Tech·Beginner

The Man Who 'Called' the AI Boom: What the Viral Leopold Reels Get Right (and Badly Wrong)

May 21, 2026·9 min read·1,763 words

#ai #leopold-aschenbrenner #situational-awareness #agi #ai-bubble

Key Takeaways

→A 23-year-old ex-OpenAI researcher, Leopold Aschenbrenner, wrote a 165-page essay in June 2024 predicting AGI by 2027 — then turned it into a hedge fund reportedly up ~2,000% in 2025 [1][2].
→The viral reels are built on real numbers stitched into a misleading story. The "$725B" and "44 jobs" figures exist — they just don't mean what the reels claim [3][4].
→His "Counting the OOMs" framework (compute gets ~10x every two years) is described accurately. His infrastructure and energy bets are tracking *ahead* of schedule [5].
→His clearest misses: AI revenue undershot his target, and open-source models thrived when he predicted they'd fade [6].
→The reels can't have it both ways: you can't say "it's a bubble" and "the smart money proves he's right" in the same breath. Record spending is also exactly what a bubble looks like [7].

The honest takeaway isn't "humanity is cooked." It's: one specific person made some falsifiable bets, and we now have two years of receipts. That's far more interesting than doom.

---

Introduction

A friend sent me a reel last week with the energy of someone who'd just seen the future. The pitch: a former OpenAI researcher quit, started a hedge fund, grew it 2,200%, and proved that AI will beat humans at most cognitive work *by next year*. The closing line — "follow for part two" — did its job. I almost shared it.

Then I did the thing the reel was counting on me not to do. I checked.

What I found is more interesting than the doom. The essay is real. The man is real. Some of his predictions are landing early. But the reel is a textbook example of how a genuinely important idea gets flattened into a dopamine hit — real facts, wrong frame. So let's separate the signal from the slot machine.

---

Who Leopold Actually Is (the reel's first small lie)

Leopold Aschenbrenner is a German researcher who graduated from Columbia as valedictorian at 19, then joined OpenAI's "Superalignment" team alongside Ilya Sutskever [1].

The reel says he "quit." He says he was fired — in April 2024, allegedly over an information leak, which he disputes [1]. Small detail, but it sets the tone: the reel rounds every fact toward the most dramatic version.

After leaving, he published *Situational Awareness: The Decade Ahead* in June 2024 — a 165-page essay arguing that almost nobody is pricing in what's coming [5]. Then he started a fund, Situational Awareness LP, seeded by Stripe's Collison brothers, Nat Friedman, and Daniel Gross, and co-run with Carl Shulman [2].

So far the reel is exaggerating, not lying. That changes.

---

The "2,200% in Two Years" Claim

Here's where precision matters.

The real reported figure is a 2025 return of roughly +2,065% — one explosive year, not two [2]. The fund only launched in September 2024 with about $225M, and grew its disclosed US equity book to $5.52B by the end of 2025 — a ~22x jump in twelve months [2].

Two things the reel leaves out, both of which make a better blog than the hype does:

1.He isn't "just betting on AI going up." He runs large *short* positions against some chip names and concentrates his longs in power and infrastructure — Bloom Energy, CoreWeave, Constellation, Vistra — not Nvidia or Microsoft [8].
2.His one-line thesis is the opposite of "buy the obvious AI stocks": "It's not algorithms that are going to be the bottleneck… It's electrons." [8]

The story isn't "kid gets rich on AI hype." It's "kid bets that the bottleneck is the power grid, and the power grid agrees." That's a thesis, not a meme.

---

"Counting the OOMs" — credit where it's due

This part of the reel is actually accurate, and it's the most important idea to understand.

OOM = "order of magnitude" = 10x. Leopold's argument is that three things compound: raw compute (~0.5 OOMs/year), algorithmic efficiency (~0.5 OOMs/year), and "unhobbling" — turning a chatbot into an agent that can use tools [5]. Stack those and you get another *preschooler-to-smart-high-schooler*-sized leap by 2027, the same size of jump we saw from GPT-2 to GPT-4 [5].

You don't have to agree with the conclusion to find the framework clarifying. The reel quotes this faithfully — which makes the next part more frustrating.

---

"$725 Billion on Just the Clusters" — real number, wrong meaning

This is the reel's most quotable stat, and it's true — about something else.

The four big hyperscalers (Google, Amazon, Microsoft, Meta) do plan to spend about $725 billion on capex in 2026, up 77% from last year [3]. Bloomberg, the Financial Times, and Statista all confirm it [3][10].

