Recently, Gao Yi Asset Management disclosed its U.S. stock holdings for the end of the second quarter on the U.S. Securities and Exchange Commission (SEC) website. According to statistics from Simuwang, Gao Yi's overseas fund held 19 U.S. stock positions in the quarter, with a total market value of approximately $979 million, equivalent to about 6.6 billion RMB.
The portfolio adjustments show that the most significant change for Gao Yi in the second quarter was a substantial increase in its position in semiconductor foundry giant TSMC. TSMC's recent performance has consistently exceeded expectations, achieving profit growth for ten consecutive quarters.
During the same period, Gao Yi also significantly increased its holdings in BOSS Zhipin and Trip.com, with share counts rising by 175% and 53% respectively.
In the memory chip sector, Gao Yi's moves coincided with the operations of the overseas fund managed by Dan Bin's Oriental Harbor.
In Q2, Gao Yi increased its holdings in Micron by over 283% and in SanDisk by about 192%, making them its sixth and seventh largest holdings respectively.
Oriental Harbor's overseas fund also newly entered a position in SanDisk during the quarter, with an ending market value exceeding $200 million.
However, when it came to AI chip leader Nvidia, several private equity firms unanimously chose to retreat.
Gao Yi reduced its Nvidia holdings by over 70% in Q2, Greenwoods Asset Management cleared its position entirely, and Oriental Harbor cut its holdings by about 15.7%.
Industry analysts point out that this is not a rejection of the AI industry trend, but rather a shift in investment logic from "expectation-driven" to "realization-driven." The market is no longer solely focused on price increases and supply-demand gaps, but is placing greater emphasis on whether companies can translate industry growth into actual orders, capacity utilization, and healthy cash flow.
This means differentiation within the AI sector will become the norm, with only companies that can truly achieve value expansion continuing to receive valuation support.
Why Has TSMC Become the Largest Holding?
In Q2, Gao Yi moved TSMC from its third-largest holding to its largest holding, with a market value of $238 million, accounting for about 24% of its total U.S. stock portfolio. This action speaks louder than any opinion.
TSMC's Q2 revenue was $39.45 billion, with net profit hitting historical highs for five consecutive quarters, and a net profit margin around 40%. In a capital-intensive, strongly cyclical industry like semiconductors, such profitability is exceptionally rare.
TSMC's moat comes from two levels.
The first level is advanced process technology. Whether it's chips from Nvidia, AMD, Broadcom, or the self-designed AI chips from Google, Microsoft, or Meta, the vast majority of advanced process chips cannot bypass TSMC.
The second level, which is more crucial, is advanced packaging CoWoS capacity. The demand for HBM and advanced packaging from AI chips has exploded, and CoWoS capacity may remain in short supply until 2026.
The process determines if a chip can be manufactured, while packaging determines if it can meet the interconnect requirements for AI computing power after being made.
The combination of these two aspects places TSMC in an extremely unique position within the AI supply chain.
From an industrial perspective, TSMC clearly belongs to a typical bottleneck asset.
The pricing power of bottleneck assets comes from supply inelasticity—demand can rise quickly, but supply cannot expand at the same pace in the medium term. The competitive landscape in AI chip design is relatively open: Nvidia has its CUDA ecosystem, AMD has cost-performance advantages, cloud providers have motives for self-designed ASICs—each wants a share of the design profits.
However, the competitive landscape in manufacturing is highly concentrated, with TSMC holding a large share of the global advanced foundry market, and advanced packaging capacity also being scarce. This constraint of physical capacity is nearly impossible to bypass in the medium term.
The capital expenditure gap further reinforces TSMC's position. Building an advanced process wafer fab requires investments of tens of billions of dollars, with a construction cycle of two to three years, plus additional time for yield ramp-up.
This means that even if competitors are willing to spend money to catch up, capacity release will have to wait a long time.
During this period, TSMC can continue to enjoy pricing power brought by supply shortages. By increasing its position in TSMC, Gao Yi is betting on a shift in AI profit allocation power from the design end to the manufacturing end. No matter who designs the most powerful chip, it ultimately has to become a silicon wafer on TSMC's production lines.
This "must-pass-through" attribute makes TSMC one of the segments in the AI supply chain with the highest visibility of profitability.
Where Exactly is the Opportunity in Memory Stocks?
The synchronized trading direction of Gao Yi and Oriental Harbor in the same period can be described as a coincidence, and it also relatively indicates that memory is no longer an ordinary sub-sector but the second bottleneck in the AI supply chain.
HBM, or High Bandwidth Memory, is likely familiar to everyone. No matter how strong an AI chip's computing power is, if data cannot be quickly moved from memory to the computing core, performance will be bottlenecked by the "memory wall."
HBM is the key to solving this problem.
However, HBM is difficult to manufacture, has slow yield ramp-up, limited capacity, and squeezes capacity for standard DRAM.
This creates a classic structural price increase logic: AI demand pulls HBM into short supply, memory manufacturers shift capacity to HBM, capacity for standard DRAM and NAND decreases, and non-AI memory also follows with price increases.
This means the memory industry is transitioning from a pure cyclical product to having dual attributes of "cycle + growth."
In the past, the logic of the memory industry was simple: oversupply led to price cuts, losses, production cuts; undersupply led to price increases, capacity expansion, profits, and the cycle repeated. But this cycle is superimposed with new AI demand; HBM's supply shock has changed the entire industry's capacity structure.
