Conviction Is Not A Trade

About three weeks ago, markets watched a historic unwind. The manager was directionally right. He didn’t survive being right. That distinction is the whole post. 

Leopold Aschenbrenner launched a hedge fund called Situational Awareness in 2024 on mathematical brilliance and an OpenAI pedigree. Performance was remarkable, until it wasn’t. Six days after asking investors for fresh capital to buy a drawdown, he met a margin call by selling the public book in one block, at a discount.¹ Reports suggest the fund peaked near $45 billion,³ and the traders copying it likely accounted for a meaningful share of the volatility and distortion we’ve absorbed over recent quarters. 
 
The insight survived July. The structure didn’t. Conviction is not a trade — it has to be expressed through securities, and those securities have to stay synced with the thesis, as it’s recognized and released. The clock that matters usually isn’t yours.  

Two Clocks That Don’t Always Match, But Must Eventually Sync

Forget what SA believed. Look at how they invested. 
 
The thesis was measured in years: a multi-year buildout of computing capacity. The vehicle was measured in days: levered public equity financed by prime brokers who get to reconsider in real time. A slow thesis in a fast wrapper. 
 
That mismatch existed from day one, and by itself it wasn’t fatal. Plenty of managers run slow theses in fast vehicles. They survive by holding more collateral, rebalancing routinely, and harvesting gains on the way up — leaving some upside unrealized, dampening the downside when it inevitably arrives. Tried and true, and reliably unrewarding when things are going well. Which is precisely when nobody questions skipping it. 
 
A prime broker doesn’t grade you on the year. It grades collateral sufficiency throughout the day, every day. 

 
Outgrowing Protection by Winning

Both sides of the trade worked. Longs in memory, data centers, and power. Shorts in the software that AI was expected to disrupt. Reported gains north of 1,000% since inception.^3,4,7 Then July unwound both ends at once — down 67% in a single month,⁴ with Citadel stepping in before it got worse,² partly for the blood in the water, partly to preserve the markets it depends on. 
 
Here’s the number worth sitting with: after the unwind, SA was still up 80% on the year;⁴ not since inception. On the year, after the month that ended its public equity adventure. Some of that is likely the write-up of private carrying values now under question. But still. 
 
What killed the fund wasn’t the insight, and it wasn’t a losing year. It was timing and leverage, seasoned with hubris that had been working too well to interrogate. 
 
Look closer and the failures compound. The long-short book wasn’t hedged — long the buildout’s suppliers, short its victims, both legs pay off if AI capex accelerates. That’s a Texas hedge: two positions, one exposure, twice the size, supercharged with leverage.⁴ 
 
The private book wasn’t ballast either. A crossover structure earns its keep when privates deliver genuinely different return streams; skeptics call the lazy version “volatility laundering.” Here the privates tripled down on the same thesis. And the ratios drifted as success compounded — the disclosed long book went from $255 million to $20 billion in six quarters, roughly eighty times, while private marks moved only when rounds were raised.⁵ The side meant to steady the portfolio kept shrinking, relatively, by winning. 
 
By June 30th, two memory names — SanDisk and Micron — were roughly 56% of the longs.^3,5 Concentration can be the residue of being right. It still has to be managed. Selling winners while the thesis holds is the hardest discipline to learn, and hardest to apply exactly when it’s working.  

Some Ideas Have No Wrapper  

SA at least had an expression to outgrow. Sometimes there isn’t one. Water scarcity is my example: the research is compelling, the direction unquestioned. But the public companies are industrials, so you’re buying industrial cyclicality with a water label. Own the rights directly and you get pure exposure — plus a government that can take them by eminent domain at peak value. 

A view you can only express by ignoring risk isn’t an opportunity. It’s a thought exercise waiting for a trade. 
 

And Some Wrappers Have No Idea  

Momentum is the inverse: an expression with no insight. A strategy explains why something is mispriced and names the evidence that would confirm or kill it. A tactic helps you enter, size, and time. Momentum makes price both the thesis and the test. I wrote in Capturing Pent-Up Alpha that early and wrong are indistinguishable in the moment; the discipline is knowing the difference.⁶ When the evidence has landed and price has run past it, that’s a trim, not a hold. 

What Actually Got Tested 

None of this settles anything about AI infrastructure. The demand is real, the buildout is real, the thesis is sound. July tested the wrapper, not the research — and the recalibration success demanded never came. 
 
Being right is the research problem, and it’s the one everyone wants to talk about. Expressing it is the structural one. That’s where risk management earns its keep.

Footnotes 
¹ David Faber, “AI investor Leopold Aschenbrenner forced to unwind all public stock positions after steep losses, sources say,” CNBC, July 30, 2026.  
² Katherine Burton and Sridhar Natarajan, “Situational Awareness Drops to $10 Billion on Citadel Pact,” Bloomberg, July 30, 2026.  
³ Yun Li, “Situational Awareness filing shows AI bets before forced portfolio sale to Citadel,” CNBC, August 14, 2026.  
⁴ Peter Rudegeair, “How a ‘Texas Hedge’ Amplified the Losses at Situational Awareness,” The Wall Street Journal, August 15, 2026.  
⁵ Situational Awareness LP, Forms 13F-HR, periods ended December 31, 2024 through June 30, 2026; most recent filed August 14, 2026.  
⁶ Mark Scalzo, Capturing Pent-Up Alpha, Validus Growth Investors LLC dba Validex Global Investing, March 30, 2026. 
⁷ Peter Rudegeair, “The 24-Year-Old AI Wiz Who Counts Jane Street as an Investor,” The Wall Street Journal, June 8, 2026.