.

The STAR 50 Index rose 2.99%, and the ChiNext Index gained 3.06%—seemingly a solid rebound day. But if you’ve seen the intraday chart, you’d know just how suffocating that day really was.

That day, the STAR 50 Index peaked at 1,731 points (up 8.99%) before closing at 1,636 points—a nearly 6-percentage-point intraday pullback. The ChiNext Index retreated from a high of +6.79% to close at +3.06%.

AI application-related stocks surged en masse—Kunlun Tech, Chinese Online, Yidiantianxia, BlueFocus, Foxit Software, and Hongjing Technology all locked in 20% daily trading limits. Meanwhile, memory hardware stocks told a different story: Puyaa Semiconductor slid from its 20% limit-up price to finish barely up 0.5%. Many tech names saw intraday gains reverse into losses of over 10% by the close; Demingli was even more dramatic—it went from hitting the upside limit straight to closing down 1.06%, making it one of the very few stocks in the entire market to finish lower amid a broad-based rebound.

On the same day, within the same tech sector, upstream and downstream segments moved in completely opposite directions.

This wasn’t a rebound. Nor was it a crash. This was—money moving elsewhere.

In the early hours of the same day, Micron Technology’s stock jumped 18.36%. On the same day, SK Hynix soared nearly 30% in Seoul, driving South Korea’s KOSPI index up 17.91%. A-shares opened sharply higher in the morning, only to begin giving back all gains within half an hour. Many tech stock holders made the same choice that day: sell into strength. They weren’t trading fundamentals—they were trading fear.

Thus, three fundamental questions emerge:

1. Is the tech sector’s rebound just a one-day rally?

2. Is the market shifting toward a new rotation theme?

3. Has the underlying logic of the tech bull market truly changed?

Let’s peel back the layers one by one.

Chapter One: What scale are global players betting with? The data doesn’t lie.

Before discussing China’s A-share market, we must first understand what global capital is doing. While sentiment in the A-share market can be volatile, real money from global industrial capital never lies.

Capital expenditures by memory giants: the largest in human history.

Samsung Electronics: total capital expenditures plus R&D spending for the full year 2026 will exceed KRW 110 trillion (approximately USD 73.3 billion). Note this isn’t revenue or profit—it’s pure investment. Kim Yong-hwan, head of Samsung Semiconductor, stated internally: 'The profit generated in 2026 alone will surpass the cumulative profits from our semiconductor business over the previous forty years.'

SK Hynix: capital expenditures will jump from KRW 30 trillion in 2025 to over KRW 40 trillion. Chairman Chey Tae-won has directly ordered the completion date for the fourth fab at the Yongin semiconductor cluster to be moved forward by 12 years—from 2045 to 2033. Twelve years—just like that, it’s been accelerated.

Micron Technology: Signed 16 strategic customer agreements, locking in approximately $100 billion in contractual performance obligations—note, these are not letters of intent, but RPOs (

Remaining Performance Obligations), which represent future revenue firmly recognized under accounting standards.

Samsung and SK Hynix jointly announced a combined investment of 800 trillion Korean won (approximately $518 billion) in the Honam region of South Korea to build four new wafer fabs.

To put this in perspective: $518 billion is equivalent to the annual GDP of a mid-sized developed country. This isn’t about 'hype'—it’s the largest single-industry investment in human history, placing a massive bet on one conviction: the AI era has only just begun.

What are tech giants racing for? Pricing data provides the answer.

Why are these giants willing to spend so much? Because they’re racing to secure one thing—capacity.

Just look at the price surges:

Samsung’s average selling prices for DRAM and NAND have surged 146% above their 2025 averages; SK Hynix reported quarterly DRAM price increases exceeding 60% and NAND increases over 70%; retail SSD prices from Samsung in Japan have skyrocketed by 300%; TrendForce forecasts DRAM contract prices will rise another 13%–18% in Q3 2026.