Micron is a core supplier of HBM, and SanDisk is a major player in NAND flash. By increasing their positions in memory, funds like Gao Yi are essentially betting on two things simultaneously: the profit recovery brought by the memory cycle bottoming and reversing, and the sustained pull of AI computing power on memory demand.
More noteworthy is the clear change in supply discipline in the memory industry over the past few years.
After multiple rounds of brutal price wars, the memory industry has formed an oligopoly. The capital expenditures of major manufacturers are more restrained compared to the previous cycle, with no blind capacity expansion.
Therefore, when demand recovers, supply cannot keep up quickly, leading to greater price elasticity.
Meanwhile, leading memory companies have experienced consecutive quarters of losses, inventory drawdowns, and production cuts, making their valuations far less crowded than Nvidia's.
Buying memory stocks at the cycle bottom naturally offers better risk-reward ("赔率").
Once a cycle reversal is confirmed, the magnitude of profit recovery often exceeds market expectations—this is a typical Davis double play, where both earnings and valuations rise simultaneously.
The simultaneous increase in memory positions by Gao Yi and Oriental Harbor indicates that institutional investors are reassessing the long-term value of the memory industry.
In the past, the market treated memory as a cyclical stock, assigning it low valuations; now, AI demand has brought a growth attribute to memory, necessitating a revaluation of the valuation system.
Such a gap in perception is often the source of excess returns.
Why Was Nvidia Collectively Reduced?
As for the collective reduction in Nvidia holdings, it's not difficult to understand either. In Q2, Gao Yi's Nvidia position was reduced to just 80,000 shares, a decrease of over 70%; Greenwoods cleared its position entirely; Oriental Harbor reduced its holdings by about 15.7%.
The synchronized selling by these three giants might lead many to think it signals "AI peaking."
The real reason, however, is that the expectation gap for Nvidia is narrowing, and its valuation is becoming more sensitive to marginal changes.
Over the past two years, Nvidia's stock price has risen nearly tenfold from its lows, and its market cap once ranked first globally.
But the higher it climbs, the more demanding the market becomes—not only requiring high growth but also growth that exceeds high expectations, and proof that gross margins won't be eroded by HBM costs, advanced packaging fees, or customer self-designed ASICs. Nvidia's biggest risk comes from profit reallocation within the supply chain.
TSMC's advanced packaging is seeing price increases, HBM prices are rising, cloud providers' self-designed chips are diverting demand, and custom ASICs are encroaching on part of the inference market. These factors may not cripple Nvidia, but they could compress its future margin elasticity.
From a valuation logic perspective, the market's pricing of AI is shifting from expectation-driven to realization-driven. In 2023-2024, the core logic of AI investment was the future imagination space: would prices rise, would shortages worsen, how much money could be made in the future. During this phase, Nvidia was the strongest play, as it benefited most directly from the computing arms race, and the market was willing to pay a high premium for future potential.
By 2025 and beyond, questions about where these profits will actually come from, how long they will last, and whether they can be realized as cash flow become more important. At this point, the valuation anchor shifts from imagination space to discounted cash flow.
TSMC has verifiable capacity, orders, and profits; memory has the price increases from cycle reversal and new AI demand; although Nvidia also has strong cash flow, its valuation already incorporates optimistic expectations for many years ahead.
When everyone knows Nvidia is good, its stock price already reflects that goodness.
Selling Nvidia now does not equate to rejecting AI.
This capital is moving from the most crowded trade to bottleneck segments with verifiable profits and cheaper valuations.
Doesn't this sound a lot like the so-called "赔率" (risk-reward ratio)?
Investment isn't just about direction; it's also about the risk-reward ratio.
Nvidia's direction is still correct, but its risk-reward ratio has significantly declined.
After a stock rises tenfold from a low, the space for continued upward movement requires fundamentals to consistently exceed expectations, and any shortfall can trigger a sharp correction.
In comparison, TSMC and memory stocks have lower valuations, higher certainty of profit recovery, and better risk-reward ratios. Institutional investors reducing Nvidia holdings might seem bearish on the surface, but in reality, they are reallocating based on risk-reward profiles.
Furthermore, the "easiest money" from Nvidia has likely been made.
The next phase enters the "profit reallocation" stage, where its excess returns might be partially shared by segments like TSMC and memory. The profit pool of the AI supply chain is being redrawn, with excess profits from the design end beginning to flow towards the manufacturing end and the memory end. This profit migration is the underlying logic behind the synchronized portfolio adjustments by multiple private equity firms in Q2.
For ordinary investors, 13F reports have a lag, and directly copying trades often means buying at high points.
However, the portfolio adjustments by funds like Gao Yi provide an important analytical insight: at different stages of a technological wave, excess profits migrate along the supply chain.
Initially, profits concentrate in the chip design end, because whoever designs the strongest computing power holds pricing power.
As computing chips scale, bottlenecks begin shifting to the manufacturing and memory ends. TSMC's CoWoS capacity, HBM supply, and the memory cycle become the new profit pools.
In the future, when computing infrastructure becomes sufficiently widespread, profits may further migrate downstream to the application end.
At that time, the real winners might be companies that use AI to reduce costs, improve efficiency, and create revenue, rather than those selling AI hardware.
Therefore, this round of portfolio adjustments, on the surface, is sector rotation; in essence, it's AI investment shifting from pricing "technological possibility" to pricing "physical constraints." Those who control capacity, who can bottleneck memory, will occupy a more advantageous position in the next phase of profit allocation.
This is perhaps the most interesting aspect of the Q2 13F reports.
This article is from the WeChat public account "Dong Zhen Shang Lue," author: Qi Qi Ai Chui Niu
