Author: Jim, MSX Maitong
Editor: Frank, MSX Maitong
The least valuable takeaway from each quarter's 13F filings is likely:
"Which stock did the big shots buy this time?"
This is because 13F filings are inherently a lagging snapshot of holdings.
According to SEC rules, institutions have up to 45 days after the end of the quarter to disclose their holdings. The latest round of Q2 2026 13F filings reflects holdings as of June 30th, with the concentrated disclosure deadline being August 14th. More importantly, 13F primarily covers eligible long positions in US-listed securities and does not fully reflect short positions in stocks, etc.
Therefore, it is not suitable for real-time "copy-trading."
But from another perspective, the value of 13F is actually very high—what have the truly big funds been buying and selling over the past three months?
After analysis, we identified a crucial signal: AI has not receded, but Wall Street is becoming "pickier" about AI.
I. The AI Consensus Remains, but the "Herd Consensus" is Loosening
If you only look at a few star funds, it's easy to be misled by individual trades.
What's truly worth paying attention to is the change across the entire institutional community. Reuters analyzed Q2 13F filings from 6,371 entities including pension funds, hedge funds, and wealth management firms and found:
- Close to 44% of institutions reduced their holdings in the "Magnificent Seven," while about 42% chose to initiate or increase positions—nearly a tie;
- However, for semiconductors, the sentiment remains noticeably bullish: about 48% of institutions were net buyers, with only 34.5% being net sellers;
- Software is the opposite. Among a group of major software companies, net sellers accounted for 28.2%, slightly higher than net buyers at 26.3%;
This indicates that the AI consensus still exists, but it is rapidly fragmenting internally.
After all, if Wall Street were to systematically reject AI, the first sign would likely be a unified retreat from the semiconductor, computing power, and data center supply chain.
That is not the case.
Chips remain a sector with clear institutional preference. AI infrastructure hasn't faced systematic selling either. Big money is simply starting to ask questions that were less critical over the past two years:
Has the stock price of this company already reflected its growth for the next two to three years? As AI CapEx continues to grow, who can truly turn capital expenditure into profits? If the market corrects, which type of assets have the most crowded institutional positions and are most likely to be the first to be liquidated?
This is also the most important change in the Q2 13F: Wall Street is visibly starting to discuss "who offers better risk/reward within AI."
And Berkshire Hathaway, Tiger Global, Third Point, and Duquesne (under Druckenmiller) happen to provide four completely different answers.
II. Four Institutions, Four "Risk/Reward Mindsets"
1. Berkshire Hathaway: Starting to Deploy Cash, Making a Big Bet on Google
Among this round of 13Fs, Berkshire Hathaway's move on Alphabet is particularly noteworthy.
At the end of Q1, Berkshire's disclosed holdings in Alphabet's A and C shares totaled approximately 57.84 million shares; by the end of Q2, that number had risen to about 106 million shares, an increase of over 80%.
Calculated by quarter-end market value alone, Alphabet has already jumped to become one of Berkshire's most important public U.S. stock assets. Meanwhile, Berkshire also increased exposure to airlines and homebuilders like Delta Air Lines and Lennar.
It's worth noting that Alphabet might be one of the "Magnificent Seven" least resembling a "pure AI trade."
Over the past few years, one of the market's biggest concerns has been whether generative AI would change the search gateway, even eroding the core commercial moat that Google Search has long held.
On the other hand, Alphabet still possesses Search, YouTube, Google Cloud, its advertising business, and a massive cash flow base.
Therefore, Berkshire's heavy bet is precisely on whether a company with still-strong cash flow and a core business not yet disproven, but long under pressure due to AI concerns, has room for repricing?
This is a completely different trade from chasing the hottest AI winner.
2. Tiger Global: Trimming Mega-Cap Tech, But Still Focused on Tech Positions
Tiger Global's portfolio provides another very typical sample.
In Q2, it reduced its Alphabet holdings from about 10.63 million shares to about 5.81 million shares, a decrease of 45.4%. Broadcom holdings were nearly halved, and TSMC also saw reductions. Microsoft, Meta, and NVIDIA were also trimmed to varying degrees.
If we stopped here, it would be easy to conclude that "Tiger is starting to exit AI."
But looking at what it bought paints an almost opposite picture.
Tiger initiated new positions in AMD, Applied Digital, Cerebras, etc., in Q2. Its portfolio also included AI computing power and data center-related assets like Cipher Digital and Core Scientific. Intel holdings increased from about 1.64 million shares to about 4.25 million shares.
This looks more like a rebalancing within AI positions—reducing exposure to extremely crowded top-tier assets and moving some chips toward the next layer of opportunities where market consensus isn't as strong.
NVIDIA is the most classic example. A fund's long-term bullishness on AI computing power doesn't mean it must perpetually increase its NVIDIA position.
As long as the position weight is already sufficiently high, or the stock price rises faster than earnings estimate revisions in a given phase, reducing can simply be portfolio management, not a reversal of the industry thesis.
This will also be an increasingly important aspect of U.S. stocks going forward: continued earnings growth does not necessarily mean the stock price will continue to rise in the same way it has over the past two years.
Because what determines the stock price isn't just "how good the results are," but also how much the market already believed in advance.
3. Third Point: Starting to Lock in Gains, Seeking AI's Next Wave
Daniel Loeb's Third Point made even more pronounced moves.
In Q2, it completely exited NVIDIA, Broadcom, KLA, Lam Research, and the VanEck Semiconductor ETF (SMH)—a group of core beneficiaries from the AI capital expenditure cycle—while also exiting Meta.
Looking solely at this set of trades, one could almost interpret it as a large-scale "AI reduction."