But Leopold's actual prediction was a single trillion-dollar cluster — one mega-project — by 2027 [5]. The $725B is *total industry capex* across four companies, covering data centers, chips, and networking. No single trillion-dollar cluster exists yet. The reel takes an industry-wide number and pins it to an individual-cluster prediction. The buildout is genuinely tracking ahead of his curve — but not the way the reel says.

---

"GPT-5.5 Aces 85% Across 44 Jobs" — the most distorted claim

This one falls apart on contact.

The "44 jobs" is GDPval, a benchmark OpenAI released in September 2025 that tests models on real professional deliverables across 44 occupations in 9 sectors [4]. That part is real. The "85%" is not.

The actual launch numbers: GPT-5 won or tied human experts about 40.6% of the time; Claude Opus 4.1 scored around 49% [11]. That's a *win-or-tie rate against seasoned professionals*, not "acing 85%." And "operating an entire computer like a person" describes a different capability (computer-use agents) bolted onto the GDPval stat for extra shock [4].

Worth noting GDPval's own blind spot: it only covers higher-paid knowledge work — construction, transport, agriculture, and hospitality are entirely absent [12]. Impressive, but not "AI beats humans at most cognitive work."

Rule of thumb for any AI reel: when a single number sounds like a movie trailer, it's usually two real numbers in a trench coat.

---

So How Did His Predictions Actually Do?

This is the section the reels never make, because a scorecard is less viral than a prophecy. In March 2026, a thorough community review graded his essay claim-by-claim [6]. The verdict is honestly *mixed* — which is the most useful thing about it:

→Tracking ahead: infrastructure investment and algorithmic efficiency [6].
→Roughly on track: raw benchmark capability — though the "shocking qualitative leap" he described hasn't quite *felt* like he promised [6].
→Clear misses: he predicted ~$100B AI revenue run-rate by mid-2026; the best is closer to $60B [6]. And his biggest error — he bet open-source would fade and a durable US "algorithmic moat" would form. Instead, capable AI diffused *faster* than his model assumed, partly thanks to Chinese labs like DeepSeek [6].
→Still unresolved: AGI-by-2027. It looked implausible six months ago; recent jumps in software-engineering ability have made it credible again [6].

That's the grown-up version of the story: a smart person made falsifiable bets, and the scorecard is part-vindication, part-correction. Far more honest than "he was right, we're cooked."

---

The Contradiction the Reels Never Resolve

Here's the part that should bother you, and it's the strongest thing you can put in your own head.

You probably reacted to the reel the way I did — half "wow," half "isn't this a bubble?" Both reactions are mainstream among serious people, and they point in opposite directions.

On the bubble side: analysts at Bernstein call a bubble the "likely outcome," Ray Dalio sees dot-com parallels, and — remarkably — Sam Altman himself acknowledged in 2025 that he thinks a bubble is underway [7]. The specific fears are *circular financing* (Nvidia invests in OpenAI, OpenAI buys Nvidia chips), extreme market concentration, and a yawning gap between spend and returns — one widely-cited MIT figure claims 95% of enterprises see zero return on generative AI so far [13].

On the boom side: BlackRock argues real earnings growth separates today from 2000, and Jensen Huang argues efficiency gains *increase* compute demand rather than killing it [7].

Here's the trap the reel walks you into: it uses record spending as proof the thesis is right. But record spending is also exactly what the top of a bubble looks like. "The institutions are spending trillions" is evidence for *both* stories. A reel that tells you it settles the debate is selling you certainty that doesn't exist.

---

A Final Thought

I don't think humanity is "cooked," and I don't think it's all hype. I think a specific 23-year-old wrote down specific predictions, put his own money behind them, and gave the rest of us something rare in this industry: a falsifiable bet with a date on it.

That deserves better than a reel that flattens it into a cliffhanger. Read the actual essay [5]. Read the scorecard [6]. Notice that the most honest answer — "tracking ahead on hardware, behind on revenue, undecided on AGI" — is the one nobody can turn into a 30-second hook.

The reels want you to feel the future is settled. The evidence says it's still being written. That should make you more curious, not more afraid.

---

If a number in an AI reel makes your stomach drop, that's the moment to open a new tab — not to hit share. Most of the doom evaporates on contact with a primary source.

---

References

// you just read about ai→ you might like

#claude

Stop Treating Claude Like Google: The Mental Shift That Unlocks 10x Better Results

Part 1

8 min read · 2026-05-05

↳ also: ai

#nextjs

Sitemaps + robots.txt: How to Tell Google Every Page of Your Portfolio

Part 4

8 min read · 2026-04-07