Research · AI infrastructure

Leopold Aschenbrenner's 13F is not "short AI." It is a bottleneck map.

Published May 18, 2026
$13.68B
Total reported value
$8.46B
Reported puts
$1.36B
Reported calls
$3.86B
Non-option longs
42
Filing rows

Bottom line

The latest Situational Awareness LP 13F does not prove that Leopold Aschenbrenner is simply "short AI." That reading is too clean.

The filing shows something more useful: a large reported put book against the semiconductor and AI infrastructure complex, sitting next to a long book that still depends on AI demand. The long exposure is concentrated in physical bottlenecks: power, memory and storage, data-center capacity, and high-performance compute conversion assets.

So the Clarion read is straightforward. The filing is a high-signal research prompt, not a copy-trade instruction. It argues for studying whether AI infrastructure leadership is rotating from accelerator scarcity toward bottleneck capture. It does not prove that AI demand is rolling over.

What the filing actually says

Situational Awareness LP filed its Q1 2026 Form 13F-HR on May 18, 2026. The report date was March 31, 2026. That date distinction matters. The filing is a quarter-end snapshot, not a live portfolio feed.

The public information table shows 42 rows and $13.68 billion of reported value after converting the 13F values into dollars. Reported puts account for $8.46 billion. Reported calls account for $1.36 billion. Non-option common equity accounts for $3.86 billion.

That split is what made the filing travel online. The put book is large and it is aimed at the heart of the AI trade: SMH, Nvidia, Broadcom, AMD, Micron, Taiwan Semiconductor, ASML, Intel, Oracle, Infosys, and Corning.

But 13F option rows do not disclose strike, expiry, premium paid, delta, spread structure, or the rest of the risk book. The filing does not show cash, swaps, direct shorts, private holdings, non-reportable positions, or short stock. A reported put is evidence of option exposure. It is not, by itself, proof of a clean net short.

That caveat is not a technicality. It is the difference between reading the filing as a risk-management document and reading it as a headline.

What he still appears long

The long book is not anti-AI. It is more physical than the popular AI basket.

The largest common-equity long is Bloom Energy at $879 million of reported value. Sandisk is $724 million. CoreWeave is $556 million. IREN is $401 million. Core Scientific is $389 million. Applied Digital is $320 million. Those are not random positions if the question is where AI infrastructure bottlenecks could migrate.

Grouped by Clarion's first-pass buckets, the non-put exposure is concentrated in:

Bucket Reported exposure Examples
Data-center and compute capacity $1.81B CoreWeave, IREN, Core Scientific, Applied Digital, CoreWeave calls
Memory and storage $1.54B Sandisk common, Sandisk calls, Micron calls, Micron common
Power and grid bottlenecks $1.09B Bloom Energy, T1 Energy, Power Solutions, Solaris Energy, Babcock & Wilcox
Miners and HPC conversion $0.32B Riot, CleanSpark, Bitfarms, Bitdeer, Hive
Residual semis and optionality $0.41B TSM calls and smaller common-equity semi positions

This still looks like an AI capex book. It just does not look like a broad "own every semiconductor winner" book.

The implied question is not whether AI demand exists. The implied question is who captures the next dollar of economics if chips stop being the only scarce asset.

What he appears short or hedged

The bearish or hedge exposure is concentrated and new versus the prior quarter in the parsed 13F comparison.

Exposure Reported puts Names
Semi hardware hedge or short $7.36B SMH, Nvidia, Broadcom, AMD, Micron, TSM, ASML, Intel
AI infrastructure and software hedge or short $1.08B Oracle, Infosys
Component supply-chain hedge or short $0.02B Corning

This is the part of the filing that deserves attention. The puts target the core AI hardware supply chain rather than one expensive stock. Broad semis through SMH sit at the top of the list. Nvidia, Broadcom, AMD, TSM, ASML, and Micron are all represented. Oracle adds a different angle: AI infrastructure spending, cloud capex, and financing intensity.

There are four plausible readings.

First, this could be a portfolio hedge. A book long power, memory, data centers, and HPC conversion still has AI beta. Puts against crowded AI hardware could protect that book from a broad AI unwind.

Second, this could be a pair trade. Long the bottleneck beneficiaries. Hedge or short the over-owned hardware layer.

Third, this could be a directional short basket. The fund may believe semis and adjacent AI infrastructure have outrun fundamentals.

Fourth, this could be an optical illusion created by 13F option reporting. Without strikes, expiries, premiums, and deltas, reported values can make an options book look cleaner than the economic exposure really is.

The filing alone does not let us choose among those four. That is the central point.

The misconception to avoid

The internet wants the simple version: "Aschenbrenner is short AI."

That is not what the first-party filing proves.

A cleaner statement is: Situational Awareness LP reported a large put book against AI semis and infrastructure, while keeping major long exposure to AI-related physical bottlenecks. That is different from being generically bearish on AI. It is closer to saying the easy semiconductor trade may be crowded, while the next scarcity rents may sit in power, storage, data-center capacity, and contracted compute infrastructure.

That distinction matters for investors. If the thesis is "AI is dead," the trade is broad de-risking. If the thesis is "AI bottlenecks are rotating," the work is more specific. You need to find the bottleneck, find who owns it, and decide whether the economics accrue to equity holders at today's price.

Why power, memory, and data centers matter

The physical bottleneck thesis has three moving parts.

