Gao Yi Asset recently disclosed its U.S. equity holdings as of the end of the second quarter on the U.S. Securities and Exchange Commission website. According to figures cited from Simuwang, Gao Yi’s offshore fund held 19 U.S. stock positions during the quarter, with a total market value of about $979 million, or roughly RMB 6.6 billion.
The most visible portfolio change was a large increase in Taiwan Semiconductor Manufacturing Co. (TSMC). Gao Yi also sharply raised its positions in Kanzhun, the operator of Boss Zhipin, and Trip.com, with share counts up 175% and 53%, respectively.
In memory chips, Gao Yi’s moves closely matched those of Dongfang Harbor’s offshore fund, which is managed by Dan Bin. Gao Yi increased its Micron holding by more than 283% and its SanDisk stake by about 192%. The two names ranked as its sixth- and seventh-largest holdings. Dongfang Harbor’s offshore fund also opened a new position in SanDisk during the quarter, with an end-period market value of more than $200 million.
By contrast, several private funds moved to reduce exposure to Nvidia, the leading AI chip company. Gao Yi cut its Nvidia position by more than 70% in the second quarter, Jinglin Asset exited the stock entirely, and Dongfang Harbor reduced its holding by about 15.7%.
The article cited industry analysis saying the shift should not be read as a denial of the AI trend. Instead, the investment logic is moving from expectation-driven trades to delivery-driven trades. The market is no longer focused only on price hikes and supply shortages, but on whether companies can turn favorable industry conditions into actual orders, capacity release and healthy cash flow.
TSMC became Gao Yi’s largest holding
Gao Yi raised TSMC from its third-largest U.S. holding to the top position in the second quarter. The stake was valued at $238 million at quarter-end, representing about 24% of its total U.S. equity portfolio.
The article said TSMC posted $39.45 billion in revenue in the second quarter, while net profit hit a record high for a fifth straight quarter and net margin stayed around 40%. In a capital-intensive and highly cyclical semiconductor industry, that level of profitability stands out.
It described TSMC’s moat in two layers. The first is advanced process technology. Most advanced-node AI chips, whether from Nvidia, Advanced Micro Devices (AMD), Broadcom, Google, Microsoft or Meta, still rely on TSMC. The second, and more important in the article’s framing, is advanced packaging capacity for CoWoS. Demand for HBM and advanced packaging in AI chips has surged, and CoWoS capacity may remain undersupplied through 2026.
Process technology determines whether a chip can be manufactured at all. Packaging determines whether the completed chip can meet AI interconnect requirements after production. Put together, those two links place TSMC in an unusual position inside the AI supply chain.
The article called TSMC a classic bottleneck asset. The pricing power of such an asset comes from rigid supply: demand can rise quickly, but supply cannot expand at the same pace in the medium term. In comparison, the AI chip design segment remains more open to competition. Nvidia has the CUDA ecosystem, AMD has a cost-performance argument, and cloud companies have incentives to develop their own application-specific integrated circuits, or ASICs.
Manufacturing is different. The article said TSMC controls most of the global advanced-node foundry market, and advanced packaging capacity is scarce as well. Those physical capacity constraints are difficult to bypass in the medium term.
Capital spending raises the barrier even more. The article noted that a leading-edge wafer fab can require investment of tens of billions of dollars, with construction taking two to three years, followed by additional time for yield ramp-up. Even if rivals are willing to spend heavily, capacity release would still come much later.
In that window, TSMC can continue to benefit from the pricing power that comes with tight supply. The article said Gao Yi’s increase in TSMC amounts to a bet that profit allocation in AI is shifting from chip design toward manufacturing. No matter which company ends up designing the strongest chip, the chip still has to be turned into silicon on TSMC’s production lines.
Memory stocks emerged as another focus
Alongside TSMC, memory was another major direction where Gao Yi and Dongfang Harbor moved in parallel during the second quarter. The article argued that memory is no longer just a routine subsegment of semiconductors, but the second bottleneck in the AI supply chain.
HBM, or high-bandwidth memory, was described as the key to solving the memory-wall problem in AI chips. Even if compute power is strong, performance can still be constrained if data cannot move fast enough from memory to the compute core.
HBM is difficult to manufacture, slow to ramp in yield and limited in capacity. It also crowds out standard DRAM capacity. That creates a structural setup: AI demand drives HBM into shortage, memory makers redirect capacity toward HBM, and the reduced supply of standard DRAM and NAND pushes up prices even outside direct AI demand.
