Leopold Aschenbrenner's portfolio: the 5 AI stocks Situational Awareness actually owns

Leopold Aschenbrenner is the 25-year-old former OpenAI Superalignment researcher who in mid-2024 published "Situational Awareness: The Decade Ahead" — the canonical "AGI by 2027-2028, the entire industrial base bends to make it happen" essay. He launched a hedge fund of the same name months later, backed by Patrick Collison, Daniel Gross, and Nat Friedman, with one job: trade the financial expression of the situational-awareness thesis.

The fund doesn't file 13Fs (sub-$100M US-equity AUM, recent vintage). On 2026-05-08, an aggregator surfaced five past trades the fund has disclosed publicly. They are the cleanest available read on how an AGI-imminence-pilled, infrastructure-bottleneck-focused trader actually positions a book.

The holdings are $SNDK, $LITE, $INTC, $COHR, and $BE. Each maps onto a different QuantAbundancia validated AI-infrastructure bubble. There is zero exposure to the obvious "AI plays" — no $NVDA, no hyperscalers, no AI-software basket. The fund is positioned exclusively in the bottlenecks downstream of the GPU.

That construction is the entire lesson. The full portfolio analysis is below.

The thesis: AGI is a buildout, not a chip

To read the portfolio you need the thesis. Aschenbrenner's claim in Situational Awareness, in 90 seconds: model intelligence is rising on a near-vertical curve, the bottleneck to deploying it is industrial — chip fab capacity, memory, electricity, optical bandwidth — and the West will mobilize a wartime-scale buildout to win the AGI race against China.

The implication for trading: the chip companies ($NVDA, $AMD) are already priced for the demand. The constraints downstream of the chips are not. Whoever holds the bottleneck collects rent during the buildout.

Five bottlenecks, ranked by how tight the constraint actually binds:

1. Memory — HBM3E and NAND flash. AI compute is memory-bound, not compute-bound, on most modern training workloads.
2. Optical interconnect — bandwidth between racks and between datacenters. Constrains cluster scaling.
3. Foundry capacity — chips need leading-edge manufacturing. The Intel turnaround thesis.
4. Datacenter power — electricity is the medium-term ceiling; covered in detail in Datacenter power stocks: the second-derivative AI trade.
5. Photonics components — the physical layer of optical interconnect.

Aschenbrenner has all five. He has nothing else.

Position 1 — $SNDK SanDisk · the largest disclosed allocation

The source language describes SNDK as where Aschenbrenner "entered the bulk of his position" — the largest single allocation in the disclosed book. Entered Q4 2025 at ~$230, last close ~$1,559 (+550%).

The bottleneck: NAND flash. Distinct from $MU's HBM3E in the DRAM/HBM Memory bubble but in the same structural position one layer over. NAND is the storage tier that holds model weights, training data, and inference KV caches at scale. The pacing constraint on AI compute is memory bandwidth and capacity at multiple tiers; SNDK plays the persistent-storage tier.

Why it's the largest position: post-WDC spinoff in 2025, SanDisk is the only US-listed pure-play NAND name. The competitive structure is an oligopoly (Samsung, SK Hynix, Kioxia, Micron, SanDisk) with the same demand-supply discipline that drove HBM3E ASPs to all-time highs in 2024-2025. The thesis bets that AI training and inference workloads expand NAND TAM by a step function over 2026-2028.

QA taxonomy fit: sits adjacent to the DRAM/HBM Memory bubble (validated, ~0.71 residualized correlation per The 12 AI bubbles ranked). The bubble's bull case is "AI compute volume × memory ASP" expanding together; SNDK is the NAND expression of that same trade.

Position 2 — $LITE Lumentum · the largest reported gain

Source claims entry ~$40 in 2024-2025, last close ~$902 (+2000%). The largest reported single-position gain in the disclosed book. Even if the initial position was small, after a 20-bagger it is now top-of-book by current market value regardless of original sizing.

The bottleneck: high-speed optical components for AI cluster fabrics. As GPUs scale beyond a single rack, the bandwidth between racks (and between datacenters in geographically-distributed training runs) becomes the binding constraint. Lumentum makes the lasers, transceivers, and pluggable optics that move multi-terabit-per-second signals between AI compute pods.

