最相信 AGI 加速的多頭，Q1 買了數十億的名義 put？

In this round of US stock 13F filings, one of the most closely watched funds by the market is neither Bridgewater nor Berkshire Hathaway, but a fund with a very unique name: Situational Awareness LP.

Its manager, Leopold Aschenbrenner, is not a traditional Wall Street veteran but a former member of the OpenAI Superalignment team. In 2024, he published a lengthy article titled "Situational Awareness: The Decade Ahead," which contained a very radical core thesis, stating bluntly that AGI might arrive much sooner than most people think, and that what will truly be scarce in the future is not just model capability itself, but compute power, electricity, data centers, chips, storage, and the national-level resource competition surrounding the AI arms race.

Two years later, it turns out he was right.

Leopold internalized a set of judgments about the next decade of AGI and then mapped these judgments onto the capital market. For this very reason, from its inception, Situational Awareness hasn't looked like an ordinary tech fund; rather, it's more like translating the AGI roadmap directly into an AI infrastructure investment map.

This is why, in the AI investment field, every move it makes attracts significant market attention. The latest disclosed 13F filing shows that this most knowledgeable AI bull seems to be quietly building a large-scale put option position.

1. SALP: A Fund Built on the Conviction of AGI

Public information shows that Leopold founded an investment firm focused on AGI, backed by prominent Silicon Valley figures such as Patrick Collison, John Collison, Nat Friedman, and Daniel Gross.

According to market reports, Situational Awareness achieved a net return of approximately +47% after fees in the first half of 2025, significantly outperforming the S&P 500 and tech hedge fund indices over the same period. Its uniqueness lies in the fact that it isn't simply bullish on "tech stocks," but is highly focused on AI infrastructure, betting on where AI capital expenditures will ultimately flow.

As mentioned at the beginning, his underlying logic is that if AGI truly accelerates, the first companies to be revalued might not be application-layer companies, but rather those controlling compute power, electricity, data centers, storage, optical communications, semiconductor equipment, and energy resources. Therefore, its high returns don't rely on simply buying indices but on leveraging a group of high-beta AI infrastructure names to widen the gap, such as Bloom Energy, Sandisk, Lumentum, CoreWeave, and Core Scientific.

First, it's necessary to explain what a 13F is.

A 13F is a quarterly filing of holdings submitted to the SEC by institutional investment managers in the US, typically used to observe quarterly changes in the holdings of large funds in US stocks, ETFs, and related options. However, it is essentially just a snapshot at the end of the quarter, telling the market "what was disclosed at a specific point in time." It cannot fully reconstruct a fund's complete trading strategy. Especially for the options part, 13F filings do not reveal strike prices, expiration dates, or whether they are paired with other positions, nor can they directly infer the fund's true net exposure.

This is the part most easily misinterpreted when reading this document.

The reporting date for this Q1 13F was March 31st. Last10K shows the filing was submitted on the evening of May 15th Eastern Time, but the SEC acceptance time was May 18th. This means it's not simply a case of "not filing"; there is a time lag between the submission and when the market actually sees the disclosed results. This is why discussions like "waiting for Leopold's 13F" appeared on social platforms.

More importantly, the disclosed results of this 13F were not exactly what the market initially imagined. Many originally thought Leopold would continue to heavily increase positions in core AI assets like NVIDIA, Broadcom, AMD, TSMC, and ASML. However, the reality is that SALP built a large number of new PUT option positions, covering core AI and semiconductor names such as the SMH Semiconductor ETF, NVIDIA, Oracle, Broadcom, AMD, Micron, TSMC, ASML, and Intel.

This led the market to reconsider a question: Why is the person who most believes in the accelerated arrival of AGI starting to buy insurance on AI leaders?

Simply attributing this to "being bearish on AI" is too crude. What is truly worth analyzing is the macroeconomic backdrop against which he made this move and the implication of how the AI trading structure has changed.