Even more critically, SK Hynix executives revealed firsthand: 'Customers are demanding a 5x to 6x increase in supply.' NVIDIA and SK Hynix have signed a multi-year agreement with a potential value of up to $500 billion. Microsoft and Google have prepaid 10%–30% of their contract amounts to SK Hynix.

The global memory chip market is experiencing the most severe supply-demand imbalance in history. This isn’t the peak of a cycle—it’s the acceleration phase of a supercycle.

How long can the price rally last? Let’s break it down layer by layer.

But we must stay calm: not all memory products can keep rising indefinitely.

This table reveals a core insight: the upcycle for memory chips has been significantly extended, but it is not perpetual. The HBM story will last the longest, while NAND faces risks earliest.

Chapter Two: Has the fundamental logic of the tech bull market changed? — Five structural shifts

If you bought memory chips before 2023, you’ve surely lived through the classic playbook: upcycle begins → manufacturers ramp up capacity aggressively → oversupply hits in 2–3 years → prices crash → losses trigger production cuts → next cycle begins. This 3–4 year boom-bust cycle is why memory chips have traditionally been labeled as 'highly cyclical stocks.'

However, this cycle has brought five unprecedented structural changes.

Change One: Long-Term Agreements (LTAs) have reshaped the pricing mechanism

This is the most critical shift. In the past, memory chips were standardized commodities, with buyers and sellers transacting at prevailing market prices—price movements driven entirely by supply and demand.

Now? SK Hynix’s LTAs cover roughly 50% of its revenue and have eliminated price caps—allowing LTA prices to rise in tandem with spot market increases. Micron Technology’s Supply Chain Agreements (SCAs) not only lock in volume and pricing for five years but also include a price floor mechanism, with approximately $22 billion already secured under these terms.

What does this mean? Even if supply-demand dynamics ease in the future and spot prices soften, the price floor embedded in LTAs will significantly slow both the depth and speed of any price decline. Whereas memory cycles used to follow a pattern of 'sharp rally → steep crash → sharp rally,' they may now evolve into 'sharp rally → gradual decline → sustained high-level volatility.'

Change #2: HBM is not a standardized commodity

HBM requires joint validation and customized design with customers, and logic dies must be tailored to customer specifications. This resembles a custom ASIC more than a standardized DRAM chip.

HBM3E → HBM4 → HBM4E → HBM5—the pace of technological iteration is rapid, with each generation delivering a significant increase in average selling price (ASP). As HBM’s share of SK Hynix’s revenue continues to rise, the overall value-added profile of the company’s product portfolio is being structurally elevated.

Change #3: AI demand represents sustained growth, not a one-time pulse

Past memory chip demand surges came from 'upgrade cycles'—smartphones transitioning from 4G to 5G, or PCs shifting from HDDs to SSDs. Such pulse-like demand arrives quickly and fades just as fast.

AI is different. From training to inference, the commercialization of Agentic AI is driving continuous compute demand. Global server shipments are projected to grow 17% year-over-year in 2026 and continue rising in 2027. This isn’t a 'device replacement cycle'—it’s about 'building data centers from scratch.'

Change #4: Capacity expansion cycles have been significantly extended

Traditional memory capacity expansions typically take about 2–3 years. But this cycle? The Yongin project was announced 12 years early, yet its first fab won’t come online until 2027, with full capacity not expected before 2030+. Critical equipment—including EUV tools, etchers, and photomasks—is in extremely tight supply, with Samsung and SK Hynix competing directly against Taiwan Semiconductor and Intel for these resources.

Supply-side response speed is far slower than in any previous cycle.

Change #5: Geopolitics is reshaping the supply landscape

The South Korean government plans to double domestic DRAM production capacity within five years. The United States is pressuring Samsung and SK Hynix to build factories in the U.S. Global memory supply remains highly concentrated in South Korea, the U.S., and Japan. Although Chinese memory chipmakers ChangXin Memory Technologies and Yangtze Memory Technologies still lag behind by about two to three years in mainstream technology—and roughly five years in HBM—their pace of catching up is rapid, and the memory industry will ultimately face a reshuffling driven by Chinese manufacturing.