Why a 20× return: the inflection in AI cluster scaling between 2024 and 2026 turned LITE from a cyclical optical-components company into a pure-play AI-buildout supplier, with multiple quarters of triple-digit-percent revenue growth in the AI-datacenter segment driving multiple expansion on top of EPS expansion.

QA taxonomy fit: the Networking / Optical bubble (validating, ~0.60 residualized). Aschenbrenner owns two of the bloc's top members (LITE + COHR), making the optical layer a 2-name basket inside the fund — which is the right construction given the bubble-level coherence of that bloc.

Position 3 — $INTC Intel · the levered foundry bet

Source claims entry via call options at premium ~$19 in Q1 2025, current premium ~$125 (+510%). This is the option premium return, not the underlying stock return. INTC stock did not move from $19 to $125 in that window; the leverage from a call option position explains a large premium move on a more modest stock move.

The bottleneck: US-domestic leading-edge foundry capacity. Intel's 18A node and the Foundry Services business are the only credible US-soil response to TSMC concentration. Aschenbrenner's thesis is that the AI buildout makes foundry capacity strategically critical, and political will (CHIPS Act, AI Diffusion Rules) backstops the capex required to make 18A real.

Why options instead of stock: the IDM-turnaround payoff distribution is right-tailed and binary at the technology-execution level. If 18A yields hit and Foundry Services lands TSMC-comparable customers, the stock multi-bags. If 18A pushes a year, the stock dies. Calls express that distribution capital-efficiently and limit downside to the premium paid.

QA taxonomy fit: straddles the Semi Equipment / Litho bubble (Intel is one of the largest equipment buyers globally) and the broader compute / GPU bloc (Intel makes Gaudi accelerators). Aschenbrenner's chosen expression is the foundry-turnaround leg, levered.

Position 4 — $COHR Coherent Corp · the optical pair to LITE

Source claims entry ~$60, last close ~$333 (+425%).

The bottleneck: same as LITE — optical components, lasers, photonics for AI cluster interconnect. COHR sits alongside LITE as the second leg of the optical-bandwidth thesis. Different product mix, similar end-market exposure, similar capacity-constrained supply dynamics during the 2024-2026 AI-buildout phase.

Why two names instead of one: the optical bloc has a 2-name internal cluster (LITE, COHR) that trades tighter with each other than with the broader Networking / Optical bubble. Sizing 50/50 across the two captures the bubble's bandwidth-bottleneck signal while diversifying the company-specific risk (product roadmap, customer-concentration, M&A).

QA taxonomy fit: same Networking / Optical validation as LITE. The two together represent Aschenbrenner's bandwidth-layer trade.

Position 5 — $BE Bloom Energy · the on-site power bet

Source claims entry ~$87 end-2024, last close ~$263 (+190%). The smallest reported gain in the book — but still a 3-bagger inside ~18 months.

The bottleneck: electricity at the rack, not the grid. Bloom builds solid-oxide fuel cells that hyperscalers can deploy on the datacenter site to generate power independently of grid interconnect timelines. This is a workaround for the multi-year grid queue backlogs that bottleneck conventional Datacenter Power constituents ($VST, $CEG) — a hyperscaler that buys Bloom doesn't need to wait for new substations.

Why the smallest gain: the on-site fuel-cell thesis is structurally earlier than the grid-side thesis. Bloom's revenue is still in ramp; the bull case is multi-year and needs hyperscaler PPA announcements at scale (not yet broadly disclosed) to validate. Lower beta, longer duration, smaller current return.

QA taxonomy fit: the Datacenter Power bubble (validated, ~0.55 residualized). BE is an alternative construction within that bubble — the on-site/distributed leg vs. the grid-scale IPP leg. Both express AI-power scarcity; different mechanisms.

What the construction tells you

Five names. Four different validated AI-infrastructure bubbles. Zero exposure to the obvious AI tickers.