But Third Point didn't leave technology.
Instead, it significantly increased Alphabet and TSMC, and established new positions in Keysight and Flex. Meanwhile, capital began flowing into media, finance, and industrial directions like Warner Bros. Discovery, Capital One, and Norfolk Southern. Warner Bros. Discovery even became its largest public U.S. stock holding at the end of Q2.
Therefore, Third Point is essentially locking in gains from the most easily understood winners of the first phase and seeking opportunities for the next phase that aren't yet fully priced in by the market.
Why did NVIDIA, Broadcom, KLA, and Lam Research become first-phase winners? Because their logic was too straightforward:
Larger models need more GPUs; advanced chip expansion needs semiconductor equipment; expanding AI clusters need networks, ASICs, and increasingly complex infrastructure.
This logic isn't wrong.
The issue is, when all investors already know this logic, what determines the returns for the next phase becomes whether the actual growth can continue to exceed the already-high market expectations.
This also means top investment institutions are starting to believe that the most obvious alpha from AI's first phase is becoming increasingly expensive.
4. Druckenmiller: Only Trading on Mispricing
If looking for a fund that best explains the thinking of institutions this round, Stanley Druckenmiller's Duquesne might be the most typical sample.
At the end of Q1, it still held Broadcom and Micron. By Q2, both had disappeared from the 13F.
But at the same time, Duquesne initiated new tech positions in Alphabet, AMD, Palo Alto Networks, etc., and continued to increase holdings in TSMC and STMicroelectronics: TSMC increased from about 495,000 shares to about 590,000 shares, and STMicroelectronics rose from about 2.61 million shares to about 3.10 million shares.
At first glance, it even seems contradictory. They are all semiconductors, why sell some and buy others?
The answer might be precisely the most important keyword of this 13F round: mispricing/expectation gaps.
If a company's stock rises too fast and the market has already priced in growth for the next two to three years, even if the long-term industry thesis remains correct, it's perfectly reasonable to take profits first.
Conversely, if another company's profit cycle is improving and the market hasn't yet formed a consensus, then even if it's not the hottest AI leader, it might offer a better risk/reward profile.
Getting the industry call right is only the first step in investing. The valuation, entry point, and how much the market already believes at that time truly determine the ultimate returns.
III. From "Buying AI" to "Calculating Risk/Reward": What Is Wall Street Trading?
When you put these four institutions together, the truly valuable signal begins to emerge.
First, Alphabet is transitioning from a consensus leader to a "divisive asset."
Alphabet might be the most interesting large-cap tech stock in this cycle. Berkshire significantly increased its stake; Third Point and Duquesne also chose to increase or re-establish positions. On the other hand, Tiger Global slashed its holdings by nearly half.
The same company has received completely different answers from top-tier capital.
The market is uncertain whether AI will ultimately weaken Google Search's moat or further unleash the potential of Alphabet's massive traffic, data, cloud, and computing infrastructure.
Therefore, from this perspective, buyers see cash flow, valuation, and AI's potential upside; sellers see changes in the search gateway, expanding capital expenditures, and the long-term structural challenges the old business model might face.
Such assets are often more worthy of study than companies where "everyone knows they're good," because true excess returns come from where the market holds disagreements.
Second, the semiconductor consensus remains, but the era of "buying chips with your eyes closed" is over.
Looking at the overall 13F data, chips remain a sector where institutions are clearly net long, with the proportion of net buyers significantly higher than net sellers.
But looking at individual star funds reveals huge internal differences. Broadcom—some are selling. TSMC—some are buying, others selling. AMD—some institutions are re-establishing positions. NVIDIA has gradually transitioned from an almost undisputed core AI asset to a stock that requires re-calculating position costs and crowdedness.
This indicates that semiconductors can no longer be traded as a monolithic beta.
GPU, ASIC, foundry, memory, semiconductor equipment, networking, data center infrastructure—they all seem to belong to AI hardware, but their respective profit cycles, supply/demand positions, and valuation states are already vastly different.
In other words, AI hardware is moving from "buying the industry beta" to a stage that truly tests "individual stock alpha."
Another easily overlooked change is the re-entry of non-AI assets into portfolios.
This isn't a negation of AI, but rather more like a hedge to reduce correlation. Assets such as homebuilders, airlines, finance, healthcare, media, railroads, and industrials are reappearing in significant adjustments by some top-tier institutions:
Berkshire increased exposure to airlines and homebuilders; Third Point directed substantial funds into media, finance, and railroads; Duquesne's portfolio itself is far from being solely centered on AI.
In a sense, precisely because AI has become the market's most conspicuous and easily understood theme, big money increasingly needs to find return sources less correlated with AI.
Over the past two years, correctly judging the broad AI direction itself contributed substantial returns. Moving forward, the importance of portfolio management will only increase.
This might be the most noteworthy aspect of the latest 13F round for ordinary investors, showing us what questions the smartest, most resource-rich funds are starting to disagree on when facing the same industry thesis.
Final Thoughts
Over the past two years, the easiest trade to understand in the U.S. stock market was to find AI and buy into it.
NVIDIA, Broadcom, Meta, Microsoft, TSMC, and the entire semiconductor supply chain all simultaneously enjoyed the multiple benefits of industry growth, earnings revisions upward, and valuation expansion.
But the latest round of 13F is sending an increasingly clear signal—AI is not over, but the phase of "as long as the direction is right, everything can rise together" is ending.
This is the most noteworthy change in the Q2 2026 13F filings:
AI has not receded, but the herd is loosening.
The next phase will be determined not by who dares to chase harder, but by who is better at calculating risk/reward.