Power is the obvious one. AI data centers need electricity, grid interconnects, backup generation, and infrastructure that can be deployed at the speed customers want. If power becomes the binding constraint, the winner is not automatically the company with the best AI model or the best GPU. The winner may be the company with scarce sites, equipment, contracts, and the balance sheet to build.

Memory and storage are the second piece. AI training gets most of the attention, but inference, data movement, and model-serving workloads can create persistent demand for memory bandwidth and storage. That does not mean every memory stock is a good investment. Memory is cyclical. Supply can catch up. Pricing can break. The diligence question is whether AI demand changes the duration or amplitude of the cycle enough to matter.

Data-center capacity is the third piece. CoreWeave, IREN, Core Scientific, and Applied Digital are different businesses, but the shared question is the same: who has contracted power, credible customers, durable financing, and economics that survive outside a perfect capital market?

That is where the research should move next. Not "what did he buy?" but "where do bottleneck economics accrue?"

The Clarion POV

Clarion's initial view is: research signal high, trade signal not yet.

The signal is high because the filing pushes against a lazy AI framework. It says the market may have over-compressed "AI infrastructure" into semiconductors, when the harder problem is physical deployment. That is a useful challenge.

The trade signal is not there yet because the 13F cannot answer the questions that matter most. It cannot tell us whether the puts are hedges or directional shorts. It cannot tell us the option economics. It cannot tell us whether the fund still owns the same exposures today. It cannot tell us whether the public long names are attractive after the filing date.

The practical conclusion is not to mirror the book. The practical conclusion is to build a research stack around bottleneck capture.

Before making similar investment decisions, we need to answer six questions:

Question Required answer
What is the causal bottleneck? It must be measurable, not thematic.
Who captures the economics? The company needs pricing power, margin capture, contract protection, asset scarcity, or some combination.
What is already priced in? The valuation must leave room for the bottleneck thesis to be wrong or early.
What breaks the thesis? The kill condition must be defined before sizing.
What is the clean expression? Direct long, pair trade, hedge, watchlist, or no action.
What does the market regime allow? A good idea can still be a bad entry if the parent tape is hostile.

The company work starts here

The long-side diligence should start with the bottleneck names, not the most talked-about shorts.

Bloom Energy needs a power-infrastructure analysis: backlog, data-center customer evidence, margins, cash burn, dilution risk, and whether fuel cells are a real AI power solution or a thematic proxy.

Sandisk needs a memory and storage-cycle analysis: NAND pricing, AI storage intensity, spin/separation context, and where it sits against Western Digital, Micron, Samsung, and SK Hynix.

CoreWeave needs a capacity and financing analysis: customer concentration, GPU financing, debt, lease obligations, and whether economics survive if AI infrastructure multiples compress.

IREN, Core Scientific, and Applied Digital need power-site and customer-contract work. The question is not whether they are exposed to AI. The question is whether they own scarce capacity with durable economics, or whether they are financing vehicles levered to capital-market appetite.

On the put side, the work is different. For Nvidia, Broadcom, AMD, TSM, ASML, Micron, Oracle, and SMH, the question is what risk the puts are underwriting: valuation, export controls, capex digestion, customer concentration, gross margin pressure, foundry geopolitics, or the semiconductor cycle.

If the answer is "all of the above," the trade is probably too vague. If the answer is specific, testable, and not already priced in, it becomes a research candidate.

What would make this publishable as a trade thesis

This note is not a trade recommendation. It is a map.

To become a trade thesis, the next note would need company-level evidence. For a long candidate, that means first-party filings, customer contracts, backlog, margin structure, balance sheet, and explicit failure modes. For a hedge or short candidate, it means a precise risk that earnings estimates or valuation have not absorbed.

The best version of the idea is not "short AI." It is more disciplined:

AI demand can remain real while the equity market changes who gets paid for it.

That sentence is the research agenda.

Sources and method

Holdings claims in this note come from first-party sources: the SEC EDGAR filing index for Situational Awareness LP's Q1 2026 Form 13F-HR, the Q1 2026 13F information table, SEC company submissions, SEC IAPD adviser records, the official Situational Awareness LP site, and Leopold Aschenbrenner's first-party thesis site.

Secondary commentary and social posts are useful for understanding what the market is saying. They are not the source of truth for holdings.

Key caveats:

  • The 13F report date is March 31, 2026. The filing date is May 18, 2026.
  • 13F filings omit cash, swaps, direct shorts, private holdings, and many non-reportable positions.
  • 13F option rows do not disclose strike, expiry, premium, delta, or whether the position is part of a spread.
  • Reported value is not the same thing as current market value.
  • A reported put is not proof of a net short.

Primary-source links:

Top reported rows

IssuerInstrumentReported valueStatus vs. Q4
VanEck Semiconductor ETFPut$2.04BNew
NvidiaPut$1.57BNew
OraclePut$1.07BNew
BroadcomPut$1.01BNew
AMDPut$0.97BNew
Bloom EnergyLong$0.88BReduced
SandiskLong$0.72BIncreased
MicronPut$0.58BNew
CoreWeaveLong$0.56BIncreased
TSMCPut$0.54BNew
ASMLPut$0.49BNew
MicronCall$0.42BNew
IRENLong$0.40BIncreased
Core ScientificLong$0.39BReduced
SandiskCall$0.39BNew