Everyone Is Reading Leopold Aschenbrenner's 13F Wrong

The $8.45 billion in semiconductor puts are being read as a doom trade. The portfolio underneath them tells a different story.

By \[Author\] | Bloomberg Markets | May 18, 2026

When Leopold Aschenbrenner's Q1 2026 13F hit the SEC's EDGAR system, the reaction was immediate and almost uniformly grim. The former OpenAI researcher whose Situational Awareness essay became required reading in both Silicon Valley and hedge-fund circles had just disclosed $8.45 billion in new put positions against the stocks most AI investors now treat as untouchable: Nvidia, Broadcom, TSMC, Oracle, AMD and the VanEck Semiconductor ETF.

The interpretation came fast. The AI insider had turned. The trade was over. The sector's most articulate accelerationist had finally joined the bears.

That reading is too fast and too literal. It treats the filing's most dramatic numbers as the thesis rather than the hedge. Read the whole portfolio, and a more interesting picture comes into focus. Aschenbrenner does not appear to be backing away from AI. He appears to be moving away from the most crowded expression of it while leaning harder into the parts of the buildout the market may still be underpricing.

Start With What He Actually Owns

The cleanest way to misread the filing is to start with the options book and stop there. A better place to begin is with the outright longs and the size of the adds.

Among the largest unhedged positions:

• Bloom Energy: $879 million
• SanDisk: $724 million
• CoreWeave: $556 million

Among the biggest increases:

• CleanSpark: shares up 648%
• Riot Platforms: shares up 87%

That is not the portfolio of someone positioning for an AI air pocket. It is a portfolio aimed at the less glamorous layers of the buildout: power, storage, compute access and energy-rich infrastructure.

That distinction matters. Those are the parts of the AI stack that can keep winning even if the highest-multiple chip names stumble, even if hyperscaler capital spending gets paced more unevenly, and even if investors decide Nvidia at current valuations deserves a lower multiple than it commands today.

The Filing Looks Different Once You Put the Longs Beside the Puts

The raw size of the put book dominated first-pass reactions:

• SMH puts: $2.04 billion
• Nvidia puts: $1.57 billion
• Oracle puts: $1.07 billion
• Broadcom puts: $1.01 billion
• AMD puts: $969 million
• Micron puts: $584 million
• TSMC puts: $535 million

But the scale only tells you something useful once you place it next to the rest of the portfolio.

On the surface, those puts look like a broadside against the AI winners. Inside the full portfolio, they read more like insurance on a long AI infrastructure book than a clean directional short on the future of the sector.

That view gets stronger once you notice what sits next to them.

The Detail Most People Missed

Aschenbrenner also disclosed:

• Micron calls: $422 million
• TSMC calls: $355 million
• SanDisk calls: $389 million

That should have slowed down the hot takes. A manager who is outright bearish on semiconductor demand does not usually pair large put positions with hundreds of millions of dollars of calls in the same ecosystem. What this looks like instead is targeted range exposure: participate if upside continues, protect capital if the multiple or the cycle breaks against you.

Micron is the clearest example. The structure reads less like a short than a defined expression of a specific view. HBM demand could keep surprising to the upside. Memory pricing could still turn volatile. Export controls could still matter. A manager can want exposure to that setup without wanting to absorb the full downside if the trade wobbles.

The same logic applies to TSMC. The company remains central to the AI supply chain. It is also central to one of the market's largest geopolitical risk concentrations. Owning upside while protecting against a sharp repricing is not confused positioning. It is what risk-managed conviction looks like.

The Real Bull Case Sits Outside the Chipmakers

The conventional AI picks-and-shovels trade has become almost too familiar. Buy Nvidia because it sells the GPUs. Buy TSMC because it manufactures the most advanced chips. Buy ASML because the entire stack runs through EUV.

Aschenbrenner's portfolio suggests he may be looking one layer over.

The bet embedded in his biggest longs appears to be that the next choke points in AI are not limited to compute. They are also:

• Power
• Storage
• Independent compute access

That is where Bloom Energy, SanDisk and CoreWeave fit.

Bloom Energy is the clearest tell. At $879 million, it is his largest disclosed long position without a matching hedge structure dominating the story. Bloom makes fuel-cell systems and distributed power infrastructure, exactly the kind of asset that becomes more valuable if AI data centers start straining regional grids and if hyperscalers increasingly need reliable on-site or near-site generation.

For all the obsession with GPUs, power has become one of the most important constraints in AI. Data centers do not run on narratives. They run on electrons. If Aschenbrenner believes the AI buildout is still early, Bloom is a more direct way to own that bottleneck than another crowded chip multiple.