What does geopolitics imply? Capacity expansion is no longer purely a commercial decision but a strategic game involving national security. This further delays the timeline for new supply coming online.

Notably, research by Liu Chenming’s strategy team at GF Securities has uncovered a key pattern: 'single peaks' are market exceptions; repeated, multiple peaks are the norm. China's A-share broad-based indices saw only one single peak—in 2015 (due to regulatory deleveraging)—and the Nasdaq experienced just one single peak—in 2000 (as Y2K-driven hardware upgrade demand was disproven). Both instances shared a common trait: a one-off disruptive factor outside underlying industry trends.

Returning to the present: does this cycle feature such a one-off disruptive factor? Microsoft’s $15 billion capex cut reflects 'slowing growth' rather than 'demand collapse.' AI applications are accelerating their rollout—OpenAI has significantly slashed prices to boost volume, DeepSeek-V4 has entered public beta, and ByteDance’s Seedance video-generation platform has launched. The fundamental driver of memory demand—AI data center expansion—has not been invalidated.

Therefore, this cycle is far more likely to follow a pattern of repeated, multiple peaks rather than a one-time crash.

Chapter Three: Is the market entering a new rotation? — Money is moving elsewhere

The market action on July 31 sent an extremely clear signal: AI applications surged with multiple stocks hitting their daily trading limits, while memory hardware stocks stalled or even closed lower. This isn’t the 'end of the tech rally'—it’s capital migration within the tech sector itself.

The AI supply chain exhibits a three-tier rotation.

To understand this migration, one must grasp the three-layer structure of the AI supply chain:

Tier One [Infrastructure Layer]: GPUs, HBM, DRAM, NAND, optical modules, and servers—the point from which capital begins taking profits.

Layer 2 [Platform Layer]: Cloud computing, AI platforms, databases, and enterprise software—this is where the money flows.

Layer 3 [Application Layer]: AI agents, SaaS, AI-powered marketing, healthcare AI, financial AI, and AI robotics—this is the endpoint, and also a new starting point.

What happened in China’s A-share market on July 31? AI application stocks in Layer 3 surged en masse. Capital hasn’t left the tech sector—it has simply rotated within it, shifting from 'shovel sellers' to 'shovel users.'

This view is globally validated: U.S. equities have already led the rotation from hardware to applications. AI software and cloud service providers like Microsoft and Salesforce have demonstrated far greater price resilience than pure-play hardware companies.

So where does this rotation ultimately end?

The real logic chain goes like this: AI applications explode → demand for inference compute surges → demand for storage spikes again → capital eventually cycles back to the infrastructure layer.

This creates a complete capital loop. OpenAI’s 80% price cut doesn’t mean 'AI has become worthless'—it means 'many more people can now afford AI.' For underlying compute and storage, this is a usage explosion, not a price collapse.

What does this mean?

For investors, this implies two key judgments:

First, if you hold pure-play hardware stocks, understand that the market is currently transitioning from 'hardware euphoria' to 'application validation.' During this transition, hardware stocks will underperform application-focused names. This doesn’t mean hardware fundamentals have deteriorated—it simply reflects that the market is waiting for the next catalyst (AI application explosion → surge in inference demand).

Second, if you’re focused on AI applications, you must distinguish between those with genuine paying customers and those that are just traditional software wrapped in an AI label. August is peak earnings season—concept stocks without real revenue support will quickly lose momentum.

Chapter Four: The Federal Reserve — The Sword Hanging Overhead

This is the largest variable among all current scenarios.

First, the logic for rate hikes is strengthening.

On July 29, the FOMC kept rates unchanged at 3.50%–3.75%, but the vote outcome was unsettling—9 to 3, with three members advocating a 25-basis-point hike. This marks the largest internal divergence since 2016.