Under that logic, memory shifts from a pure cyclical product into a sector with both cyclical and growth characteristics. The article said the old pattern in memory was straightforward: oversupply led to price declines, losses and production cuts; undersupply led to price increases, expansion and profits; then the cycle repeated. This time, AI demand is layered onto the cycle, and the supply impact from HBM changes the capacity structure of the entire industry.
Micron was identified as a core HBM supplier, while SanDisk was described as an important NAND flash player. In the article’s view, funds increasing exposure to memory were effectively betting on two things at once: earnings recovery from a cyclical bottom and sustained demand from expanding AI compute.
The article also said supply discipline in memory has changed in recent years. After multiple rounds of fierce price competition, the industry has become more oligopolistic, and major producers are showing more restraint in capital spending than in the previous cycle. That means when demand recovers, supply may not catch up quickly, making price elasticity larger.
At the same time, leading memory companies have already gone through several quarters of losses, inventory reduction and output cuts, leaving valuations far less crowded than Nvidia’s. The article’s conclusion was that buying memory stocks near the bottom of a cycle offers better odds, and once a confirmed turnaround appears, the scale of earnings recovery can exceed market expectations, lifting both profits and valuation.
Gao Yi and Dongfang Harbor increasing memory exposure at the same time was presented as a sign that institutional investors are rethinking the sector’s long-term value. The market used to price memory mainly as a cyclical trade. AI demand now adds a growth element, opening the door to a valuation reset.
Why Nvidia was cut
As for Nvidia, Gao Yi’s second-quarter holding was reduced to 80,000 shares after a cut of more than 70%. Jinglin Asset sold out completely, and Dongfang Harbor cut its position by about 15.7%. Those synchronized reductions became the most closely watched part of the filings.
The article argued that the reason is not difficult to understand. Nvidia’s expectation gap is narrowing, and its valuation has become more sensitive to marginal changes. It noted that Nvidia’s share price had risen nearly tenfold from its low over the past two years, and its market capitalization was at one point the largest in the world.
At that level, the market’s demands become tougher. The article said Nvidia now has to deliver not only high growth, but growth that continues to exceed already elevated expectations, while also proving that gross margin will not be eroded by higher HBM costs, advanced packaging expenses and customers’ self-developed ASICs. In that framing, the main risk comes from a redistribution of profits across the supply chain.
The article listed several pressure points: rising prices for TSMC’s advanced packaging, rising HBM prices, cloud companies diverting some demand with in-house chips, and custom ASICs taking part of the inference market. None of those factors necessarily break Nvidia’s business, but they could compress the flexibility of its future profit margins.
On valuation, the article divided AI pricing into two phases. From 2023 to 2024, the market focused on long-dated upside, asking whether prices could keep rising, whether shortages would deepen and how much money companies might make in the future. In that phase, Nvidia was the strongest target because it benefited most directly from the race to build compute capacity, and investors were willing to pay a high premium for that future potential.
In the article’s view, the focus changes after 2025. Questions about where profits come from, how long they last and whether they can be turned into cash flow become more important. At that point, the valuation anchor shifts from imagination to discounted cash flow. TSMC has verifiable capacity, orders and profits. Memory has both cyclical pricing recovery and incremental AI demand. Nvidia also has strong cash flow, but its valuation already reflects optimistic assumptions for years ahead.
From that perspective, cutting Nvidia does not mean rejecting AI. The money is moving from the most crowded trade toward bottleneck links with more verifiable earnings and lower valuations. The direction for Nvidia may still be intact, but after such a large rise, more upside requires fundamentals to keep beating expectations, while any disappointment could trigger a sharp pullback.
What the 13F filings suggest about AI profit migration
The article ended by noting that 13F filings are delayed, so simply copying positions can leave ordinary investors buying at elevated levels. Even so, the portfolio moves by Gao Yi and other funds offer a way to think about where excess profits are moving inside the AI supply chain.
Its framework is that excess profits migrate along the chain at different stages of a technology wave. At the start, profits concentrate in chip design because the company that can produce the strongest compute enjoys pricing power. As AI chips scale, bottlenecks shift toward manufacturing and memory. CoWoS capacity at TSMC, HBM supply and the broader memory cycle become the next pools of profit.
The article added that profits may later move further downstream to application companies once compute infrastructure becomes widely available. At that stage, the eventual winners may be firms that use AI to cut costs, improve efficiency and generate revenue, rather than companies that sell AI hardware.
Under that interpretation, the second-quarter rebalancing looks like sector rotation on the surface, but in substance it reflects a change in AI investing from pricing technological possibility to pricing physical constraints. In the next round of profit distribution, the companies that control capacity and memory supply may be in a stronger position.
The article was sourced from the WeChat public account Dongzhen Shanglue and credited to the author Qiqi Aichuiniu.