2. Deciphering SALP's Latest 13F: From Betting on AI to Managing AI Volatility

The most striking action revealed in this 13F is undoubtedly SALP's massive new put option positions:

• The largest is the SMH Semiconductor ETF PUT, with a disclosed value of approximately $2.043 billion;
• Followed by NVDA PUTs, valued at about $1.568 billion;
• Then ORCL PUTs, approximately $1.073 billion;
• AVGO PUTs, about $1.006 billion;
• And AMD PUTs, approximately $969 million;
• Additionally, it established new positions in MU PUTs, TSM PUTs, ASML PUTs, INTC PUTs, etc.;

On the surface, this looks very much like betting against AI leaders. However, the issue is that PUTs do not necessarily represent unilateral short-selling. After all, the option amounts in a 13F are more like the notional value disclosed based on the size of the underlying securities, not the actual premium cost incurred by the fund. Crucially, the 13F does not show strike prices, expiration dates, whether they are paired with other positions, or the true net exposure within the portfolio.

Therefore, it is imprecise to directly say Leopold is "completely bearish on NVIDIA and semiconductors." A more reasonable interpretation is that he is buying "insurance" for his long portfolio of AI infrastructure names. Many of SALP's original holdings are themselves high-beta, highly volatile companies sensitive to interest rates, such as Bloom Energy, CoreWeave, Core Scientific, IREN, Applied Digital, and Sandisk. While their long-term logic is tied to AI infrastructure, their short-term stock prices often heavily depend on risk appetite and the valuation environment.

Once the market starts de-risking due to rising oil prices, recurring inflation, higher interest rates, or geopolitical conflicts, these high-beta assets tend to be sold off first. This is also related to the macroeconomic backdrop at the end of March. On one hand, the Middle East situation and risks of US-Iran conflict pushed up oil price expectations; on the other hand, rising oil prices can exacerbate inflation stickiness, reducing market confidence in interest rate cuts.

For high-valuation growth stocks, this represents a "double pressure": oil prices push up inflation, inflation suppresses rate cuts, interest rates remain elevated, and the valuations of long-duration tech assets face compression.

Viewed against this backdrop, Leopold's move to build a large PUT position becomes easier to understand. It is not a denial of AI, but an acknowledgment that no matter how strong the long-term AI logic is, it cannot completely ignore macroeconomic headwinds.

Especially for a fund like SALP, the portfolio contains many high-beta assets. If it only holds offensive positions, the portfolio's net value will fluctuate significantly during a systemic market downturn. By buying PUTs on liquid, representative core AI assets like SMH, NVDA, AVGO, AMD, and ORCL, it can use relatively standardized tools to hedge against systemic downside risk across the entire AI trade.

The real meaning behind this is that Leopold hasn't turned from an AI bull into an AI bear; he has shifted from "aggressively going long on AI in a unilateral way" to "continuing to bet on AI infrastructure, but starting to manage path volatility."

This is a more mature portfolio management approach.

3. Where is Leopold's Offensive Direction?

If the new PUT positions address the "defensive question," then the lists of increased, reduced, and liquidated holdings truly tell us where Leopold's offensive direction lies.

According to the disclosure, SALP retained and increased positions in several AI infrastructure-related names. For example, it slightly increased its position in Sandisk common stock, and added to positions in CoreWeave common stock, IREN, Applied Digital, Riot Platforms, CleanSpark, Bitfarms, and Bitdeer. Other significant long positions retained include Bloom Energy, Sandisk, CoreWeave, IREN, Core Scientific, and Applied Digital.

This indicates that it hasn't abandoned AI. On the contrary, it continues to bet on the same long-term thesis: AI capital expenditures will continue to trickle down, and the true beneficiaries are the companies that control electricity, data centers, storage, compute capacity, and infrastructure bottlenecks.

This is very close to MSX's Q2 mainline judgment. In our article "AI infrastructure rose all through Q1. In Q2, who can still sustain 'high valuations'?", we emphasized that the center of gravity in the AI trade has shifted from pure GPUs to networking, storage, and electricity. The market now cares more about where the continuously expanding capital expenditures of major tech companies ultimately flow in terms of orders, revenue, and profit. Segments like equipment, networking, storage, and electricity are more advantageous not because they are sexier, but because they better align with the market's current aesthetic for deliverability.

From this perspective, SALP's long positions are quite representative: Bloom Energy corresponds to electricity and independent energy supply; CoreWeave, Applied Digital, Core Scientific, and IREN correspond to data centers, compute hosting, and infrastructure capacity; positions related to Sandisk, Micron, and TSM correspond to storage, semiconductor manufacturing, and hardware supply.