More critically, the yield on the 30-year U.S. Treasury has surged to 5.27%, a 19-year high. The 10-year yield stands at 4.74%, and the 2-year at 4.29%. Mortgage rates are approaching 6.66%, and analysts expect they could surpass 7.5% by year-end.

Core PCE remains elevated at 3.4% (well above the 2% target), and inflation expectations are rising rather than falling (1-year expectation at 3.7%, 3-year at 3.3%). Moreover, the intermittent flare-ups in U.S.-Iran tensions could at any moment drive oil prices higher and trigger a new round of imported inflation.

Second, the base case is to hold steady. The most likely path over the next 3–6 months is continued inaction. Waller (the new Fed Chair) has a very clear strategy: extreme data dependence, rejection of forward guidance, and emphasis on assessing 'trends' rather than isolated data points. He will not act prematurely until clear trend signals emerge in both inflation and employment.

Therefore, we must also clearly understand the three transmission channels through which Fed policy affects China’s A-share market.

The good news is this: regardless of what the U.S. does, China’s own policy direction is clear. The Political Bureau meeting on July 30 set the tone for 'a more proactive fiscal policy and a moderately accommodative monetary policy.' The probability of both a reserve requirement ratio (RRR) cut and an interest rate cut is rising in the second half of the year. The People’s Bank of China’s easing stance will not be fundamentally altered by the Fed’s actions.

Chapter Five: Has the A-Share Market Already Bottomed Out? — A Multidimensional Framework for Confirming the Bottom

The STAR 50 Index fell from 2,100 to 1,588 points, a decline of approximately 24%. At the individual stock level, 54 stocks posted monthly declines exceeding 50%, with an average drop of -54.8%.

This drawdown ranks second in A-share history, surpassed only by the 2015 deleveraging crash and the quant meltdown in early 2024. However, the key difference lies in the underlying causes: the 2015 crash stemmed from leveraged capital collapse, while the 2024 event was driven by liquidity evaporation—whereas in July 2026, these companies are reporting their strongest quarterly earnings on record.

We assess this from five dimensions used to confirm market bottoms.

Comprehensive assessment: the market has entered the bottoming zone (valuation and sentiment have aligned), but the precise bottom has not yet been confirmed (technical indicators and capital flows have not converged). Current entry constitutes left-side trading; certainty will rise from 50% to over 80% once a volume-drying secondary bottom confirmation occurs.

Meanwhile, we can reference historical rebounds following extreme oversold conditions. In A-share history, after technology sectors dropped more than 30% in a single month, the average rebound over the subsequent three months was approximately +35%. Examples include the 2018 trade war (ChiNext: -37% → +50% rebound), April 2022 (STAR 50: -45% → +55% rebound), and February 2024 (CSI 1000: -35% → +40% rebound).

Moreover, historical data shows that in every crash not driven by fundamental deterioration, the hardest-hit sectors exhibited the strongest rebounds—not the so-called 'resilient' or 'defensive' sectors.

Chapter Six: Competitive Landscape Projections Across the Industrial Chain

It is especially important to note that the following analysis focuses solely on industry positioning, economic moats, and sectoral prospects within each segment, and does not constitute any investment recommendation.

In the memory chip segment, differentiation is key.

The memory chip industry is undergoing unprecedented divergence—

HBM (High Bandwidth Memory) segment: Only SK Hynix, Samsung, and Micron Technology can produce HBM globally, reflecting extremely high technological barriers. There are no A-share companies directly manufacturing HBM, but memory interface chips—the critical components connecting CPUs and DRAM—are among the most technically demanding segments in the HBM supply chain. Globally, only two companies can supply these chips, one of which is listed on the A-share market and holds approximately 45% of the global DDR5 memory interface chip market share.