The pattern:

• Memory: 1 name (SNDK), one of two largest by current value
• Optical interconnect: 2 names (LITE, COHR), basketed
• Foundry: 1 name (INTC), levered with options
• On-site power: 1 name (BE), early-stage
• Compute (NVDA, AMD): zero names
• Hyperscalers (MSFT, GOOGL, AMZN, META): zero names
• AI software (PLTR, SNOW, etc.): zero names

This is exactly the construction that the QuantAbundancia bubble-validation methodology argues for. Read Why correlation > narrative in thematic investing for the full framework, but the 30-second version: editorial taxonomies group by story (anything with "AI" in it) while real co-movement clusters group by exposure (companies that share the bottleneck binding their P&L). Aschenbrenner's portfolio is the bottleneck-exposure version of the AI trade.

Compare with the Hyperscalers failed bubble: five "AI capex" mega-caps that look like an AI bloc on raw correlation but collapse to ~0.05 residualized — they're SPY in a costume. Aschenbrenner owns none of them despite having literally founded a fund on AGI imminence. That gap (the obvious AI play vs. the empirically real AI play) is the trade thesis the portfolio expresses in capital allocation terms.

What's missing from the disclosed book

The five public positions don't span the full QA taxonomy. Bubbles Aschenbrenner has not disclosed exposure to:

• Quantum (validated 0.76, the highest in the entire taxonomy). Aschenbrenner's thesis is AGI by ~2027-2028; quantum-utility commercialization is a 2030+ event. Off-thesis-window.
• Semi Equipment (validated 0.82, the tightest bloc in coverage). Conspicuous absence — $ASML, $AMAT, $LRCX, $KLAC would be obvious vehicles for "wartime AI capex." Possibly held but not disclosed; possibly avoided due to China-export-rule overhang.
• Cooling / DC Infrastructure (validated 0.70, "Boring. Wins."). $VRT in particular would fit the bottleneck-thesis cleanly. Same possibility — held but undisclosed.
• Nuclear / SMR (speculative, 0.70+ on the pre-revenue subset). Strategic-baseload-electricity thesis. Could fit Aschenbrenner's framework.

The undisclosed-but-likely set is informative: any of these could land in the next disclosure (the source teases new releases ~2026-05-15). The QuantAbundancia Situational Awareness portfolio page auto-refreshes when new positions ingest.

Why this matters for retail traders

Two takeaways:

1. The names are public; the construction is the alpha. Anyone could have bought SNDK, LITE, INTC, COHR, BE in 2024. Most retail traders bought NVDA and called it AI exposure. The construction Aschenbrenner ran — bottleneck-mapped, no-overlap-with-the-obvious-trade, sized by bubble coherence — is the difference between "long AI" and "long the binding constraint." Position sizing matters more than ticker selection.

2. The disclosed gains are real but the entries are gone. SNDK at $230 is not coming back; LITE at $40 is not coming back. The actionable read is not "buy these now" but "where is the next bottleneck that the obvious AI trade hasn't priced yet?" The QA taxonomy has 12 bubbles ranked by empirical realness; four of Aschenbrenner's five positions sit in tier-1 or tier-2 validated bubbles. The next position will likely sit in another tier-1 or tier-2 bubble that the market hasn't fully repriced — current candidates from the unfilled list: cooling, semi equipment (if the China overhang lifts), nuclear/SMR.

Bottom line

Aschenbrenner's disclosed book is a 5-position concentrated long expression of the AI-bottleneck thesis: memory (SNDK), optical (LITE + COHR), foundry (INTC, levered), on-site power (BE). The construction maps onto four different empirically-validated QuantAbundancia AI bubbles with zero exposure to the failed-validation hyperscalers or AI-software baskets.

For a retail trader trying to express the same thesis in 2026, the playbook isn't to copy these specific entries — it's to copy the construction logic. Map AI capital flow to bottleneck constraints, size by bubble coherence rather than by name, and avoid the high-narrative low-residualized blocs (Hyperscalers, AI Software, Robotics) where the editorial label does no analytical work.

---

Browse all 12 bubbles ranked by empirical realness for the full taxonomy. Or go deeper on individual bottlenecks: Datacenter power stocks, the Quantum bubble, why Hyperscalers fail validation.

The live Situational Awareness portfolio page shows the current marks and per-name notes, refreshed nightly.