SanDisk fits the same logic from another angle. Training and inference at scale do not only require faster chips. They require vast amounts of fast, reliable storage. A freshly separated storage company can look more interesting than a consensus mega-cap if the market has not fully repriced what the next wave of AI workloads implies for data movement and retention.

CoreWeave is more obvious but still important. It remains one of the clearest direct equity bets on AI compute demand outside the hyperscalers. Even after trimming the call exposure sharply, Aschenbrenner kept a large common-equity position. That suggests discipline on valuation and structure, not a loss of faith in the underlying company.

Why the Bitcoin Miner Adds Matter

The most aggressive move in the filing may not be the put book. It may be the additions to CleanSpark and Riot.

Those companies are still widely filed under the Bitcoin miner label. That framing may already be stale. What they also own is power access, physical infrastructure, operating experience in energy-intensive facilities and, in some cases, optionality to repurpose sites for adjacent compute uses.

The key adds were hard to miss:

• CleanSpark: +648%
• Riot Platforms: +87%

There has been a circulating thesis for two years that parts of the mining sector could evolve into AI infrastructure landlords, hosting providers or energy-arbitrage vehicles. Most investors have treated that as an interesting side conversation. Aschenbrenner's sizing suggests he may take it more seriously.

If he does, the miner adds belong in the same basket as Bloom. They extend the power thesis. They also offer a way to own AI demand through infrastructure that remains cheaper, less loved and less fully explained by the market than the headline semiconductor winners.

The Exits Tell You What He No Longer Wants

The full exits are as revealing as the new positions:

• Intel calls: previously $747 million
• Lumentum: previously $479 million
• EQT Corp: previously $133 million
• Tower Semiconductor: previously $85 million

This looks less like a retreat from AI than a refinement of where he thinks the best risk-adjusted exposure now sits.

Intel was a recovery story. Lumentum was a second-order component bet. Tower was a specialty foundry angle. Those positions depended on a more complicated chain of things going right.

Bloom, SanDisk, CoreWeave, CleanSpark and Riot are conceptually cleaner. If AI demand keeps compounding, the case for them can strengthen even if the market becomes less willing to pay 25 to 30 times sales for the best-known semiconductor names.

The Valuation Hedge Is the Point

The bear case on this filing confuses a hedge on valuation with a hedge on the underlying technology trend.

A manager can believe all of the following at once:

• AI capex is still in the early innings
• Data-center demand will keep expanding
• Power constraints will worsen
• Storage demand will rise materially
• Nvidia, Broadcom and other winners are also expensive enough to justify protection

That combination is more sophisticated than simple bullishness. It accepts that secular trends and stock prices do not move in lockstep. An investor can be structurally right on AI and still lose money buying the wrong expression of it at the wrong price.

That appears to be the lesson embedded in this filing.

The puts hedge valuation risk. They hedge the possibility that the market has crowded into the same names at the same time and assigned them the same immaculate future. They do not, on their own, tell you the manager has changed his view on the scale of the AI buildout.

The Aschenbrenner Framework Still Fits

None of this is separable from what Aschenbrenner has written and said publicly over the past two years. His view has never been that AI will matter in some broad, eventually true sense. His framework has been sharper than that. He has argued that advanced AI arrives on a far faster timetable than institutions, governments and markets are prepared for.

If that remains his worldview, the filing makes more sense than the panic around it.

A manager with that view would not necessarily want maximum exposure to the names everyone already agrees are the winners. He might prefer exposure to the constraints the market still treats as secondary. He might want the companies that benefit if the buildout accelerates harder than expected and the bottlenecks move away from chips alone.

That is the logic connecting Bloom, SanDisk, CoreWeave and the miner adds. It is a portfolio shaped around what happens if AI demand keeps arriving faster than consensus expects.

The Filing Reads Bullish, Just Not in the Obvious Way

The fast-twitch interpretation of this 13F is easy to understand. It is also probably too shallow.

People saw:

• $8.45 billion in new puts
• Big hedges against the semiconductor complex
• Trims in CoreWeave calls

They concluded that Aschenbrenner had turned defensive on AI.

The deeper read is narrower and more interesting. The AI boom can continue. The market's favorite stocks can still be overowned. The better place to be may be the infrastructure behind the infrastructure: the power suppliers, the storage layer, the alternative compute platforms, the energy-rich edge cases the market still treats as speculative.

That is not a portfolio built for AI winter. It is a portfolio built for a world in which AI keeps scaling and the market has misidentified where the next bottlenecks, and the next equity upside, will actually be.

The 13F data cited in this article is drawn from Aschenbrenner's Q1 2026 SEC filing. Position sizes reflect market values as reported at the filing date.