DRAM segment: Among domestic players, only ChangXin Memory, which recently went public, can independently manufacture DRAM. ChangXin’s LPDDR6 technology is nearing mass production; if successfully launched, it would become the 'benchmark stock' for A-share memory chips. Existing A-share DRAM-related companies are primarily module assemblers and packaging/testing firms—they benefit from rising DRAM prices, but their pricing power and technological barriers are far weaker than those at the manufacturing level.

NAND segment: Yangtze Memory Technologies (YMTC) is China’s main NAND manufacturer, with a technology gap of roughly 2–3 years compared to global leaders. NAND module makers face a supply-demand inflection point risk in the second half of 2027—prior to that, they enjoy profit elasticity from price increases; afterward, inventory impairment risks cannot be ignored.

NOR Flash segment: This is the most domestically self-reliant segment within China’s memory chip industry. The leading company ranks among the top three globally and has simultaneously developed an MCU (microcontroller unit) product line, establishing a platform strategy integrating NOR + MCU. Its Q2 revenue surged 220% quarter-over-quarter, strongly validating sector momentum.

From the perspective of AI applications, the key challenge is bridging the gap from narrative to actual revenue.

Behind the batch of AI application stocks hitting their daily trading limits on July 31 lies an ongoing industry inflection point—

On the cost side: Post-training techniques for open-source large language models have matured, enabling AI application companies to fine-tune proprietary models using their own business data, significantly reducing inference costs. OpenAI’s 80% price cut means more users can afford AI—lowering customer acquisition costs for applications.

On the revenue side: AI-powered digital marketing (cross-border ad placement), AI-driven short videos and content creation tools, and B2B SaaS solutions have already begun generating stable recurring revenue. Market valuation logic is shifting from 'concept-based' to metrics-driven—focusing on ARR (Annual Recurring Revenue), customer retention rates, and gross margins.

On the policy front: The 'Fifteenth Five-Year Plan' allocates RMB 4 trillion in direct investment for national computing infrastructure. Local governments are rolling out subsidies for token-based economies and specialized incentives for AI agents, signaling a clear policy pivot—from 'developing large models' to 'commercializing real-world applications.'

Differentiated directions worth watching:

1. AI-powered digital marketing and cross-border ad placement: The segment with the strongest near-term elasticity, as its performance-based business model has already been validated;

2. Enterprise SaaS + AI Agents: Highest customer stickiness, with concentrated government and enterprise procurement expected in the second half of the year;

3. AIGC-driven entertainment and content: Mature consumer (C-end) monetization models make this segment highly susceptible to market catalysts;

4. AI in finance and government sectors: Ample budgets, high compliance barriers, and low volatility.

From the perspective of semiconductor equipment and materials, domestic substitution remains the overarching long-term theme.

China’s anticipated breakthrough in lithography machines reaching mass production is one of the biggest potential catalysts for the second half of the year. If China truly achieves mass production of ArF immersion lithography tools capable of 28nm–14nm nodes, it would signify that the final bottleneck in China’s chip manufacturing supply chain has been resolved.

Order of sectoral benefits along the supply chain: Lithography machine components (light sources, lenses, wafer stages) → Semiconductor equipment (etching, deposition, cleaning, inspection) → Semiconductor materials (photoresists, sputtering targets, specialty gases) → Wafer fabrication → Chip design.

This is a long-term strategic theme spanning China’s 14th and 15th Five-Year Plans, unaffected by short-term market sentiment.

From the perspective of AI-powered robotics, this represents the AI-driven innovation application with the greatest future growth potential.

Every major technological revolution has aimed to liberate productivity and enhance efficiency, fundamentally rooted in labor substitution. Humanoid robots, in particular, represent the most transformative new product of this era’s technological revolution. They will unlock entirely new use cases capable of directly replacing human labor—performing tasks previously unattainable by conventional machinery—in both households and factories.

Throughout the roughly 5,000–6,000-year history of human civilization on Earth, our planet has been shaped around the human form. The future widespread adoption of humanoid robots will once again elevate humanity’s GDP-generating capacity by an order of magnitude. During the Industrial Revolution, the time required for GDP to double shortened to 30–50 years. Subsequent technological waves—electrification and the information revolution—further accelerated growth, reducing the doubling time to 15–20 years. The current AI wave is projected to shorten this interval further to just 5–10 years.

