NVIDIA is facing collective stake reduction, what are private equity firms sensing?

Recently, GaoYi Asset disclosed its US stock positions as of the end of the second quarter on the official website of the U.S. Securities and Exchange Commission (SEC). According to statistics from Simuwang, the overseas fund of GaoYi Asset held a total of 19 US stock targets in this quarter, with a total market value of positions of about 979 million US dollars, equivalent to about 6.6 billion RMB.

The position adjustment shows that the most notable change of GaoYi Asset in the second quarter was the massive increase in positions in TSMC, the semiconductor foundry giant. TSMC's recent performance has continuously exceeded expectations, and it has achieved profit growth for ten consecutive quarters.

In the same period, GaoYi Asset also significantly increased its holdings in BOSS Zhipin and Trip.com, with the number of shares held increasing by 175% and 53% respectively.

In the field of memory chips, the operation of GaoYi Asset coincided with that of the overseas fund of Oriental Harbour led by Dan Bin.

In the second quarter, the number of shares GaoYi Asset held in Micron increased by more than 283%, and the number of shares held in SanDisk increased by about 192%. The two companies ranked its sixth and seventh largest heavyweight stocks respectively.

The overseas fund of Oriental Harbour also newly bought SanDisk in this quarter, and the market value of its shares at the end of the period exceeded 200 million US dollars.

However, for NVIDIA, the leader of AI chips, many private equity firms unanimously chose to reduce their positions.

GaoYi Asset reduced its holdings of NVIDIA by more than 70% in the second quarter, Greenwoods Asset directly liquidated its position, and Oriental Harbour also reduced its holdings by about 15.7%.

Industry insiders pointed out in the analysis that this does not deny the trend of the AI industry, but the investment logic has shifted from "expectation-driven" to "realization-driven". The market no longer only focuses on price increases and supply-demand gaps, but pays more attention to whether enterprises can transform industry prosperity into actual orders, capacity release and healthy cash flow.

This means that internal differentiation in the AI sector will become the norm, and only companies that can truly achieve value expansion can continuously obtain valuation support.

Why Has TSMC

Become The Largest Heavyweight Stock?

GaoYi bought TSMC from the third largest heavyweight stock to the first largest heavyweight stock in the second quarter, with a shareholding market value of 238 million US dollars, accounting for about 24% of the total US stock positions. This action is more convincing than any point of view.

TSMC's revenue in the second quarter was 39.45 billion US dollars, its net profit hit a record high for five consecutive quarters, and its net profit margin was around 40%. In the semiconductor industry, which is asset-heavy and strongly cyclical, this profitability is very rare.

TSMC's moat comes from two levels.

The first level is the advanced process. No matter the self-developed AI chips of NVIDIA, AMD, Broadcom, Google, Microsoft and Meta, the vast majority of advanced process chips cannot bypass TSMC.

The second level is more critical, which is the advanced CoWoS packaging capacity. The demand for HBM and advanced packaging for AI chips is exploding, and CoWoS capacity may be in short supply until 2026.

The process determines whether the chip can be manufactured, and the packaging determines whether the manufactured chip can meet the interconnection requirements of AI computing power.

The superposition of these two links puts TSMC in an extremely special position in the AI industrial chain.

From the perspective of the industry, TSMC is obviously a typical bottleneck asset.

The pricing power of bottleneck assets comes from supply rigidity, that is, demand can rise rapidly, but supply cannot expand synchronously in the medium term. The competitive landscape of the AI chip design link is relatively open. NVIDIA has the CUDA ecosystem, AMD has the advantage of cost performance, and cloud vendors have the motivation to develop their own ASICs, and everyone wants to take a share of the design profit.

However, the competitive landscape of the manufacturing link is highly concentrated. TSMC accounts for most of the global share of advanced process foundry, and the advanced packaging capacity is equally scarce. This physical capacity constraint can hardly be bypassed in the medium term.

The gap in capital expenditure further strengthens TSMC's position. The investment in an advanced process wafer fab often amounts to tens of billions of dollars, the construction period lasts two to three years, and the yield ramp-up requires additional time.

This means that even if competitors are willing to spend money to catch up, capacity release will have to wait for a long time.

During this period of time, TSMC can continuously enjoy the pricing power brought by short supply. By increasing its positions in TSMC, GaoYi is betting that the AI profit distribution right will shift from the design end to the manufacturing end. No matter who designs the strongest chip, it will eventually be turned into silicon wafers on TSMC's production lines.

This "must-pass" attribute makes TSMC one of the links with the highest earnings visibility in the AI industrial chain.

Where Exactly Is The Opportunity For Memory Stocks?

The synchronous operation direction of GaoYi and Oriental Harbour can be described as coinciding, which also relatively shows that memory is no longer an ordinary branch, but the second bottleneck of the AI industrial chain.

HBM, namely High Bandwidth Memory, is believed to be familiar to everyone. No matter how strong the computing power of AI chips is, if data cannot be quickly transferred from memory to the computing core, the performance will be stuck at the "memory wall".

HBM is the key to solving this problem.

However, HBM is difficult to manufacture, the yield ramps up slowly, the capacity is limited, and it will squeeze the capacity of standard DRAM.

Thus, a classic structural price increase logic emerges: AI demand drives HBM to be in short supply, memory manufacturers shift capacity to HBM, the capacity of standard DRAM and NAND decreases, and non-AI memory also rises in price accordingly.

This means that the memory industry has changed from a pure cyclical product in the past to a dual attribute of "cyclical + growth".

In the past, the logic of the memory industry was very simple: oversupply led to price cuts, losses and production cuts; undersupply led to price increases, capacity expansion and profit, and the cycle repeated. But this round of cycle is superimposed with new AI demand, and the supply shock of HBM has changed the capacity structure of the entire industry.