In other words, Leopold hasn't stopped buying AI; he is more concerned about where AI money will ultimately be spent and who can turn that money into revenue on their financial statements.

Looking at the reductions and liquidations, they are also very informative. SALP liquidated its positions in INTC CALL, Lumentum, and Cipher Mining, and reduced its positions in CoreWeave CALL, Bloom Energy, and Core Scientific. The most notable point here is that it is not simply withdrawing from a particular sector, but reducing some positions that have already appreciated significantly, are highly volatile, or have higher leverage characteristics.

For example, with CoreWeave, it reduced CALLs but still holds common stock. This suggests it hasn't completely given up on CoreWeave, but has shifted from a more aggressive options expression back to a relatively controllable common stock expression. Similarly, reducing Bloom Energy and Core Scientific doesn't necessarily mean the logic has failed; it's more likely risk control and profit-taking at the portfolio level.

The liquidation of Lumentum is even more intriguing. In MSX's Q1 review, AI hardware and optical communications were the two strongest themes, with stocks like AXTI, AAOI, LITE, and LWLG achieving double-digit percentage gains. The strength of optical communications essentially stems from the explosive demand for optical interconnects, optical modules, and network links driven by AI data centers. However, the problem is that the stronger a theme runs in Q1, the more prone it becomes to crowded trades and declining risk-reward ratios entering Q2.

Therefore, Leopold selling off LITE and reducing some high-beta AI infrastructure positions doesn't necessarily mean he is bearish on the direction; it might be a more realistic acknowledgment: The most successful trade of Q1 isn't necessarily the most cost-effective trade of Q2.

This is the most important aspect of this portfolio adjustment. It is not a denial of AI, but an active structural shift: moving from buying anything along the AI chain to retaining only those assets that can better absorb long-term capital expenditures, possess more infrastructural characteristics, and are better able to navigate macroeconomic volatility.

What he is giving up is not AI itself, but the linear fantasy that "all AI will rise together."

Essentially, this 13F is just a snapshot as of March 31st and doesn't mean Leopold still holds the exact same positions in May. Nevertheless, it still offers strong insights for the current market conditions.

First, the long-term AI theme is not over, but the trading structure has changed. Going forward, it's not about buying any AI and seeing it rise; it's about who can deliver, who gets the premium, who is too crowded, and who needs hedging.

Second, in an environment of high oil prices, high interest rates, and high volatility, the truly effective strategy is not simply all-out offense or all-out defense, but playing offense with defense. The core positions bet on certainty, marginal positions bet on elasticity, while using hedging tools to control portfolio drawdowns. In essence, Leopold's recent actions demonstrated this logic using his real portfolio.

Third, this confirms a major change in the US stock market for 2026: Index Beta weakens, Structural Alpha strengthens. In the past, just buying the Magnificent Seven or NVIDIA might have guaranteed a win. But the market is now more discerning; it will question every company: Can your AI story ultimately turn into orders? Can it turn into revenue? Can it turn into profits? If not, no matter how high the valuation, it will be compressed.

This is why AI Infrastructure 2.0 becomes important. Future capital won't just look at GPUs; it will track down the chain from compute power → interconnectivity → storage → electricity → data center infrastructure to find the segments that can truly deliver.

Final Thoughts

On the surface, the most eye-catching part of this 13F filing is the string of massive PUT positions.

But if you truly examine the entire portfolio, you'll find that Leopold hasn't "turned from an AI bull into a bear"; instead, he has undergone a more mature upgrade: Long-term, still betting on AI infrastructure; short-term, starting to acknowledge the volatility risks of high-valuation, high-beta assets.

This is the most important takeaway from this 13F. It tells us that the direction of AI might still be correct, but the path to that direction will certainly not be a straight line.

For a true fund manager, the important thing has never been just to bet correctly on the destination, but also to survive the volatility and navigate the journey along the way.

For ordinary investors, the biggest lesson from this 13F is also clear: The AI trade in 2026 has shifted from "buying stories" to "buying delivery"; from "buying leaders" to "finding bottlenecks"; from "unilateral offense" to "playing offense with defense."

This is the most intriguing signal and one that absolutely should not be ignored.