Unitree Robotics, a domestic humanoid robotics company, is preparing for its IPO. 2026 will be a pivotal 'year of mass production' and the key inflection point marking the industry's transition from lab prototypes to factory output—the realization of 'zero to one.' The true 'iPhone moment'—when humanoid robots become as ubiquitous as smartphones and transform daily life—may still take another 1–3 years or even longer. This hinges on breakthroughs in 'brain-level' intelligence, further cost reductions, and the emergence of killer applications. The future of humanoid robots has already arrived; it simply hasn't gone mainstream yet. This 'iPhone moment' is well worth our collective anticipation.

Conclusion: Answering Three Questions

Based on a comprehensive analysis of global industrial capital flows, AI supply chain rotation patterns, long-term supply-demand dynamics in memory chips, the Federal Reserve’s monetary policy trajectory, and the framework for confirming the bottom of China’s A-share market, we can now address the three questions posed at the outset.

Question 1: Is the tech stock rally just a one-day surge?

No. The high open followed by a decline on July 31 was the result of concentrated selling by trapped positions, not the end of the rebound narrative. When valuations are low and fundamentals continue to deliver strong earnings or even high growth, the room for a rebound is only a matter of time. However, the path of the rebound will be winding—it won’t happen all at once but will require a process of 'panic selling exhaustion → volume contraction and stabilization → second bottoming → confirmation → primary uptrend.' The week of August 3–7 is a critical window.

Question 2: Has the market shifted to a new rotation direction?

Yes, but the rotation is occurring within the tech sector itself. Capital is migrating from the AI infrastructure layer (hardware) to the application layer (software/AI agents/SaaS). This reflects a global industry trend, not a phenomenon unique to China’s A-share market. The U.S. market has already begun this rotation, while the A-share market is just starting and still requires market validation, as capital investment and paying customers in this segment have yet to scale significantly. This means the tech rally isn’t over—it’s simply changing posture and continuing forward.

Question 3: Has the fundamental logic behind the tech bull market changed?

It has been extended, but not overturned. Five structural shifts—repricing under LTA frameworks, HBM customization, sustained AI demand, prolonged capacity expansion cycles, and geopolitical constraints—collectively stretch this memory industry cycle from the traditional 3–4 years to 5–7 years. However, don’t mythologize it as 'perpetual growth.' NAND is expected to reach a supply-demand inflection point in the second half of 2027—that’s currently the clearest industry forecast. The tech bull market is still intact.

Finally, three most important points:

First: The AI technology industry trend won’t vanish just because of a one-month crash. Memory giants have $518 billion under construction, $100 billion in locked-in orders, and a 146% price surge—these figures are far more real than any sentiment.

Second: 'Not dead' doesn’t mean 'imminent rally.' What August needs isn’t wishful thinking about a V-shaped recovery but patience to await secondary confirmation of the bottom. Only when the STAR Market 50 Index stabilizes with shrinking volume between 1,588 and 1,636 should serious consideration begin.

Third: Recognizing the direction of sector rotation is more important than fixating on short-term gains or losses in any single segment. The AI application wave is taking shape, the domestic substitution logic for semiconductor equipment is accelerating, and the memory chip supercycle continues—once you see the direction clearly, you won’t be blinded by a single day’s candlestick.

Important Disclaimer: All analysis in this article is based on publicly available data and industry logic, aiming to provide an analytical framework and market trend assessment. Companies and industry directions mentioned herein are for research reference only and do not constitute any investment advice. The market entails significant uncertainty; please make all investment decisions independently based on your own risk tolerance. Investing involves risks—proceed with caution.

The views expressed in this article represent the author’s personal judgment based on publicly available information and do not reflect the position of any institution.

15K Views