According to Wallstreetcn.com, the 13F U.S. equity holdings filings submitted for the second quarter by Wall Street’s top hedge funds or family offices show a significant divergence in their investment stances toward global AI computing power industry leaders such as the “chip foundry king” TSMC (TSM.US) and AI chip superpower Nvidia (NVDA.US). The most notable increase comes from Soros Capital Management, the alternative asset management family office founded by Robert Soros, eldest son of Wall Street financier and billionaire George Soros. This institution comprehensively increased its exposure to AI computing power, with 60% of its portfolio invested across memory, foundries, GPUs, and new AI cloud leaders, notably increasing its TSMC holdings by over 1,000% to 78,430 shares, worth about $37.46 million.
Third Point, run by billionaire Daniel Loeb, also increased its TSMC holdings by 67%, expanding the number of shares held from 275,000 to 460,000 in the second quarter.
Other large hedge funds or family offices also chose to increase their positions in this Asian giant, the world's largest chip manufacturer. Appaloosa raised its TSMC holding by 24%, Duquesne by 19%, Soroban Capital by 12%, Lansdowne Partners by 10%, and Altimeter Capital increased its position by a modest 4%.
Lone Pine, led by Stephen Mandel, made the most significant reduction, cutting its TSMC holdings by nearly 93% to 101,242 shares. Japanese investment legend Masayoshi Son’s SoftBank Group also cut its TSMC stake by 71%, reducing its share count from nearly 2 million shares to 565,000. TSMC now accounts for only 1.5% of SoftBank’s total holdings with a market value close to $270 million.
Other funds with significant TSMC reductions include Viking Global, SRS Investment Management, and Maverick Capital, which cut their holdings by 50%, 48%, and 43%, respectively. Although legendary asset manager Tiger Global reduced its TSMC position by about 20% to 4.88 million shares, TSMC still accounts for 9.7% of its total portfolio, remaining as its largest holding.
Soros' Eldest Son Reshapes Portfolio into an "AI Computing Power Industry Chain Encyclopedia"
Data shows Soros Capital’s total holdings were valued at $560 million in the second quarter, compared to just $230 million in the previous quarter, meaning that Soros Capital’s holdings more than doubled quarter-over-quarter during a period of profit-taking pressure in the stock market. In Q2, this financial giant added 18 new stock positions, increased 14, reduced 21, and fully sold off 7. The top ten holdings comprised about 60% of total market value.

Specifically, the Korean market ETF iShares MSCI Korea ETF (EWY.US), seen as a "weathervane for AI computing power investment," became Soros Capital’s largest holding, with about 212,200 shares worth about $42.84 million, accounting for 7.66% of the portfolio—a newly added position this quarter. This layout is akin to a concentrated bet on the bull run of the South Korean stock market via a country ETF, indirectly gaining upward exposure to memory chip and AI hardware supply chain leaders such as Samsung Electronics and SK Hynix.
American "memory chip giant" Micron Technology (MU.US) is the second largest holding, with approximately 36,800 shares worth $42.45 million, comprising 7.59% of the portfolio—the second largest new position in Q2. The simultaneous presence of Micron and EWY among top positions highlights Soros Capital’s concentrated bet on the super-cycle of HBM, server DRAM, and NAND storage fueled by AI servers.
“Global chip foundry king” TSMC (TSM.US) is the third largest holding, with 78,430 shares worth about $37.46 million and a portfolio weight of 6.70%. During Q2, 73,100 shares were added, causing the share count to surge 1371.48% quarter-over-quarter, and the portfolio weight leapt from 0.78% last quarter to 6.70%. This is not just an ordinary increase, but a high-conviction directional bet on advanced manufacturing processes, advanced packaging, and a certain increase in global AI chip capacity.
Global engineering machinery and power equipment giant Caterpillar (CAT.US) ranked fourth in holdings, with 32,800 shares valued at approximately $34.93 million, accounting for 6.25% of the portfolio; Q2 saw an increase of 16,300 shares, a surge of 98.79%, raising the portfolio weight from 5.08% to 6.25%. This allocation shows that, beyond chips, Soros Capital is also betting on “shovel sellers” benefiting from expansions in data center, power equipment, mining, and energy infrastructure capex.
Semiconductor equipment giant Applied Materials (AMAT.US) ranked fifth, with 47,750 shares valued around $34.52 million, accounting for 6.17% of the portfolio, and is a new Q2 position. The world’s largest supplier of semiconductor equipment, Applied Materials, covers key chip manufacturing segments including advanced logic chips, HBM, wafer-level advanced packaging, and materials engineering—making it a major beneficiary of capex expansion from chip design toward manufacturing equipment for AI computing power.