Micron is the core supplier of HBM, and SanDisk is an important player in NAND flash memory. The essence of GaoYi and other institutions increasing their positions in memory is to bet on two things at the same time: the profit repair brought by the bottoming and reversal of the memory cycle, and the continuous pull of AI computing power on memory demand.

What is more noteworthy is that the supply discipline of the memory industry has changed significantly in the past few years.

After several rounds of brutal price wars, the memory industry has formed an oligopoly pattern, and the capital expenditure of major manufacturers is more restrained than the previous cycle, and they no longer expand capacity blindly.

Then, it means that when demand picks up, supply cannot keep up quickly, and the price elasticity will be greater.

At the same time, leading memory manufacturers have experienced consecutive quarters of losses, destocking and production cuts, and their valuations are far less crowded than NVIDIA's.

Buying memory stocks at the bottom of the cycle naturally has better odds.

Once the cycle reversal is confirmed, the magnitude of earnings repair often exceeds market expectations. This is the typical Davis double-click, that is, the simultaneous improvement of performance and valuation.

GaoYi and Oriental Harbour increased their positions in memory at the same time, indicating that institutional investors are re-evaluating the long-term value of the memory industry.

In the past, the market regarded memory as a cyclical stock, and the valuation was very low; now AI demand brings growth attributes to memory, and the valuation system is facing revaluation.

This cognitive difference is often the source of excess returns.

Why Is NVIDIA Collectively Reduced In Positions?

As for the collective reduction of positions in NVIDIA, it is actually not difficult to understand. GaoYi only held 80,000 shares of NVIDIA in the second quarter, with a reduction ratio of more than 70%; Greenwoods directly liquidated its position; Oriental Harbour reduced its holdings by about 15.7%.

Three giants reduced their positions at the same time, and people generally think that it is "AI peaking".

In fact, the real reason is that the expectation gap of NVIDIA is narrowing, and the sensitivity of valuation to marginal changes is rising.

In the past two years, NVIDIA's stock price has risen nearly ten times from the low point, and its market value once ranked first in the world.

However, the higher the price goes, the more demanding the market is for it. It not only needs high growth, but also high growth exceeding expectations, and it needs to prove that its gross profit margin will not be eroded by HBM costs, advanced packaging expenses and customers' self-developed ASICs. The biggest risk NVIDIA faces comes from the redistribution of industrial chain profits.

TSMC's advanced packaging is rising in price, HBM is rising in price, cloud vendors' self-developed chips are diverting, and custom ASICs are eroding part of the reasoning market. These factors may not bring down NVIDIA, but will compress its profit margin elasticity in the future.

From the perspective of valuation logic, the market's pricing of AI is shifting from expectation-driven to realization-driven. From 2023 to 2024, the core logic of AI investment is the long-term imagination space: will prices rise, will there be more shortages, and how much money can be made in the future. At this stage, NVIDIA is the strongest target, because it benefits most directly from the computing power arms race, and the market is willing to pay a high premium for the long-term space.

After 2025, questions such as where these profits come from, how long they will last, and whether they can be realized as cash flow will become more important. At this time, the valuation anchor will shift from imagination space to discounted cash flow.

TSMC has verifiable capacity, orders and profits; memory has cyclical reversal price increases and new AI demand; although NVIDIA also has strong cash flow, its valuation has included optimistic expectations for many years in the future.

When everyone knows that NVIDIA is doing well, its stock price has already reflected this good performance.

Reducing positions in NVIDIA at this time does not mean denying AI.

These funds are moving their positions from the most crowded transactions to bottleneck links with verifiable earnings and cheaper valuations.

Isn't this very similar to the so-called "odds"?

Investment depends not only on the direction, but also on the odds.

NVIDIA's direction is still correct, but the odds have dropped significantly.

After a stock rises ten times from the low level, the room for continued rise needs to be supported by continuous performance exceeding expectations, and any underperformance may trigger a sharp correction.

In contrast, TSMC and memory stocks have lower valuations, higher certainty of earnings repair, and better odds. Institutional investors reducing their positions in NVIDIA seem to be bearish on the surface, but in fact they are reallocating the risk-return ratio.

Furthermore, NVIDIA's "easiest money to make" has been earned.

The next stage will enter the "profit redistribution" phase, and part of its excess returns may be taken away by links such as TSMC and memory. The profit pool of the AI industrial chain is being re-divided, and the excess profits at the design end begin to flow to the manufacturing end and the memory end. This profit migration is the underlying logic for many private equity firms to adjust their positions simultaneously in the second quarter.

For ordinary investors, the 13F report is lagging, and directly copying positions often leads to buying at a high level.

However, the position adjustments of institutions like GaoYi provide an important analysis method: at different stages of the technology wave, excess profits will migrate along the industrial chain.

In the earliest stage, profits were concentrated at the chip design end, because whoever could design the strongest computing power would have pricing power.

As computing power chips are released in large quantities, the bottleneck begins to shift to the manufacturing end and the memory end. TSMC's CoWoS capacity, HBM supply, and memory cycles have become new profit pools.

In the future, when computing power infrastructure is sufficiently popularized, profits may further migrate to downstream application ends.

At that time, the real winners may be companies that use AI to reduce costs, increase efficiency and generate revenue, rather than companies that sell AI hardware.

Therefore, this round of position adjustment is superficially industry rotation, but in essence, AI investment is shifting from "technical possibility" pricing to "physical constraint" pricing. Whoever can control the capacity and grasp the memory can occupy a more favorable position in the profit distribution of the next stage.

This may be the interesting part of the second quarter 13F report.

This article is from the WeChat official account "Dongzhen Strategy", the author is Qiqi Loves Bragging, and is published with authorization from 36Kr.