Among the top ten holdings, Soros Capital’s sixth to tenth largest positions as of Q2 are: Ranpak Holdings (PACK.US), Flex (FLEX.US), Advanced Micro Devices (AMD.US), ProPetro Holding (PUMP.US), and ON Semiconductor (ON.US). PACK: approx. 4.63 million shares, worth about $33.85 million (6.05% of the portfolio), unchanged in quantity; FLEX: new 198,900 shares, $32.24 million (5.76%); AMD: new 52,015 shares, $30.22 million (5.40%); PUMP: about 1.44 million shares, $20.70 million (3.70%), 352,539 shares added in Q2 (up 32.31%); ON: new 217,800 shares, $20.59 million (3.68%). The top ten holdings totaled about $329.8 million, accounting for 58.96% of the 13F portfolio.
In other significant Soros Capital positions, the firm newly established a 30,700 share position in VanEck Semiconductor ETF (SMH.US) worth about $20.14 million (3.60%); increased its ASML (ASML.US) holding by 191.11% to 9,956 shares, worth about $19.81 million (3.54%); massively increased Nvidia (NVDA.US) by 566.21% to 77,300 shares, worth about $15.47 million (2.77%); and established a new position in 55,200 shares of “AI new cloud leader” Nebius (NBIS.US), worth about $15.24 million (2.73%).
Together with TSMC, Micron, Applied Materials, AMD, ON Semiconductor, and EWY, Soros Capital has built a multi-layered “encyclopedia of AI computing power infrastructure” that covers “Korea memory - foundry - semiconductor equipment leaders - AI chip giants - chip testing and manufacturing agents - cloud inference power infrastructure.”
At the same time, Soros Capital also increased positions in Amer Sports (AS.US) by 1420.97% and Generac (GNRC.US) by 470.99%, and slightly increased TeraWulf (WULF.US) by 6.93% and American Tower (AMT.US) by 6.41%. On the other hand, it sharply reduced GE Vernova (GEV.US) by 52.17%, and trimmed Amazon (AMZN.US) by 15.29%, Texas Instruments (TXN.US) by 9.88%, Alcon (ALC.US) by 16.99%, and Visa (V.US) by 29.00%. Overall, this 13F report does not simply chase the high beta of Nvidia alone but significantly shifts investment focus toward AI storage, advanced manufacturing, semiconductor equipment, data center infrastructure, and industrial energy capex. Through country ETF, Micron, AMD, ASML, Applied Materials, Caterpillar, oil services, and consumer beneficiaries, the portfolio diversifies AI positions and demonstrates a strategy of “betting on the main wave of AI computing power while controlling the concentration risk of single tech giants.”
Is This a Big Bet on Memory Chips? Soros' Eldest Son Sharply Boosts Korean Stocks and Micron
Soros Capital seeks to simultaneously capture the multiple profit levers brought by HBM/DRAM price increases, advanced process expansion, wafer equipment orders, GPU competition, and rising AI cloud demand, while diversifying single tech stock risk through Korean ETF, Caterpillar, oil services, and consumer assets. However, a 59% broad AI exposure also means the portfolio is highly sensitive to capex cuts, rising long-end yields, and the reversal of the semiconductor cycle. Capital-intensive stocks such as WULF and NBIS come with additional financing and execution risk.
The latest portfolio and investment directions from Soros Capital, overseen by Soros’ eldest son, highlight that Wall Street’s AI computing power thematic trades are spreading from single GPU leaders to storage, foundries, equipment, server systems, and computing cloud platforms. Stock selection standards are also shifting from concept elasticity to order visibility, financing ability, and free cash flow realization.
Based on changes in portfolio weights, the top five buys for Soros Capital in Q2 under Robert Soros were: iShares MSCI Korea ETF (EWY.US), Micron Technology (MU.US), Applied Materials (AMAT.US), TSMC (TSM.US), and Flex (FLEX.US), with respective portfolio weight increases of 7.66, 7.59, 6.17, 5.91, and 5.76 percentage points—highlighting the company’s massive bet on the computing power infrastructure supply chain spanning Korean memory chips, HBM and server DRAM, advanced manufacturing, semiconductor equipment, and AI server power, liquid cooling, and systems integration. Micron, Applied Materials, and Flex were newly established top holdings, and TSMC’s share count soared 1371.48%.

The top five sell-downs by Soros Capital in Q2, directed by the eldest Soros, were: Comfort Systems USA (FIX.US), a leader in data center HVAC and MEP engineering; data center power equipment giant GE Vernova (GEV.US); ophthalmic medical device firm Alcon (ALC.US); analog chip giant Analog Devices (ADI.US); and MEP engineering contractor EMCOR Group (EME.US). Each saw respective portfolio weight declines of 6.91, 5.15, 2.15, 2.14, and 1.97 percentage points, further demonstrating that the firm is taking profits from earlier high-valued and fast-growing AI power, HVAC, and engineering beneficiaries, while reducing medical and traditional analog chip holdings—rotating capital from "secondary AI data center construction beneficiaries" toward core AI computing power bottleneck assets such as memory, foundry, and semiconductor equipment with more direct connections to AI capex. This doesn’t mean a bearish stance on AI, but rather aims to reprioritize risk-reward inside the AI computing power industry chain.